Upload 28 files

app.py

@@ -0,0 +1,19 @@
+from  hubconf import custom
+model = custom(path_or_model='best.pt')  # custom example
+# model = create(name='yolov7', pretrained=True, channels=3, classes=80, autoshape=True)  # pretrained example
+
+# Verify inference
+import numpy as np
+from PIL import Image
+import gradio as gr
+
+# imgs = [np.zeros((640, 480, 3))]
+# imgs = 'inference/images/meal.jpg'
+
+# results = model(imgs)  # batched inference
+# results.print()
+# results.save()
+
+
+
+gr.Interface(inputs=["image"],outputs=["image"],fn=lambda img:model(img).render()[0]).launch()

best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e649c278fc179eb0b75e136f2ea86a58a2dd84a7f77dba4fe2b93987d56e63df
+size 75247994

hubconf.py

@@ -0,0 +1,97 @@
+"""PyTorch Hub models
+
+Usage:
+    import torch
+    model = torch.hub.load('repo', 'model')
+"""
+
+from pathlib import Path
+
+import torch
+
+from models.yolo import Model
+from utils.general import check_requirements, set_logging
+from utils.google_utils import attempt_download
+from utils.torch_utils import select_device
+
+dependencies = ['torch', 'yaml']
+check_requirements(Path(__file__).parent / 'requirements.txt', exclude=('pycocotools', 'thop'))
+set_logging()
+
+
+def create(name, pretrained, channels, classes, autoshape):
+    """Creates a specified model
+
+    Arguments:
+        name (str): name of model, i.e. 'yolov7'
+        pretrained (bool): load pretrained weights into the model
+        channels (int): number of input channels
+        classes (int): number of model classes
+
+    Returns:
+        pytorch model
+    """
+    try:
+        cfg = list((Path(__file__).parent / 'cfg').rglob(f'{name}.yaml'))[0]  # model.yaml path
+        model = Model(cfg, channels, classes)
+        if pretrained:
+            fname = f'{name}.pt'  # checkpoint filename
+            attempt_download(fname)  # download if not found locally
+            ckpt = torch.load(fname, map_location=torch.device('cpu'))  # load
+            msd = model.state_dict()  # model state_dict
+            csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+            csd = {k: v for k, v in csd.items() if msd[k].shape == v.shape}  # filter
+            model.load_state_dict(csd, strict=False)  # load
+            if len(ckpt['model'].names) == classes:
+                model.names = ckpt['model'].names  # set class names attribute
+            if autoshape:
+                model = model.autoshape()  # for file/URI/PIL/cv2/np inputs and NMS
+        device = select_device('0' if torch.cuda.is_available() else 'cpu')  # default to GPU if available
+        return model.to(device)
+
+    except Exception as e:
+        s = 'Cache maybe be out of date, try force_reload=True.'
+        raise Exception(s) from e
+
+
+def custom(path_or_model='path/to/model.pt', autoshape=True):
+    """custom mode
+
+    Arguments (3 options):
+        path_or_model (str): 'path/to/model.pt'
+        path_or_model (dict): torch.load('path/to/model.pt')
+        path_or_model (nn.Module): torch.load('path/to/model.pt')['model']
+
+    Returns:
+        pytorch model
+    """
+    model = torch.load(path_or_model, map_location=torch.device('cpu')) if isinstance(path_or_model, str) else path_or_model  # load checkpoint
+    if isinstance(model, dict):
+        model = model['ema' if model.get('ema') else 'model']  # load model
+
+    hub_model = Model(model.yaml).to(next(model.parameters()).device)  # create
+    hub_model.load_state_dict(model.float().state_dict())  # load state_dict
+    hub_model.names = model.names  # class names
+    if autoshape:
+        hub_model = hub_model.autoshape()  # for file/URI/PIL/cv2/np inputs and NMS
+    device = select_device('0' if torch.cuda.is_available() else 'cpu')  # default to GPU if available
+    return hub_model.to(device)
+
+
+def yolov7(pretrained=True, channels=3, classes=80, autoshape=True):
+    return create('yolov7', pretrained, channels, classes, autoshape)
+
+
+if __name__ == '__main__':
+    model = custom(path_or_model='yolov7.pt')  # custom example
+    # model = create(name='yolov7', pretrained=True, channels=3, classes=80, autoshape=True)  # pretrained example
+
+    # Verify inference
+    import numpy as np
+    from PIL import Image
+
+    imgs = [np.zeros((640, 480, 3))]
+
+    results = model(imgs)  # batched inference
+    results.print()
+    results.save()

models/__init__.py

@@ -0,0 +1 @@
+# init

models/common.py

@@ -0,0 +1,2019 @@
+import math
+from copy import copy
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+import requests
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.ops import DeformConv2d
+from PIL import Image
+from torch.cuda import amp
+
+from utils.datasets import letterbox
+from utils.general import non_max_suppression, make_divisible, scale_coords, increment_path, xyxy2xywh
+from utils.plots import color_list, plot_one_box
+from utils.torch_utils import time_synchronized
+
+
+##### basic ####
+
+def autopad(k, p=None):  # kernel, padding
+    # Pad to 'same'
+    if p is None:
+        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
+    return p
+
+
+class MP(nn.Module):
+    def __init__(self, k=2):
+        super(MP, self).__init__()
+        self.m = nn.MaxPool2d(kernel_size=k, stride=k)
+
+    def forward(self, x):
+        return self.m(x)
+
+
+class SP(nn.Module):
+    def __init__(self, k=3, s=1):
+        super(SP, self).__init__()
+        self.m = nn.MaxPool2d(kernel_size=k, stride=s, padding=k // 2)
+
+    def forward(self, x):
+        return self.m(x)
+    
+    
+class ReOrg(nn.Module):
+    def __init__(self):
+        super(ReOrg, self).__init__()
+
+    def forward(self, x):  # x(b,c,w,h) -> y(b,4c,w/2,h/2)
+        return torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)
+
+
+class Concat(nn.Module):
+    def __init__(self, dimension=1):
+        super(Concat, self).__init__()
+        self.d = dimension
+
+    def forward(self, x):
+        return torch.cat(x, self.d)
+
+
+class Chuncat(nn.Module):
+    def __init__(self, dimension=1):
+        super(Chuncat, self).__init__()
+        self.d = dimension
+
+    def forward(self, x):
+        x1 = []
+        x2 = []
+        for xi in x:
+            xi1, xi2 = xi.chunk(2, self.d)
+            x1.append(xi1)
+            x2.append(xi2)
+        return torch.cat(x1+x2, self.d)
+
+
+class Shortcut(nn.Module):
+    def __init__(self, dimension=0):
+        super(Shortcut, self).__init__()
+        self.d = dimension
+
+    def forward(self, x):
+        return x[0]+x[1]
+
+
+class Foldcut(nn.Module):
+    def __init__(self, dimension=0):
+        super(Foldcut, self).__init__()
+        self.d = dimension
+
+    def forward(self, x):
+        x1, x2 = x.chunk(2, self.d)
+        return x1+x2
+
+
+class Conv(nn.Module):
+    # Standard convolution
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
+        super(Conv, self).__init__()
+        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)
+        self.bn = nn.BatchNorm2d(c2)
+        self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
+
+    def forward(self, x):
+        return self.act(self.bn(self.conv(x)))
+
+    def fuseforward(self, x):
+        return self.act(self.conv(x))
+    
+
+class RobustConv(nn.Module):
+    # Robust convolution (use high kernel size 7-11 for: downsampling and other layers). Train for 300 - 450 epochs.
+    def __init__(self, c1, c2, k=7, s=1, p=None, g=1, act=True, layer_scale_init_value=1e-6):  # ch_in, ch_out, kernel, stride, padding, groups
+        super(RobustConv, self).__init__()
+        self.conv_dw = Conv(c1, c1, k=k, s=s, p=p, g=c1, act=act)
+        self.conv1x1 = nn.Conv2d(c1, c2, 1, 1, 0, groups=1, bias=True)
+        self.gamma = nn.Parameter(layer_scale_init_value * torch.ones(c2)) if layer_scale_init_value > 0 else None
+
+    def forward(self, x):
+        x = x.to(memory_format=torch.channels_last)
+        x = self.conv1x1(self.conv_dw(x))
+        if self.gamma is not None:
+            x = x.mul(self.gamma.reshape(1, -1, 1, 1)) 
+        return x
+
+
+class RobustConv2(nn.Module):
+    # Robust convolution 2 (use [32, 5, 2] or [32, 7, 4] or [32, 11, 8] for one of the paths in CSP).
+    def __init__(self, c1, c2, k=7, s=4, p=None, g=1, act=True, layer_scale_init_value=1e-6):  # ch_in, ch_out, kernel, stride, padding, groups
+        super(RobustConv2, self).__init__()
+        self.conv_strided = Conv(c1, c1, k=k, s=s, p=p, g=c1, act=act)
+        self.conv_deconv = nn.ConvTranspose2d(in_channels=c1, out_channels=c2, kernel_size=s, stride=s, 
+                                              padding=0, bias=True, dilation=1, groups=1
+        )
+        self.gamma = nn.Parameter(layer_scale_init_value * torch.ones(c2)) if layer_scale_init_value > 0 else None
+
+    def forward(self, x):
+        x = self.conv_deconv(self.conv_strided(x))
+        if self.gamma is not None:
+            x = x.mul(self.gamma.reshape(1, -1, 1, 1)) 
+        return x
+    
+
+def DWConv(c1, c2, k=1, s=1, act=True):
+    # Depthwise convolution
+    return Conv(c1, c2, k, s, g=math.gcd(c1, c2), act=act)
+
+
+class GhostConv(nn.Module):
+    # Ghost Convolution https://github.com/huawei-noah/ghostnet
+    def __init__(self, c1, c2, k=1, s=1, g=1, act=True):  # ch_in, ch_out, kernel, stride, groups
+        super(GhostConv, self).__init__()
+        c_ = c2 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, k, s, None, g, act)
+        self.cv2 = Conv(c_, c_, 5, 1, None, c_, act)
+
+    def forward(self, x):
+        y = self.cv1(x)
+        return torch.cat([y, self.cv2(y)], 1)
+
+
+class Stem(nn.Module):
+    # Stem
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
+        super(Stem, self).__init__()
+        c_ = int(c2/2)  # hidden channels
+        self.cv1 = Conv(c1, c_, 3, 2)
+        self.cv2 = Conv(c_, c_, 1, 1)
+        self.cv3 = Conv(c_, c_, 3, 2)
+        self.pool = torch.nn.MaxPool2d(2, stride=2)
+        self.cv4 = Conv(2 * c_, c2, 1, 1)
+
+    def forward(self, x):
+        x = self.cv1(x)
+        return self.cv4(torch.cat((self.cv3(self.cv2(x)), self.pool(x)), dim=1))
+
+
+class DownC(nn.Module):
+    # Spatial pyramid pooling layer used in YOLOv3-SPP
+    def __init__(self, c1, c2, n=1, k=2):
+        super(DownC, self).__init__()
+        c_ = int(c1)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_, c2//2, 3, k)
+        self.cv3 = Conv(c1, c2//2, 1, 1)
+        self.mp = nn.MaxPool2d(kernel_size=k, stride=k)
+
+    def forward(self, x):
+        return torch.cat((self.cv2(self.cv1(x)), self.cv3(self.mp(x))), dim=1)
+
+
+class SPP(nn.Module):
+    # Spatial pyramid pooling layer used in YOLOv3-SPP
+    def __init__(self, c1, c2, k=(5, 9, 13)):
+        super(SPP, self).__init__()
+        c_ = c1 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1)
+        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
+
+    def forward(self, x):
+        x = self.cv1(x)
+        return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))
+    
+
+class Bottleneck(nn.Module):
+    # Darknet bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
+        super(Bottleneck, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_, c2, 3, 1, g=g)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
+
+
+class Res(nn.Module):
+    # ResNet bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
+        super(Res, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_, c_, 3, 1, g=g)
+        self.cv3 = Conv(c_, c2, 1, 1)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv3(self.cv2(self.cv1(x))) if self.add else self.cv3(self.cv2(self.cv1(x)))
+
+
+class ResX(Res):
+    # ResNet bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=32, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
+        super().__init__(c1, c2, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+
+
+class Ghost(nn.Module):
+    # Ghost Bottleneck https://github.com/huawei-noah/ghostnet
+    def __init__(self, c1, c2, k=3, s=1):  # ch_in, ch_out, kernel, stride
+        super(Ghost, self).__init__()
+        c_ = c2 // 2
+        self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1),  # pw
+                                  DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(),  # dw
+                                  GhostConv(c_, c2, 1, 1, act=False))  # pw-linear
+        self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False),
+                                      Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity()
+
+    def forward(self, x):
+        return self.conv(x) + self.shortcut(x)
+
+##### end of basic #####
+
+
+##### cspnet #####
+
+class SPPCSPC(nn.Module):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, k=(5, 9, 13)):
+        super(SPPCSPC, self).__init__()
+        c_ = int(2 * c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(c_, c_, 3, 1)
+        self.cv4 = Conv(c_, c_, 1, 1)
+        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
+        self.cv5 = Conv(4 * c_, c_, 1, 1)
+        self.cv6 = Conv(c_, c_, 3, 1)
+        self.cv7 = Conv(2 * c_, c2, 1, 1)
+
+    def forward(self, x):
+        x1 = self.cv4(self.cv3(self.cv1(x)))
+        y1 = self.cv6(self.cv5(torch.cat([x1] + [m(x1) for m in self.m], 1)))
+        y2 = self.cv2(x)
+        return self.cv7(torch.cat((y1, y2), dim=1))
+
+class GhostSPPCSPC(SPPCSPC):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, k=(5, 9, 13)):
+        super().__init__(c1, c2, n, shortcut, g, e, k)
+        c_ = int(2 * c2 * e)  # hidden channels
+        self.cv1 = GhostConv(c1, c_, 1, 1)
+        self.cv2 = GhostConv(c1, c_, 1, 1)
+        self.cv3 = GhostConv(c_, c_, 3, 1)
+        self.cv4 = GhostConv(c_, c_, 1, 1)
+        self.cv5 = GhostConv(4 * c_, c_, 1, 1)
+        self.cv6 = GhostConv(c_, c_, 3, 1)
+        self.cv7 = GhostConv(2 * c_, c2, 1, 1)
+
+
+class GhostStem(Stem):
+    # Stem
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__(c1, c2, k, s, p, g, act)
+        c_ = int(c2/2)  # hidden channels
+        self.cv1 = GhostConv(c1, c_, 3, 2)
+        self.cv2 = GhostConv(c_, c_, 1, 1)
+        self.cv3 = GhostConv(c_, c_, 3, 2)
+        self.cv4 = GhostConv(2 * c_, c2, 1, 1)
+        
+
+class BottleneckCSPA(nn.Module):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super(BottleneckCSPA, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(2 * c_, c2, 1, 1)
+        self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+    def forward(self, x):
+        y1 = self.m(self.cv1(x))
+        y2 = self.cv2(x)
+        return self.cv3(torch.cat((y1, y2), dim=1))
+
+
+class BottleneckCSPB(nn.Module):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super(BottleneckCSPB, self).__init__()
+        c_ = int(c2)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_, c_, 1, 1)
+        self.cv3 = Conv(2 * c_, c2, 1, 1)
+        self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+    def forward(self, x):
+        x1 = self.cv1(x)
+        y1 = self.m(x1)
+        y2 = self.cv2(x1)
+        return self.cv3(torch.cat((y1, y2), dim=1))
+
+
+class BottleneckCSPC(nn.Module):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super(BottleneckCSPC, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(c_, c_, 1, 1)
+        self.cv4 = Conv(2 * c_, c2, 1, 1)
+        self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+    def forward(self, x):
+        y1 = self.cv3(self.m(self.cv1(x)))
+        y2 = self.cv2(x)
+        return self.cv4(torch.cat((y1, y2), dim=1))
+
+
+class ResCSPA(BottleneckCSPA):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
+
+
+class ResCSPB(BottleneckCSPB):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2)  # hidden channels
+        self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
+
+
+class ResCSPC(BottleneckCSPC):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
+
+
+class ResXCSPA(ResCSPA):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=32, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+
+class ResXCSPB(ResCSPB):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=32, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2)  # hidden channels
+        self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+
+class ResXCSPC(ResCSPC):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=32, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+
+class GhostCSPA(BottleneckCSPA):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[Ghost(c_, c_) for _ in range(n)])
+
+
+class GhostCSPB(BottleneckCSPB):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2)  # hidden channels
+        self.m = nn.Sequential(*[Ghost(c_, c_) for _ in range(n)])
+
+
+class GhostCSPC(BottleneckCSPC):
+    # CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[Ghost(c_, c_) for _ in range(n)])
+
+##### end of cspnet #####
+
+
+##### yolor #####
+
+class ImplicitA(nn.Module):
+    def __init__(self, channel, mean=0., std=.02):
+        super(ImplicitA, self).__init__()
+        self.channel = channel
+        self.mean = mean
+        self.std = std
+        self.implicit = nn.Parameter(torch.zeros(1, channel, 1, 1))
+        nn.init.normal_(self.implicit, mean=self.mean, std=self.std)
+
+    def forward(self, x):
+        return self.implicit + x
+    
+
+class ImplicitM(nn.Module):
+    def __init__(self, channel, mean=1., std=.02):
+        super(ImplicitM, self).__init__()
+        self.channel = channel
+        self.mean = mean
+        self.std = std
+        self.implicit = nn.Parameter(torch.ones(1, channel, 1, 1))
+        nn.init.normal_(self.implicit, mean=self.mean, std=self.std)
+
+    def forward(self, x):
+        return self.implicit * x
+    
+##### end of yolor #####
+
+
+##### repvgg #####
+
+class RepConv(nn.Module):
+    # Represented convolution
+    # https://arxiv.org/abs/2101.03697
+
+    def __init__(self, c1, c2, k=3, s=1, p=None, g=1, act=True, deploy=False):
+        super(RepConv, self).__init__()
+
+        self.deploy = deploy
+        self.groups = g
+        self.in_channels = c1
+        self.out_channels = c2
+
+        assert k == 3
+        assert autopad(k, p) == 1
+
+        padding_11 = autopad(k, p) - k // 2
+
+        self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
+
+        if deploy:
+            self.rbr_reparam = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=True)
+
+        else:
+            self.rbr_identity = (nn.BatchNorm2d(num_features=c1) if c2 == c1 and s == 1 else None)
+
+            self.rbr_dense = nn.Sequential(
+                nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False),
+                nn.BatchNorm2d(num_features=c2),
+            )
+
+            self.rbr_1x1 = nn.Sequential(
+                nn.Conv2d( c1, c2, 1, s, padding_11, groups=g, bias=False),
+                nn.BatchNorm2d(num_features=c2),
+            )
+
+    def forward(self, inputs):
+        if hasattr(self, "rbr_reparam"):
+            return self.act(self.rbr_reparam(inputs))
+
+        if self.rbr_identity is None:
+            id_out = 0
+        else:
+            id_out = self.rbr_identity(inputs)
+
+        return self.act(self.rbr_dense(inputs) + self.rbr_1x1(inputs) + id_out)
+    
+    def get_equivalent_kernel_bias(self):
+        kernel3x3, bias3x3 = self._fuse_bn_tensor(self.rbr_dense)
+        kernel1x1, bias1x1 = self._fuse_bn_tensor(self.rbr_1x1)
+        kernelid, biasid = self._fuse_bn_tensor(self.rbr_identity)
+        return (
+            kernel3x3 + self._pad_1x1_to_3x3_tensor(kernel1x1) + kernelid,
+            bias3x3 + bias1x1 + biasid,
+        )
+
+    def _pad_1x1_to_3x3_tensor(self, kernel1x1):
+        if kernel1x1 is None:
+            return 0
+        else:
+            return nn.functional.pad(kernel1x1, [1, 1, 1, 1])
+
+    def _fuse_bn_tensor(self, branch):
+        if branch is None:
+            return 0, 0
+        if isinstance(branch, nn.Sequential):
+            kernel = branch[0].weight
+            running_mean = branch[1].running_mean
+            running_var = branch[1].running_var
+            gamma = branch[1].weight
+            beta = branch[1].bias
+            eps = branch[1].eps
+        else:
+            assert isinstance(branch, nn.BatchNorm2d)
+            if not hasattr(self, "id_tensor"):
+                input_dim = self.in_channels // self.groups
+                kernel_value = np.zeros(
+                    (self.in_channels, input_dim, 3, 3), dtype=np.float32
+                )
+                for i in range(self.in_channels):
+                    kernel_value[i, i % input_dim, 1, 1] = 1
+                self.id_tensor = torch.from_numpy(kernel_value).to(branch.weight.device)
+            kernel = self.id_tensor
+            running_mean = branch.running_mean
+            running_var = branch.running_var
+            gamma = branch.weight
+            beta = branch.bias
+            eps = branch.eps
+        std = (running_var + eps).sqrt()
+        t = (gamma / std).reshape(-1, 1, 1, 1)
+        return kernel * t, beta - running_mean * gamma / std
+
+    def repvgg_convert(self):
+        kernel, bias = self.get_equivalent_kernel_bias()
+        return (
+            kernel.detach().cpu().numpy(),
+            bias.detach().cpu().numpy(),
+        )
+
+    def fuse_conv_bn(self, conv, bn):
+
+        std = (bn.running_var + bn.eps).sqrt()
+        bias = bn.bias - bn.running_mean * bn.weight / std
+
+        t = (bn.weight / std).reshape(-1, 1, 1, 1)
+        weights = conv.weight * t
+
+        bn = nn.Identity()
+        conv = nn.Conv2d(in_channels = conv.in_channels,
+                              out_channels = conv.out_channels,
+                              kernel_size = conv.kernel_size,
+                              stride=conv.stride,
+                              padding = conv.padding,
+                              dilation = conv.dilation,
+                              groups = conv.groups,
+                              bias = True,
+                              padding_mode = conv.padding_mode)
+
+        conv.weight = torch.nn.Parameter(weights)
+        conv.bias = torch.nn.Parameter(bias)
+        return conv
+
+    def fuse_repvgg_block(self):    
+        if self.deploy:
+            return
+        print(f"RepConv.fuse_repvgg_block")
+                
+        self.rbr_dense = self.fuse_conv_bn(self.rbr_dense[0], self.rbr_dense[1])
+        
+        self.rbr_1x1 = self.fuse_conv_bn(self.rbr_1x1[0], self.rbr_1x1[1])
+        rbr_1x1_bias = self.rbr_1x1.bias
+        weight_1x1_expanded = torch.nn.functional.pad(self.rbr_1x1.weight, [1, 1, 1, 1])
+        
+        # Fuse self.rbr_identity
+        if (isinstance(self.rbr_identity, nn.BatchNorm2d) or isinstance(self.rbr_identity, nn.modules.batchnorm.SyncBatchNorm)):
+            # print(f"fuse: rbr_identity == BatchNorm2d or SyncBatchNorm")
+            identity_conv_1x1 = nn.Conv2d(
+                    in_channels=self.in_channels,
+                    out_channels=self.out_channels,
+                    kernel_size=1,
+                    stride=1,
+                    padding=0,
+                    groups=self.groups, 
+                    bias=False)
+            identity_conv_1x1.weight.data = identity_conv_1x1.weight.data.to(self.rbr_1x1.weight.data.device)
+            identity_conv_1x1.weight.data = identity_conv_1x1.weight.data.squeeze().squeeze()
+            # print(f" identity_conv_1x1.weight = {identity_conv_1x1.weight.shape}")
+            identity_conv_1x1.weight.data.fill_(0.0)
+            identity_conv_1x1.weight.data.fill_diagonal_(1.0)
+            identity_conv_1x1.weight.data = identity_conv_1x1.weight.data.unsqueeze(2).unsqueeze(3)
+            # print(f" identity_conv_1x1.weight = {identity_conv_1x1.weight.shape}")
+
+            identity_conv_1x1 = self.fuse_conv_bn(identity_conv_1x1, self.rbr_identity)
+            bias_identity_expanded = identity_conv_1x1.bias
+            weight_identity_expanded = torch.nn.functional.pad(identity_conv_1x1.weight, [1, 1, 1, 1])            
+        else:
+            # print(f"fuse: rbr_identity != BatchNorm2d, rbr_identity = {self.rbr_identity}")
+            bias_identity_expanded = torch.nn.Parameter( torch.zeros_like(rbr_1x1_bias) )
+            weight_identity_expanded = torch.nn.Parameter( torch.zeros_like(weight_1x1_expanded) )            
+        
+
+        #print(f"self.rbr_1x1.weight = {self.rbr_1x1.weight.shape}, ")
+        #print(f"weight_1x1_expanded = {weight_1x1_expanded.shape}, ")
+        #print(f"self.rbr_dense.weight = {self.rbr_dense.weight.shape}, ")
+
+        self.rbr_dense.weight = torch.nn.Parameter(self.rbr_dense.weight + weight_1x1_expanded + weight_identity_expanded)
+        self.rbr_dense.bias = torch.nn.Parameter(self.rbr_dense.bias + rbr_1x1_bias + bias_identity_expanded)
+                
+        self.rbr_reparam = self.rbr_dense
+        self.deploy = True
+
+        if self.rbr_identity is not None:
+            del self.rbr_identity
+            self.rbr_identity = None
+
+        if self.rbr_1x1 is not None:
+            del self.rbr_1x1
+            self.rbr_1x1 = None
+
+        if self.rbr_dense is not None:
+            del self.rbr_dense
+            self.rbr_dense = None
+
+
+class RepBottleneck(Bottleneck):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
+        super().__init__(c1, c2, shortcut=True, g=1, e=0.5)
+        c_ = int(c2 * e)  # hidden channels
+        self.cv2 = RepConv(c_, c2, 3, 1, g=g)
+
+
+class RepBottleneckCSPA(BottleneckCSPA):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[RepBottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+
+class RepBottleneckCSPB(BottleneckCSPB):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2)  # hidden channels
+        self.m = nn.Sequential(*[RepBottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+
+class RepBottleneckCSPC(BottleneckCSPC):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[RepBottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+
+class RepRes(Res):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
+        super().__init__(c1, c2, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.cv2 = RepConv(c_, c_, 3, 1, g=g)
+
+
+class RepResCSPA(ResCSPA):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[RepRes(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
+
+
+class RepResCSPB(ResCSPB):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2)  # hidden channels
+        self.m = nn.Sequential(*[RepRes(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
+
+
+class RepResCSPC(ResCSPC):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[RepRes(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
+
+
+class RepResX(ResX):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=32, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
+        super().__init__(c1, c2, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.cv2 = RepConv(c_, c_, 3, 1, g=g)
+
+
+class RepResXCSPA(ResXCSPA):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=32, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[RepResX(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
+
+
+class RepResXCSPB(ResXCSPB):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=False, g=32, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2)  # hidden channels
+        self.m = nn.Sequential(*[RepResX(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
+
+
+class RepResXCSPC(ResXCSPC):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=32, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*[RepResX(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
+
+##### end of repvgg #####
+
+
+##### transformer #####
+
+class TransformerLayer(nn.Module):
+    # Transformer layer https://arxiv.org/abs/2010.11929 (LayerNorm layers removed for better performance)
+    def __init__(self, c, num_heads):
+        super().__init__()
+        self.q = nn.Linear(c, c, bias=False)
+        self.k = nn.Linear(c, c, bias=False)
+        self.v = nn.Linear(c, c, bias=False)
+        self.ma = nn.MultiheadAttention(embed_dim=c, num_heads=num_heads)
+        self.fc1 = nn.Linear(c, c, bias=False)
+        self.fc2 = nn.Linear(c, c, bias=False)
+
+    def forward(self, x):
+        x = self.ma(self.q(x), self.k(x), self.v(x))[0] + x
+        x = self.fc2(self.fc1(x)) + x
+        return x
+
+
+class TransformerBlock(nn.Module):
+    # Vision Transformer https://arxiv.org/abs/2010.11929
+    def __init__(self, c1, c2, num_heads, num_layers):
+        super().__init__()
+        self.conv = None
+        if c1 != c2:
+            self.conv = Conv(c1, c2)
+        self.linear = nn.Linear(c2, c2)  # learnable position embedding
+        self.tr = nn.Sequential(*[TransformerLayer(c2, num_heads) for _ in range(num_layers)])
+        self.c2 = c2
+
+    def forward(self, x):
+        if self.conv is not None:
+            x = self.conv(x)
+        b, _, w, h = x.shape
+        p = x.flatten(2)
+        p = p.unsqueeze(0)
+        p = p.transpose(0, 3)
+        p = p.squeeze(3)
+        e = self.linear(p)
+        x = p + e
+
+        x = self.tr(x)
+        x = x.unsqueeze(3)
+        x = x.transpose(0, 3)
+        x = x.reshape(b, self.c2, w, h)
+        return x
+
+##### end of transformer #####
+
+
+##### yolov5 #####
+
+class Focus(nn.Module):
+    # Focus wh information into c-space
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
+        super(Focus, self).__init__()
+        self.conv = Conv(c1 * 4, c2, k, s, p, g, act)
+        # self.contract = Contract(gain=2)
+
+    def forward(self, x):  # x(b,c,w,h) -> y(b,4c,w/2,h/2)
+        return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))
+        # return self.conv(self.contract(x))
+        
+
+class SPPF(nn.Module):
+    # Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
+    def __init__(self, c1, c2, k=5):  # equivalent to SPP(k=(5, 9, 13))
+        super().__init__()
+        c_ = c1 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_ * 4, c2, 1, 1)
+        self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)
+
+    def forward(self, x):
+        x = self.cv1(x)
+        y1 = self.m(x)
+        y2 = self.m(y1)
+        return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1))
+    
+    
+class Contract(nn.Module):
+    # Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40)
+    def __init__(self, gain=2):
+        super().__init__()
+        self.gain = gain
+
+    def forward(self, x):
+        N, C, H, W = x.size()  # assert (H / s == 0) and (W / s == 0), 'Indivisible gain'
+        s = self.gain
+        x = x.view(N, C, H // s, s, W // s, s)  # x(1,64,40,2,40,2)
+        x = x.permute(0, 3, 5, 1, 2, 4).contiguous()  # x(1,2,2,64,40,40)
+        return x.view(N, C * s * s, H // s, W // s)  # x(1,256,40,40)
+
+
+class Expand(nn.Module):
+    # Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160)
+    def __init__(self, gain=2):
+        super().__init__()
+        self.gain = gain
+
+    def forward(self, x):
+        N, C, H, W = x.size()  # assert C / s ** 2 == 0, 'Indivisible gain'
+        s = self.gain
+        x = x.view(N, s, s, C // s ** 2, H, W)  # x(1,2,2,16,80,80)
+        x = x.permute(0, 3, 4, 1, 5, 2).contiguous()  # x(1,16,80,2,80,2)
+        return x.view(N, C // s ** 2, H * s, W * s)  # x(1,16,160,160)
+
+
+class NMS(nn.Module):
+    # Non-Maximum Suppression (NMS) module
+    conf = 0.25  # confidence threshold
+    iou = 0.45  # IoU threshold
+    classes = None  # (optional list) filter by class
+
+    def __init__(self):
+        super(NMS, self).__init__()
+
+    def forward(self, x):
+        return non_max_suppression(x[0], conf_thres=self.conf, iou_thres=self.iou, classes=self.classes)
+
+
+class autoShape(nn.Module):
+    # input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS
+    conf = 0.25  # NMS confidence threshold
+    iou = 0.45  # NMS IoU threshold
+    classes = None  # (optional list) filter by class
+
+    def __init__(self, model):
+        super(autoShape, self).__init__()
+        self.model = model.eval()
+
+    def autoshape(self):
+        print('autoShape already enabled, skipping... ')  # model already converted to model.autoshape()
+        return self
+
+    @torch.no_grad()
+    def forward(self, imgs, size=640, augment=False, profile=False):
+        # Inference from various sources. For height=640, width=1280, RGB images example inputs are:
+        #   filename:   imgs = 'data/samples/zidane.jpg'
+        #   URI:             = 'https://github.com/ultralytics/yolov5/releases/download/v1.0/zidane.jpg'
+        #   OpenCV:          = cv2.imread('image.jpg')[:,:,::-1]  # HWC BGR to RGB x(640,1280,3)
+        #   PIL:             = Image.open('image.jpg')  # HWC x(640,1280,3)
+        #   numpy:           = np.zeros((640,1280,3))  # HWC
+        #   torch:           = torch.zeros(16,3,320,640)  # BCHW (scaled to size=640, 0-1 values)
+        #   multiple:        = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...]  # list of images
+
+        t = [time_synchronized()]
+        p = next(self.model.parameters())  # for device and type
+        if isinstance(imgs, torch.Tensor):  # torch
+            with amp.autocast(enabled=p.device.type != 'cpu'):
+                return self.model(imgs.to(p.device).type_as(p), augment, profile)  # inference
+
+        # Pre-process
+        n, imgs = (len(imgs), imgs) if isinstance(imgs, list) else (1, [imgs])  # number of images, list of images
+        shape0, shape1, files = [], [], []  # image and inference shapes, filenames
+        for i, im in enumerate(imgs):
+            f = f'image{i}'  # filename
+            if isinstance(im, str):  # filename or uri
+                im, f = np.asarray(Image.open(requests.get(im, stream=True).raw if im.startswith('http') else im)), im
+            elif isinstance(im, Image.Image):  # PIL Image
+                im, f = np.asarray(im), getattr(im, 'filename', f) or f
+            files.append(Path(f).with_suffix('.jpg').name)
+            if im.shape[0] < 5:  # image in CHW
+                im = im.transpose((1, 2, 0))  # reverse dataloader .transpose(2, 0, 1)
+            im = im[:, :, :3] if im.ndim == 3 else np.tile(im[:, :, None], 3)  # enforce 3ch input
+            s = im.shape[:2]  # HWC
+            shape0.append(s)  # image shape
+            g = (size / max(s))  # gain
+            shape1.append([y * g for y in s])
+            imgs[i] = im  # update
+        shape1 = [make_divisible(x, int(self.stride.max())) for x in np.stack(shape1, 0).max(0)]  # inference shape
+        x = [letterbox(im, new_shape=shape1, auto=False)[0] for im in imgs]  # pad
+        x = np.stack(x, 0) if n > 1 else x[0][None]  # stack
+        x = np.ascontiguousarray(x.transpose((0, 3, 1, 2)))  # BHWC to BCHW
+        x = torch.from_numpy(x).to(p.device).type_as(p) / 255.  # uint8 to fp16/32
+        t.append(time_synchronized())
+
+        with amp.autocast(enabled=p.device.type != 'cpu'):
+            # Inference
+            y = self.model(x, augment, profile)[0]  # forward
+            t.append(time_synchronized())
+
+            # Post-process
+            y = non_max_suppression(y, conf_thres=self.conf, iou_thres=self.iou, classes=self.classes)  # NMS
+            for i in range(n):
+                scale_coords(shape1, y[i][:, :4], shape0[i])
+
+            t.append(time_synchronized())
+            return Detections(imgs, y, files, t, self.names, x.shape)
+
+
+class Detections:
+    # detections class for YOLOv5 inference results
+    def __init__(self, imgs, pred, files, times=None, names=None, shape=None):
+        super(Detections, self).__init__()
+        d = pred[0].device  # device
+        gn = [torch.tensor([*[im.shape[i] for i in [1, 0, 1, 0]], 1., 1.], device=d) for im in imgs]  # normalizations
+        self.imgs = imgs  # list of images as numpy arrays
+        self.pred = pred  # list of tensors pred[0] = (xyxy, conf, cls)
+        self.names = names  # class names
+        self.files = files  # image filenames
+        self.xyxy = pred  # xyxy pixels
+        self.xywh = [xyxy2xywh(x) for x in pred]  # xywh pixels
+        self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)]  # xyxy normalized
+        self.xywhn = [x / g for x, g in zip(self.xywh, gn)]  # xywh normalized
+        self.n = len(self.pred)  # number of images (batch size)
+        self.t = tuple((times[i + 1] - times[i]) * 1000 / self.n for i in range(3))  # timestamps (ms)
+        self.s = shape  # inference BCHW shape
+
+    def display(self, pprint=False, show=False, save=False, render=False, save_dir=''):
+        colors = color_list()
+        for i, (img, pred) in enumerate(zip(self.imgs, self.pred)):
+            str = f'image {i + 1}/{len(self.pred)}: {img.shape[0]}x{img.shape[1]} '
+            if pred is not None:
+                for c in pred[:, -1].unique():
+                    n = (pred[:, -1] == c).sum()  # detections per class
+                    str += f"{n} {self.names[int(c)]}{'s' * (n > 1)}, "  # add to string
+                if show or save or render:
+                    for *box, conf, cls in pred:  # xyxy, confidence, class
+                        label = f'{self.names[int(cls)]} {conf:.2f}'
+                        plot_one_box(box, img, label=label, color=colors[int(cls) % 10])
+            img = Image.fromarray(img.astype(np.uint8)) if isinstance(img, np.ndarray) else img  # from np
+            if pprint:
+                print(str.rstrip(', '))
+            if show:
+                img.show(self.files[i])  # show
+            if save:
+                f = self.files[i]
+                img.save(Path(save_dir) / f)  # save
+                print(f"{'Saved' * (i == 0)} {f}", end=',' if i < self.n - 1 else f' to {save_dir}\n')
+            if render:
+                self.imgs[i] = np.asarray(img)
+
+    def print(self):
+        self.display(pprint=True)  # print results
+        print(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {tuple(self.s)}' % self.t)
+
+    def show(self):
+        self.display(show=True)  # show results
+
+    def save(self, save_dir='runs/hub/exp'):
+        save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/hub/exp')  # increment save_dir
+        Path(save_dir).mkdir(parents=True, exist_ok=True)
+        self.display(save=True, save_dir=save_dir)  # save results
+
+    def render(self):
+        self.display(render=True)  # render results
+        return self.imgs
+
+    def pandas(self):
+        # return detections as pandas DataFrames, i.e. print(results.pandas().xyxy[0])
+        new = copy(self)  # return copy
+        ca = 'xmin', 'ymin', 'xmax', 'ymax', 'confidence', 'class', 'name'  # xyxy columns
+        cb = 'xcenter', 'ycenter', 'width', 'height', 'confidence', 'class', 'name'  # xywh columns
+        for k, c in zip(['xyxy', 'xyxyn', 'xywh', 'xywhn'], [ca, ca, cb, cb]):
+            a = [[x[:5] + [int(x[5]), self.names[int(x[5])]] for x in x.tolist()] for x in getattr(self, k)]  # update
+            setattr(new, k, [pd.DataFrame(x, columns=c) for x in a])
+        return new
+
+    def tolist(self):
+        # return a list of Detections objects, i.e. 'for result in results.tolist():'
+        x = [Detections([self.imgs[i]], [self.pred[i]], self.names, self.s) for i in range(self.n)]
+        for d in x:
+            for k in ['imgs', 'pred', 'xyxy', 'xyxyn', 'xywh', 'xywhn']:
+                setattr(d, k, getattr(d, k)[0])  # pop out of list
+        return x
+
+    def __len__(self):
+        return self.n
+
+
+class Classify(nn.Module):
+    # Classification head, i.e. x(b,c1,20,20) to x(b,c2)
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1):  # ch_in, ch_out, kernel, stride, padding, groups
+        super(Classify, self).__init__()
+        self.aap = nn.AdaptiveAvgPool2d(1)  # to x(b,c1,1,1)
+        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g)  # to x(b,c2,1,1)
+        self.flat = nn.Flatten()
+
+    def forward(self, x):
+        z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1)  # cat if list
+        return self.flat(self.conv(z))  # flatten to x(b,c2)
+
+##### end of yolov5 ######
+
+
+##### orepa #####
+
+def transI_fusebn(kernel, bn):
+    gamma = bn.weight
+    std = (bn.running_var + bn.eps).sqrt()
+    return kernel * ((gamma / std).reshape(-1, 1, 1, 1)), bn.bias - bn.running_mean * gamma / std
+    
+    
+class ConvBN(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size,
+                             stride=1, padding=0, dilation=1, groups=1, deploy=False, nonlinear=None):
+        super().__init__()
+        if nonlinear is None:
+            self.nonlinear = nn.Identity()
+        else:
+            self.nonlinear = nonlinear
+        if deploy:
+            self.conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size,
+                                      stride=stride, padding=padding, dilation=dilation, groups=groups, bias=True)
+        else:
+            self.conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size,
+                                            stride=stride, padding=padding, dilation=dilation, groups=groups, bias=False)
+            self.bn = nn.BatchNorm2d(num_features=out_channels)
+
+    def forward(self, x):
+        if hasattr(self, 'bn'):
+            return self.nonlinear(self.bn(self.conv(x)))
+        else:
+            return self.nonlinear(self.conv(x))
+
+    def switch_to_deploy(self):
+        kernel, bias = transI_fusebn(self.conv.weight, self.bn)
+        conv = nn.Conv2d(in_channels=self.conv.in_channels, out_channels=self.conv.out_channels, kernel_size=self.conv.kernel_size,
+                                      stride=self.conv.stride, padding=self.conv.padding, dilation=self.conv.dilation, groups=self.conv.groups, bias=True)
+        conv.weight.data = kernel
+        conv.bias.data = bias
+        for para in self.parameters():
+            para.detach_()
+        self.__delattr__('conv')
+        self.__delattr__('bn')
+        self.conv = conv    
+
+class OREPA_3x3_RepConv(nn.Module):
+
+    def __init__(self, in_channels, out_channels, kernel_size,
+                 stride=1, padding=0, dilation=1, groups=1,
+                 internal_channels_1x1_3x3=None,
+                 deploy=False, nonlinear=None, single_init=False):
+        super(OREPA_3x3_RepConv, self).__init__()
+        self.deploy = deploy
+
+        if nonlinear is None:
+            self.nonlinear = nn.Identity()
+        else:
+            self.nonlinear = nonlinear
+
+        self.kernel_size = kernel_size
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.groups = groups
+        assert padding == kernel_size // 2
+
+        self.stride = stride
+        self.padding = padding
+        self.dilation = dilation
+
+        self.branch_counter = 0
+
+        self.weight_rbr_origin = nn.Parameter(torch.Tensor(out_channels, int(in_channels/self.groups), kernel_size, kernel_size))
+        nn.init.kaiming_uniform_(self.weight_rbr_origin, a=math.sqrt(1.0))
+        self.branch_counter += 1
+
+
+        if groups < out_channels:
+            self.weight_rbr_avg_conv = nn.Parameter(torch.Tensor(out_channels, int(in_channels/self.groups), 1, 1))
+            self.weight_rbr_pfir_conv = nn.Parameter(torch.Tensor(out_channels, int(in_channels/self.groups), 1, 1))
+            nn.init.kaiming_uniform_(self.weight_rbr_avg_conv, a=1.0)
+            nn.init.kaiming_uniform_(self.weight_rbr_pfir_conv, a=1.0)
+            self.weight_rbr_avg_conv.data
+            self.weight_rbr_pfir_conv.data
+            self.register_buffer('weight_rbr_avg_avg', torch.ones(kernel_size, kernel_size).mul(1.0/kernel_size/kernel_size))
+            self.branch_counter += 1
+
+        else:
+            raise NotImplementedError
+        self.branch_counter += 1
+
+        if internal_channels_1x1_3x3 is None:
+            internal_channels_1x1_3x3 = in_channels if groups < out_channels else 2 * in_channels   # For mobilenet, it is better to have 2X internal channels
+
+        if internal_channels_1x1_3x3 == in_channels:
+            self.weight_rbr_1x1_kxk_idconv1 = nn.Parameter(torch.zeros(in_channels, int(in_channels/self.groups), 1, 1))
+            id_value = np.zeros((in_channels, int(in_channels/self.groups), 1, 1))
+            for i in range(in_channels):
+                id_value[i, i % int(in_channels/self.groups), 0, 0] = 1
+            id_tensor = torch.from_numpy(id_value).type_as(self.weight_rbr_1x1_kxk_idconv1)
+            self.register_buffer('id_tensor', id_tensor)
+
+        else:
+            self.weight_rbr_1x1_kxk_conv1 = nn.Parameter(torch.Tensor(internal_channels_1x1_3x3, int(in_channels/self.groups), 1, 1))
+            nn.init.kaiming_uniform_(self.weight_rbr_1x1_kxk_conv1, a=math.sqrt(1.0))
+        self.weight_rbr_1x1_kxk_conv2 = nn.Parameter(torch.Tensor(out_channels, int(internal_channels_1x1_3x3/self.groups), kernel_size, kernel_size))
+        nn.init.kaiming_uniform_(self.weight_rbr_1x1_kxk_conv2, a=math.sqrt(1.0))
+        self.branch_counter += 1
+
+        expand_ratio = 8
+        self.weight_rbr_gconv_dw = nn.Parameter(torch.Tensor(in_channels*expand_ratio, 1, kernel_size, kernel_size))
+        self.weight_rbr_gconv_pw = nn.Parameter(torch.Tensor(out_channels, in_channels*expand_ratio, 1, 1))
+        nn.init.kaiming_uniform_(self.weight_rbr_gconv_dw, a=math.sqrt(1.0))
+        nn.init.kaiming_uniform_(self.weight_rbr_gconv_pw, a=math.sqrt(1.0))
+        self.branch_counter += 1
+
+        if out_channels == in_channels and stride == 1:
+            self.branch_counter += 1
+
+        self.vector = nn.Parameter(torch.Tensor(self.branch_counter, self.out_channels))
+        self.bn = nn.BatchNorm2d(out_channels)
+
+        self.fre_init()
+
+        nn.init.constant_(self.vector[0, :], 0.25)    #origin
+        nn.init.constant_(self.vector[1, :], 0.25)      #avg
+        nn.init.constant_(self.vector[2, :], 0.0)      #prior
+        nn.init.constant_(self.vector[3, :], 0.5)    #1x1_kxk
+        nn.init.constant_(self.vector[4, :], 0.5)     #dws_conv
+
+
+    def fre_init(self):
+        prior_tensor = torch.Tensor(self.out_channels, self.kernel_size, self.kernel_size)
+        half_fg = self.out_channels/2
+        for i in range(self.out_channels):
+            for h in range(3):
+                for w in range(3):
+                    if i < half_fg:
+                        prior_tensor[i, h, w] = math.cos(math.pi*(h+0.5)*(i+1)/3)
+                    else:
+                        prior_tensor[i, h, w] = math.cos(math.pi*(w+0.5)*(i+1-half_fg)/3)
+
+        self.register_buffer('weight_rbr_prior', prior_tensor)
+
+    def weight_gen(self):
+
+        weight_rbr_origin = torch.einsum('oihw,o->oihw', self.weight_rbr_origin, self.vector[0, :])
+
+        weight_rbr_avg = torch.einsum('oihw,o->oihw', torch.einsum('oihw,hw->oihw', self.weight_rbr_avg_conv, self.weight_rbr_avg_avg), self.vector[1, :])
+        
+        weight_rbr_pfir = torch.einsum('oihw,o->oihw', torch.einsum('oihw,ohw->oihw', self.weight_rbr_pfir_conv, self.weight_rbr_prior), self.vector[2, :])
+
+        weight_rbr_1x1_kxk_conv1 = None
+        if hasattr(self, 'weight_rbr_1x1_kxk_idconv1'):
+            weight_rbr_1x1_kxk_conv1 = (self.weight_rbr_1x1_kxk_idconv1 + self.id_tensor).squeeze()
+        elif hasattr(self, 'weight_rbr_1x1_kxk_conv1'):
+            weight_rbr_1x1_kxk_conv1 = self.weight_rbr_1x1_kxk_conv1.squeeze()
+        else:
+            raise NotImplementedError
+        weight_rbr_1x1_kxk_conv2 = self.weight_rbr_1x1_kxk_conv2
+
+        if self.groups > 1:
+            g = self.groups
+            t, ig = weight_rbr_1x1_kxk_conv1.size()
+            o, tg, h, w = weight_rbr_1x1_kxk_conv2.size()
+            weight_rbr_1x1_kxk_conv1 = weight_rbr_1x1_kxk_conv1.view(g, int(t/g), ig)
+            weight_rbr_1x1_kxk_conv2 = weight_rbr_1x1_kxk_conv2.view(g, int(o/g), tg, h, w)
+            weight_rbr_1x1_kxk = torch.einsum('gti,gothw->goihw', weight_rbr_1x1_kxk_conv1, weight_rbr_1x1_kxk_conv2).view(o, ig, h, w)
+        else:
+            weight_rbr_1x1_kxk = torch.einsum('ti,othw->oihw', weight_rbr_1x1_kxk_conv1, weight_rbr_1x1_kxk_conv2)
+
+        weight_rbr_1x1_kxk = torch.einsum('oihw,o->oihw', weight_rbr_1x1_kxk, self.vector[3, :])
+
+        weight_rbr_gconv = self.dwsc2full(self.weight_rbr_gconv_dw, self.weight_rbr_gconv_pw, self.in_channels)
+        weight_rbr_gconv = torch.einsum('oihw,o->oihw', weight_rbr_gconv, self.vector[4, :])    
+
+        weight = weight_rbr_origin + weight_rbr_avg + weight_rbr_1x1_kxk + weight_rbr_pfir + weight_rbr_gconv
+
+        return weight
+
+    def dwsc2full(self, weight_dw, weight_pw, groups):
+        
+        t, ig, h, w = weight_dw.size()
+        o, _, _, _ = weight_pw.size()
+        tg = int(t/groups)
+        i = int(ig*groups)
+        weight_dw = weight_dw.view(groups, tg, ig, h, w)
+        weight_pw = weight_pw.squeeze().view(o, groups, tg)
+        
+        weight_dsc = torch.einsum('gtihw,ogt->ogihw', weight_dw, weight_pw)
+        return weight_dsc.view(o, i, h, w)
+
+    def forward(self, inputs):
+        weight = self.weight_gen()
+        out = F.conv2d(inputs, weight, bias=None, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups)
+
+        return self.nonlinear(self.bn(out))
+
+class RepConv_OREPA(nn.Module):
+
+    def __init__(self, c1, c2, k=3, s=1, padding=1, dilation=1, groups=1, padding_mode='zeros', deploy=False, use_se=False, nonlinear=nn.SiLU()):
+        super(RepConv_OREPA, self).__init__()
+        self.deploy = deploy
+        self.groups = groups
+        self.in_channels = c1
+        self.out_channels = c2
+
+        self.padding = padding
+        self.dilation = dilation
+        self.groups = groups
+
+        assert k == 3
+        assert padding == 1
+
+        padding_11 = padding - k // 2
+
+        if nonlinear is None:
+            self.nonlinearity = nn.Identity()
+        else:
+            self.nonlinearity = nonlinear
+
+        if use_se:
+            self.se = SEBlock(self.out_channels, internal_neurons=self.out_channels // 16)
+        else:
+            self.se = nn.Identity()
+
+        if deploy:
+            self.rbr_reparam = nn.Conv2d(in_channels=self.in_channels, out_channels=self.out_channels, kernel_size=k, stride=s,
+                                      padding=padding, dilation=dilation, groups=groups, bias=True, padding_mode=padding_mode)
+
+        else:
+            self.rbr_identity = nn.BatchNorm2d(num_features=self.in_channels) if self.out_channels == self.in_channels and s == 1 else None
+            self.rbr_dense = OREPA_3x3_RepConv(in_channels=self.in_channels, out_channels=self.out_channels, kernel_size=k, stride=s, padding=padding, groups=groups, dilation=1)
+            self.rbr_1x1 = ConvBN(in_channels=self.in_channels, out_channels=self.out_channels, kernel_size=1, stride=s, padding=padding_11, groups=groups, dilation=1)
+            print('RepVGG Block, identity = ', self.rbr_identity)
+
+
+    def forward(self, inputs):
+        if hasattr(self, 'rbr_reparam'):
+            return self.nonlinearity(self.se(self.rbr_reparam(inputs)))
+
+        if self.rbr_identity is None:
+            id_out = 0
+        else:
+            id_out = self.rbr_identity(inputs)
+
+        out1 = self.rbr_dense(inputs)
+        out2 = self.rbr_1x1(inputs)
+        out3 = id_out
+        out = out1 + out2 + out3
+
+        return self.nonlinearity(self.se(out))
+
+
+    #   Optional. This improves the accuracy and facilitates quantization.
+    #   1.  Cancel the original weight decay on rbr_dense.conv.weight and rbr_1x1.conv.weight.
+    #   2.  Use like this.
+    #       loss = criterion(....)
+    #       for every RepVGGBlock blk:
+    #           loss += weight_decay_coefficient * 0.5 * blk.get_cust_L2()
+    #       optimizer.zero_grad()
+    #       loss.backward()
+
+    # Not used for OREPA
+    def get_custom_L2(self):
+        K3 = self.rbr_dense.weight_gen()
+        K1 = self.rbr_1x1.conv.weight
+        t3 = (self.rbr_dense.bn.weight / ((self.rbr_dense.bn.running_var + self.rbr_dense.bn.eps).sqrt())).reshape(-1, 1, 1, 1).detach()
+        t1 = (self.rbr_1x1.bn.weight / ((self.rbr_1x1.bn.running_var + self.rbr_1x1.bn.eps).sqrt())).reshape(-1, 1, 1, 1).detach()
+
+        l2_loss_circle = (K3 ** 2).sum() - (K3[:, :, 1:2, 1:2] ** 2).sum()      # The L2 loss of the "circle" of weights in 3x3 kernel. Use regular L2 on them.
+        eq_kernel = K3[:, :, 1:2, 1:2] * t3 + K1 * t1                           # The equivalent resultant central point of 3x3 kernel.
+        l2_loss_eq_kernel = (eq_kernel ** 2 / (t3 ** 2 + t1 ** 2)).sum()        # Normalize for an L2 coefficient comparable to regular L2.
+        return l2_loss_eq_kernel + l2_loss_circle
+
+    def get_equivalent_kernel_bias(self):
+        kernel3x3, bias3x3 = self._fuse_bn_tensor(self.rbr_dense)
+        kernel1x1, bias1x1 = self._fuse_bn_tensor(self.rbr_1x1)
+        kernelid, biasid = self._fuse_bn_tensor(self.rbr_identity)
+        return kernel3x3 + self._pad_1x1_to_3x3_tensor(kernel1x1) + kernelid, bias3x3 + bias1x1 + biasid
+
+    def _pad_1x1_to_3x3_tensor(self, kernel1x1):
+        if kernel1x1 is None:
+            return 0
+        else:
+            return torch.nn.functional.pad(kernel1x1, [1,1,1,1])
+
+    def _fuse_bn_tensor(self, branch):
+        if branch is None:
+            return 0, 0
+        if not isinstance(branch, nn.BatchNorm2d):
+            if isinstance(branch, OREPA_3x3_RepConv):
+                kernel = branch.weight_gen()
+            elif isinstance(branch, ConvBN):
+                kernel = branch.conv.weight
+            else:
+                raise NotImplementedError
+            running_mean = branch.bn.running_mean
+            running_var = branch.bn.running_var
+            gamma = branch.bn.weight
+            beta = branch.bn.bias
+            eps = branch.bn.eps
+        else:
+            if not hasattr(self, 'id_tensor'):
+                input_dim = self.in_channels // self.groups
+                kernel_value = np.zeros((self.in_channels, input_dim, 3, 3), dtype=np.float32)
+                for i in range(self.in_channels):
+                    kernel_value[i, i % input_dim, 1, 1] = 1
+                self.id_tensor = torch.from_numpy(kernel_value).to(branch.weight.device)
+            kernel = self.id_tensor
+            running_mean = branch.running_mean
+            running_var = branch.running_var
+            gamma = branch.weight
+            beta = branch.bias
+            eps = branch.eps
+        std = (running_var + eps).sqrt()
+        t = (gamma / std).reshape(-1, 1, 1, 1)
+        return kernel * t, beta - running_mean * gamma / std
+
+    def switch_to_deploy(self):
+        if hasattr(self, 'rbr_reparam'):
+            return
+        print(f"RepConv_OREPA.switch_to_deploy")
+        kernel, bias = self.get_equivalent_kernel_bias()
+        self.rbr_reparam = nn.Conv2d(in_channels=self.rbr_dense.in_channels, out_channels=self.rbr_dense.out_channels,
+                                     kernel_size=self.rbr_dense.kernel_size, stride=self.rbr_dense.stride,
+                                     padding=self.rbr_dense.padding, dilation=self.rbr_dense.dilation, groups=self.rbr_dense.groups, bias=True)
+        self.rbr_reparam.weight.data = kernel
+        self.rbr_reparam.bias.data = bias
+        for para in self.parameters():
+            para.detach_()
+        self.__delattr__('rbr_dense')
+        self.__delattr__('rbr_1x1')
+        if hasattr(self, 'rbr_identity'):
+            self.__delattr__('rbr_identity') 
+
+##### end of orepa #####
+
+
+##### swin transformer #####    
+    
+class WindowAttention(nn.Module):
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        nn.init.normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        # print(attn.dtype, v.dtype)
+        try:
+            x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        except:
+            #print(attn.dtype, v.dtype)
+            x = (attn.half() @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+class Mlp(nn.Module):
+
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.SiLU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+def window_partition(x, window_size):
+
+    B, H, W, C = x.shape
+    assert H % window_size == 0, 'feature map h and w can not divide by window size'
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+def window_reverse(windows, window_size, H, W):
+    
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class SwinTransformerLayer(nn.Module):
+
+    def __init__(self, dim, num_heads, window_size=8, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.SiLU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        # if min(self.input_resolution) <= self.window_size:
+        #     # if window size is larger than input resolution, we don't partition windows
+        #     self.shift_size = 0
+        #     self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=(self.window_size, self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def create_mask(self, H, W):
+        # calculate attention mask for SW-MSA
+        img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        return attn_mask
+
+    def forward(self, x):
+        # reshape x[b c h w] to x[b l c]
+        _, _, H_, W_ = x.shape
+
+        Padding = False
+        if min(H_, W_) < self.window_size or H_ % self.window_size!=0 or W_ % self.window_size!=0:
+            Padding = True
+            # print(f'img_size {min(H_, W_)} is less than (or not divided by) window_size {self.window_size}, Padding.')
+            pad_r = (self.window_size - W_ % self.window_size) % self.window_size
+            pad_b = (self.window_size - H_ % self.window_size) % self.window_size
+            x = F.pad(x, (0, pad_r, 0, pad_b))
+
+        # print('2', x.shape)
+        B, C, H, W = x.shape
+        L = H * W
+        x = x.permute(0, 2, 3, 1).contiguous().view(B, L, C)  # b, L, c
+
+        # create mask from init to forward
+        if self.shift_size > 0:
+            attn_mask = self.create_mask(H, W).to(x.device)
+        else:
+            attn_mask = None
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        x = x.permute(0, 2, 1).contiguous().view(-1, C, H, W)  # b c h w
+
+        if Padding:
+            x = x[:, :, :H_, :W_]  # reverse padding
+
+        return x
+
+
+class SwinTransformerBlock(nn.Module):
+    def __init__(self, c1, c2, num_heads, num_layers, window_size=8):
+        super().__init__()
+        self.conv = None
+        if c1 != c2:
+            self.conv = Conv(c1, c2)
+
+        # remove input_resolution
+        self.blocks = nn.Sequential(*[SwinTransformerLayer(dim=c2, num_heads=num_heads, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2) for i in range(num_layers)])
+
+    def forward(self, x):
+        if self.conv is not None:
+            x = self.conv(x)
+        x = self.blocks(x)
+        return x
+
+
+class STCSPA(nn.Module):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super(STCSPA, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(2 * c_, c2, 1, 1)
+        num_heads = c_ // 32
+        self.m = SwinTransformerBlock(c_, c_, num_heads, n)
+        #self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+    def forward(self, x):
+        y1 = self.m(self.cv1(x))
+        y2 = self.cv2(x)
+        return self.cv3(torch.cat((y1, y2), dim=1))
+
+
+class STCSPB(nn.Module):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super(STCSPB, self).__init__()
+        c_ = int(c2)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_, c_, 1, 1)
+        self.cv3 = Conv(2 * c_, c2, 1, 1)
+        num_heads = c_ // 32
+        self.m = SwinTransformerBlock(c_, c_, num_heads, n)
+        #self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+    def forward(self, x):
+        x1 = self.cv1(x)
+        y1 = self.m(x1)
+        y2 = self.cv2(x1)
+        return self.cv3(torch.cat((y1, y2), dim=1))
+
+
+class STCSPC(nn.Module):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super(STCSPC, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(c_, c_, 1, 1)
+        self.cv4 = Conv(2 * c_, c2, 1, 1)
+        num_heads = c_ // 32
+        self.m = SwinTransformerBlock(c_, c_, num_heads, n)
+        #self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+    def forward(self, x):
+        y1 = self.cv3(self.m(self.cv1(x)))
+        y2 = self.cv2(x)
+        return self.cv4(torch.cat((y1, y2), dim=1))
+
+##### end of swin transformer #####   
+
+
+##### swin transformer v2 ##### 
+  
+class WindowAttention_v2(nn.Module):
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, attn_drop=0., proj_drop=0.,
+                 pretrained_window_size=[0, 0]):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.pretrained_window_size = pretrained_window_size
+        self.num_heads = num_heads
+
+        self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True)
+
+        # mlp to generate continuous relative position bias
+        self.cpb_mlp = nn.Sequential(nn.Linear(2, 512, bias=True),
+                                     nn.ReLU(inplace=True),
+                                     nn.Linear(512, num_heads, bias=False))
+
+        # get relative_coords_table
+        relative_coords_h = torch.arange(-(self.window_size[0] - 1), self.window_size[0], dtype=torch.float32)
+        relative_coords_w = torch.arange(-(self.window_size[1] - 1), self.window_size[1], dtype=torch.float32)
+        relative_coords_table = torch.stack(
+            torch.meshgrid([relative_coords_h,
+                            relative_coords_w])).permute(1, 2, 0).contiguous().unsqueeze(0)  # 1, 2*Wh-1, 2*Ww-1, 2
+        if pretrained_window_size[0] > 0:
+            relative_coords_table[:, :, :, 0] /= (pretrained_window_size[0] - 1)
+            relative_coords_table[:, :, :, 1] /= (pretrained_window_size[1] - 1)
+        else:
+            relative_coords_table[:, :, :, 0] /= (self.window_size[0] - 1)
+            relative_coords_table[:, :, :, 1] /= (self.window_size[1] - 1)
+        relative_coords_table *= 8  # normalize to -8, 8
+        relative_coords_table = torch.sign(relative_coords_table) * torch.log2(
+            torch.abs(relative_coords_table) + 1.0) / np.log2(8)
+
+        self.register_buffer("relative_coords_table", relative_coords_table)
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=False)
+        if qkv_bias:
+            self.q_bias = nn.Parameter(torch.zeros(dim))
+            self.v_bias = nn.Parameter(torch.zeros(dim))
+        else:
+            self.q_bias = None
+            self.v_bias = None
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        
+        B_, N, C = x.shape
+        qkv_bias = None
+        if self.q_bias is not None:
+            qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))
+        qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+        qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        # cosine attention
+        attn = (F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1))
+        logit_scale = torch.clamp(self.logit_scale, max=torch.log(torch.tensor(1. / 0.01))).exp()
+        attn = attn * logit_scale
+
+        relative_position_bias_table = self.cpb_mlp(self.relative_coords_table).view(-1, self.num_heads)
+        relative_position_bias = relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        relative_position_bias = 16 * torch.sigmoid(relative_position_bias)
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        try:
+            x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        except:
+            x = (attn.half() @ v).transpose(1, 2).reshape(B_, N, C)
+            
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}, window_size={self.window_size}, ' \
+               f'pretrained_window_size={self.pretrained_window_size}, num_heads={self.num_heads}'
+
+    def flops(self, N):
+        # calculate flops for 1 window with token length of N
+        flops = 0
+        # qkv = self.qkv(x)
+        flops += N * self.dim * 3 * self.dim
+        # attn = (q @ k.transpose(-2, -1))
+        flops += self.num_heads * N * (self.dim // self.num_heads) * N
+        #  x = (attn @ v)
+        flops += self.num_heads * N * N * (self.dim // self.num_heads)
+        # x = self.proj(x)
+        flops += N * self.dim * self.dim
+        return flops
+    
+class Mlp_v2(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.SiLU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition_v2(x, window_size):
+    
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse_v2(windows, window_size, H, W):
+    
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class SwinTransformerLayer_v2(nn.Module):
+
+    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.SiLU, norm_layer=nn.LayerNorm, pretrained_window_size=0):
+        super().__init__()
+        self.dim = dim
+        #self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        #if min(self.input_resolution) <= self.window_size:
+        #    # if window size is larger than input resolution, we don't partition windows
+        #    self.shift_size = 0
+        #    self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention_v2(
+            dim, window_size=(self.window_size, self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop,
+            pretrained_window_size=(pretrained_window_size, pretrained_window_size))
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp_v2(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def create_mask(self, H, W):
+        # calculate attention mask for SW-MSA
+        img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        return attn_mask
+
+    def forward(self, x):
+        # reshape x[b c h w] to x[b l c]
+        _, _, H_, W_ = x.shape
+
+        Padding = False
+        if min(H_, W_) < self.window_size or H_ % self.window_size!=0 or W_ % self.window_size!=0:
+            Padding = True
+            # print(f'img_size {min(H_, W_)} is less than (or not divided by) window_size {self.window_size}, Padding.')
+            pad_r = (self.window_size - W_ % self.window_size) % self.window_size
+            pad_b = (self.window_size - H_ % self.window_size) % self.window_size
+            x = F.pad(x, (0, pad_r, 0, pad_b))
+
+        # print('2', x.shape)
+        B, C, H, W = x.shape
+        L = H * W
+        x = x.permute(0, 2, 3, 1).contiguous().view(B, L, C)  # b, L, c
+
+        # create mask from init to forward
+        if self.shift_size > 0:
+            attn_mask = self.create_mask(H, W).to(x.device)
+        else:
+            attn_mask = None
+
+        shortcut = x
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition_v2(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse_v2(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+        x = shortcut + self.drop_path(self.norm1(x))
+
+        # FFN
+        x = x + self.drop_path(self.norm2(self.mlp(x)))
+        x = x.permute(0, 2, 1).contiguous().view(-1, C, H, W)  # b c h w
+        
+        if Padding:
+            x = x[:, :, :H_, :W_]  # reverse padding
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
+               f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
+
+    def flops(self):
+        flops = 0
+        H, W = self.input_resolution
+        # norm1
+        flops += self.dim * H * W
+        # W-MSA/SW-MSA
+        nW = H * W / self.window_size / self.window_size
+        flops += nW * self.attn.flops(self.window_size * self.window_size)
+        # mlp
+        flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
+        # norm2
+        flops += self.dim * H * W
+        return flops
+
+
+class SwinTransformer2Block(nn.Module):
+    def __init__(self, c1, c2, num_heads, num_layers, window_size=7):
+        super().__init__()
+        self.conv = None
+        if c1 != c2:
+            self.conv = Conv(c1, c2)
+
+        # remove input_resolution
+        self.blocks = nn.Sequential(*[SwinTransformerLayer_v2(dim=c2, num_heads=num_heads, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2) for i in range(num_layers)])
+
+    def forward(self, x):
+        if self.conv is not None:
+            x = self.conv(x)
+        x = self.blocks(x)
+        return x
+
+
+class ST2CSPA(nn.Module):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super(ST2CSPA, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(2 * c_, c2, 1, 1)
+        num_heads = c_ // 32
+        self.m = SwinTransformer2Block(c_, c_, num_heads, n)
+        #self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+    def forward(self, x):
+        y1 = self.m(self.cv1(x))
+        y2 = self.cv2(x)
+        return self.cv3(torch.cat((y1, y2), dim=1))
+
+
+class ST2CSPB(nn.Module):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super(ST2CSPB, self).__init__()
+        c_ = int(c2)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_, c_, 1, 1)
+        self.cv3 = Conv(2 * c_, c2, 1, 1)
+        num_heads = c_ // 32
+        self.m = SwinTransformer2Block(c_, c_, num_heads, n)
+        #self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+    def forward(self, x):
+        x1 = self.cv1(x)
+        y1 = self.m(x1)
+        y2 = self.cv2(x1)
+        return self.cv3(torch.cat((y1, y2), dim=1))
+
+
+class ST2CSPC(nn.Module):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super(ST2CSPC, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c1, c_, 1, 1)
+        self.cv3 = Conv(c_, c_, 1, 1)
+        self.cv4 = Conv(2 * c_, c2, 1, 1)
+        num_heads = c_ // 32
+        self.m = SwinTransformer2Block(c_, c_, num_heads, n)
+        #self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
+
+    def forward(self, x):
+        y1 = self.cv3(self.m(self.cv1(x)))
+        y2 = self.cv2(x)
+        return self.cv4(torch.cat((y1, y2), dim=1))
+
+##### end of swin transformer v2 #####   

models/experimental.py

@@ -0,0 +1,272 @@
+import numpy as np
+import random
+import torch
+import torch.nn as nn
+
+from models.common import Conv, DWConv
+from utils.google_utils import attempt_download
+
+
+class CrossConv(nn.Module):
+    # Cross Convolution Downsample
+    def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False):
+        # ch_in, ch_out, kernel, stride, groups, expansion, shortcut
+        super(CrossConv, self).__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, (1, k), (1, s))
+        self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
+
+
+class Sum(nn.Module):
+    # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070
+    def __init__(self, n, weight=False):  # n: number of inputs
+        super(Sum, self).__init__()
+        self.weight = weight  # apply weights boolean
+        self.iter = range(n - 1)  # iter object
+        if weight:
+            self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True)  # layer weights
+
+    def forward(self, x):
+        y = x[0]  # no weight
+        if self.weight:
+            w = torch.sigmoid(self.w) * 2
+            for i in self.iter:
+                y = y + x[i + 1] * w[i]
+        else:
+            for i in self.iter:
+                y = y + x[i + 1]
+        return y
+
+
+class MixConv2d(nn.Module):
+    # Mixed Depthwise Conv https://arxiv.org/abs/1907.09595
+    def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True):
+        super(MixConv2d, self).__init__()
+        groups = len(k)
+        if equal_ch:  # equal c_ per group
+            i = torch.linspace(0, groups - 1E-6, c2).floor()  # c2 indices
+            c_ = [(i == g).sum() for g in range(groups)]  # intermediate channels
+        else:  # equal weight.numel() per group
+            b = [c2] + [0] * groups
+            a = np.eye(groups + 1, groups, k=-1)
+            a -= np.roll(a, 1, axis=1)
+            a *= np.array(k) ** 2
+            a[0] = 1
+            c_ = np.linalg.lstsq(a, b, rcond=None)[0].round()  # solve for equal weight indices, ax = b
+
+        self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)])
+        self.bn = nn.BatchNorm2d(c2)
+        self.act = nn.LeakyReLU(0.1, inplace=True)
+
+    def forward(self, x):
+        return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1)))
+
+
+class Ensemble(nn.ModuleList):
+    # Ensemble of models
+    def __init__(self):
+        super(Ensemble, self).__init__()
+
+    def forward(self, x, augment=False):
+        y = []
+        for module in self:
+            y.append(module(x, augment)[0])
+        # y = torch.stack(y).max(0)[0]  # max ensemble
+        # y = torch.stack(y).mean(0)  # mean ensemble
+        y = torch.cat(y, 1)  # nms ensemble
+        return y, None  # inference, train output
+
+
+
+
+
+class ORT_NMS(torch.autograd.Function):
+    '''ONNX-Runtime NMS operation'''
+    @staticmethod
+    def forward(ctx,
+                boxes,
+                scores,
+                max_output_boxes_per_class=torch.tensor([100]),
+                iou_threshold=torch.tensor([0.45]),
+                score_threshold=torch.tensor([0.25])):
+        device = boxes.device
+        batch = scores.shape[0]
+        num_det = random.randint(0, 100)
+        batches = torch.randint(0, batch, (num_det,)).sort()[0].to(device)
+        idxs = torch.arange(100, 100 + num_det).to(device)
+        zeros = torch.zeros((num_det,), dtype=torch.int64).to(device)
+        selected_indices = torch.cat([batches[None], zeros[None], idxs[None]], 0).T.contiguous()
+        selected_indices = selected_indices.to(torch.int64)
+        return selected_indices
+
+    @staticmethod
+    def symbolic(g, boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold):
+        return g.op("NonMaxSuppression", boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold)
+
+
+class TRT_NMS(torch.autograd.Function):
+    '''TensorRT NMS operation'''
+    @staticmethod
+    def forward(
+        ctx,
+        boxes,
+        scores,
+        background_class=-1,
+        box_coding=1,
+        iou_threshold=0.45,
+        max_output_boxes=100,
+        plugin_version="1",
+        score_activation=0,
+        score_threshold=0.25,
+    ):
+        batch_size, num_boxes, num_classes = scores.shape
+        num_det = torch.randint(0, max_output_boxes, (batch_size, 1), dtype=torch.int32)
+        det_boxes = torch.randn(batch_size, max_output_boxes, 4)
+        det_scores = torch.randn(batch_size, max_output_boxes)
+        det_classes = torch.randint(0, num_classes, (batch_size, max_output_boxes), dtype=torch.int32)
+        return num_det, det_boxes, det_scores, det_classes
+
+    @staticmethod
+    def symbolic(g,
+                 boxes,
+                 scores,
+                 background_class=-1,
+                 box_coding=1,
+                 iou_threshold=0.45,
+                 max_output_boxes=100,
+                 plugin_version="1",
+                 score_activation=0,
+                 score_threshold=0.25):
+        out = g.op("TRT::EfficientNMS_TRT",
+                   boxes,
+                   scores,
+                   background_class_i=background_class,
+                   box_coding_i=box_coding,
+                   iou_threshold_f=iou_threshold,
+                   max_output_boxes_i=max_output_boxes,
+                   plugin_version_s=plugin_version,
+                   score_activation_i=score_activation,
+                   score_threshold_f=score_threshold,
+                   outputs=4)
+        nums, boxes, scores, classes = out
+        return nums, boxes, scores, classes
+
+
+class ONNX_ORT(nn.Module):
+    '''onnx module with ONNX-Runtime NMS operation.'''
+    def __init__(self, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=640, device=None, n_classes=80):
+        super().__init__()
+        self.device = device if device else torch.device("cpu")
+        self.max_obj = torch.tensor([max_obj]).to(device)
+        self.iou_threshold = torch.tensor([iou_thres]).to(device)
+        self.score_threshold = torch.tensor([score_thres]).to(device)
+        self.max_wh = max_wh # if max_wh != 0 : non-agnostic else : agnostic
+        self.convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]],
+                                           dtype=torch.float32,
+                                           device=self.device)
+        self.n_classes=n_classes
+
+    def forward(self, x):
+        boxes = x[:, :, :4]
+        conf = x[:, :, 4:5]
+        scores = x[:, :, 5:]
+        if self.n_classes == 1:
+            scores = conf # for models with one class, cls_loss is 0 and cls_conf is always 0.5,
+                                 # so there is no need to multiplicate.
+        else:
+            scores *= conf  # conf = obj_conf * cls_conf
+        boxes @= self.convert_matrix
+        max_score, category_id = scores.max(2, keepdim=True)
+        dis = category_id.float() * self.max_wh
+        nmsbox = boxes + dis
+        max_score_tp = max_score.transpose(1, 2).contiguous()
+        selected_indices = ORT_NMS.apply(nmsbox, max_score_tp, self.max_obj, self.iou_threshold, self.score_threshold)
+        X, Y = selected_indices[:, 0], selected_indices[:, 2]
+        selected_boxes = boxes[X, Y, :]
+        selected_categories = category_id[X, Y, :].float()
+        selected_scores = max_score[X, Y, :]
+        X = X.unsqueeze(1).float()
+        return torch.cat([X, selected_boxes, selected_categories, selected_scores], 1)
+
+class ONNX_TRT(nn.Module):
+    '''onnx module with TensorRT NMS operation.'''
+    def __init__(self, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=None ,device=None, n_classes=80):
+        super().__init__()
+        assert max_wh is None
+        self.device = device if device else torch.device('cpu')
+        self.background_class = -1,
+        self.box_coding = 1,
+        self.iou_threshold = iou_thres
+        self.max_obj = max_obj
+        self.plugin_version = '1'
+        self.score_activation = 0
+        self.score_threshold = score_thres
+        self.n_classes=n_classes
+
+    def forward(self, x):
+        boxes = x[:, :, :4]
+        conf = x[:, :, 4:5]
+        scores = x[:, :, 5:]
+        if self.n_classes == 1:
+            scores = conf # for models with one class, cls_loss is 0 and cls_conf is always 0.5,
+                                 # so there is no need to multiplicate.
+        else:
+            scores *= conf  # conf = obj_conf * cls_conf
+        num_det, det_boxes, det_scores, det_classes = TRT_NMS.apply(boxes, scores, self.background_class, self.box_coding,
+                                                                    self.iou_threshold, self.max_obj,
+                                                                    self.plugin_version, self.score_activation,
+                                                                    self.score_threshold)
+        return num_det, det_boxes, det_scores, det_classes
+
+
+class End2End(nn.Module):
+    '''export onnx or tensorrt model with NMS operation.'''
+    def __init__(self, model, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=None, device=None, n_classes=80):
+        super().__init__()
+        device = device if device else torch.device('cpu')
+        assert isinstance(max_wh,(int)) or max_wh is None
+        self.model = model.to(device)
+        self.model.model[-1].end2end = True
+        self.patch_model = ONNX_TRT if max_wh is None else ONNX_ORT
+        self.end2end = self.patch_model(max_obj, iou_thres, score_thres, max_wh, device, n_classes)
+        self.end2end.eval()
+
+    def forward(self, x):
+        x = self.model(x)
+        x = self.end2end(x)
+        return x
+
+
+
+
+
+def attempt_load(weights, map_location=None):
+    # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a
+    model = Ensemble()
+    for w in weights if isinstance(weights, list) else [weights]:
+        attempt_download(w)
+        ckpt = torch.load(w, map_location=map_location)  # load
+        model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().fuse().eval())  # FP32 model
+    
+    # Compatibility updates
+    for m in model.modules():
+        if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]:
+            m.inplace = True  # pytorch 1.7.0 compatibility
+        elif type(m) is nn.Upsample:
+            m.recompute_scale_factor = None  # torch 1.11.0 compatibility
+        elif type(m) is Conv:
+            m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
+    
+    if len(model) == 1:
+        return model[-1]  # return model
+    else:
+        print('Ensemble created with %s\n' % weights)
+        for k in ['names', 'stride']:
+            setattr(model, k, getattr(model[-1], k))
+        return model  # return ensemble
+
+

models/yolo.py

@@ -0,0 +1,843 @@
+import argparse
+import logging
+import sys
+from copy import deepcopy
+
+sys.path.append('./')  # to run '$ python *.py' files in subdirectories
+logger = logging.getLogger(__name__)
+import torch
+from models.common import *
+from models.experimental import *
+from utils.autoanchor import check_anchor_order
+from utils.general import make_divisible, check_file, set_logging
+from utils.torch_utils import time_synchronized, fuse_conv_and_bn, model_info, scale_img, initialize_weights, \
+    select_device, copy_attr
+from utils.loss import SigmoidBin
+
+try:
+    import thop  # for FLOPS computation
+except ImportError:
+    thop = None
+
+
+class Detect(nn.Module):
+    stride = None  # strides computed during build
+    export = False  # onnx export
+    end2end = False
+    include_nms = False
+    concat = False
+
+    def __init__(self, nc=80, anchors=(), ch=()):  # detection layer
+        super(Detect, self).__init__()
+        self.nc = nc  # number of classes
+        self.no = nc + 5  # number of outputs per anchor
+        self.nl = len(anchors)  # number of detection layers
+        self.na = len(anchors[0]) // 2  # number of anchors
+        self.grid = [torch.zeros(1)] * self.nl  # init grid
+        a = torch.tensor(anchors).float().view(self.nl, -1, 2)
+        self.register_buffer('anchors', a)  # shape(nl,na,2)
+        self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2))  # shape(nl,1,na,1,1,2)
+        self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch)  # output conv
+
+    def forward(self, x):
+        # x = x.copy()  # for profiling
+        z = []  # inference output
+        self.training |= self.export
+        for i in range(self.nl):
+            x[i] = self.m[i](x[i])  # conv
+            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
+            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+
+            if not self.training:  # inference
+                if self.grid[i].shape[2:4] != x[i].shape[2:4]:
+                    self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
+                y = x[i].sigmoid()
+                if not torch.onnx.is_in_onnx_export():
+                    y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+                else:
+                    xy, wh, conf = y.split((2, 2, self.nc + 1), 4)  # y.tensor_split((2, 4, 5), 4)  # torch 1.8.0
+                    xy = xy * (2. * self.stride[i]) + (self.stride[i] * (self.grid[i] - 0.5))  # new xy
+                    wh = wh ** 2 * (4 * self.anchor_grid[i].data)  # new wh
+                    y = torch.cat((xy, wh, conf), 4)
+                z.append(y.view(bs, -1, self.no))
+
+        if self.training:
+            out = x
+        elif self.end2end:
+            out = torch.cat(z, 1)
+        elif self.include_nms:
+            z = self.convert(z)
+            out = (z, )
+        elif self.concat:
+            out = torch.cat(z, 1)
+        else:
+            out = (torch.cat(z, 1), x)
+
+        return out
+
+    @staticmethod
+    def _make_grid(nx=20, ny=20):
+        yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
+        return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
+
+    def convert(self, z):
+        z = torch.cat(z, 1)
+        box = z[:, :, :4]
+        conf = z[:, :, 4:5]
+        score = z[:, :, 5:]
+        score *= conf
+        convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]],
+                                           dtype=torch.float32,
+                                           device=z.device)
+        box @= convert_matrix                          
+        return (box, score)
+
+
+class IDetect(nn.Module):
+    stride = None  # strides computed during build
+    export = False  # onnx export
+    end2end = False
+    include_nms = False
+    concat = False
+
+    def __init__(self, nc=80, anchors=(), ch=()):  # detection layer
+        super(IDetect, self).__init__()
+        self.nc = nc  # number of classes
+        self.no = nc + 5  # number of outputs per anchor
+        self.nl = len(anchors)  # number of detection layers
+        self.na = len(anchors[0]) // 2  # number of anchors
+        self.grid = [torch.zeros(1)] * self.nl  # init grid
+        a = torch.tensor(anchors).float().view(self.nl, -1, 2)
+        self.register_buffer('anchors', a)  # shape(nl,na,2)
+        self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2))  # shape(nl,1,na,1,1,2)
+        self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch)  # output conv
+        
+        self.ia = nn.ModuleList(ImplicitA(x) for x in ch)
+        self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch)
+
+    def forward(self, x):
+        # x = x.copy()  # for profiling
+        z = []  # inference output
+        self.training |= self.export
+        for i in range(self.nl):
+            x[i] = self.m[i](self.ia[i](x[i]))  # conv
+            x[i] = self.im[i](x[i])
+            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
+            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+
+            if not self.training:  # inference
+                if self.grid[i].shape[2:4] != x[i].shape[2:4]:
+                    self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
+
+                y = x[i].sigmoid()
+                y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+                z.append(y.view(bs, -1, self.no))
+
+        return x if self.training else (torch.cat(z, 1), x)
+    
+    def fuseforward(self, x):
+        # x = x.copy()  # for profiling
+        z = []  # inference output
+        self.training |= self.export
+        for i in range(self.nl):
+            x[i] = self.m[i](x[i])  # conv
+            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
+            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+
+            if not self.training:  # inference
+                if self.grid[i].shape[2:4] != x[i].shape[2:4]:
+                    self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
+
+                y = x[i].sigmoid()
+                if not torch.onnx.is_in_onnx_export():
+                    y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+                else:
+                    xy, wh, conf = y.split((2, 2, self.nc + 1), 4)  # y.tensor_split((2, 4, 5), 4)  # torch 1.8.0
+                    xy = xy * (2. * self.stride[i]) + (self.stride[i] * (self.grid[i] - 0.5))  # new xy
+                    wh = wh ** 2 * (4 * self.anchor_grid[i].data)  # new wh
+                    y = torch.cat((xy, wh, conf), 4)
+                z.append(y.view(bs, -1, self.no))
+
+        if self.training:
+            out = x
+        elif self.end2end:
+            out = torch.cat(z, 1)
+        elif self.include_nms:
+            z = self.convert(z)
+            out = (z, )
+        elif self.concat:
+            out = torch.cat(z, 1)            
+        else:
+            out = (torch.cat(z, 1), x)
+
+        return out
+    
+    def fuse(self):
+        print("IDetect.fuse")
+        # fuse ImplicitA and Convolution
+        for i in range(len(self.m)):
+            c1,c2,_,_ = self.m[i].weight.shape
+            c1_,c2_, _,_ = self.ia[i].implicit.shape
+            self.m[i].bias += torch.matmul(self.m[i].weight.reshape(c1,c2),self.ia[i].implicit.reshape(c2_,c1_)).squeeze(1)
+
+        # fuse ImplicitM and Convolution
+        for i in range(len(self.m)):
+            c1,c2, _,_ = self.im[i].implicit.shape
+            self.m[i].bias *= self.im[i].implicit.reshape(c2)
+            self.m[i].weight *= self.im[i].implicit.transpose(0,1)
+            
+    @staticmethod
+    def _make_grid(nx=20, ny=20):
+        yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
+        return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
+
+    def convert(self, z):
+        z = torch.cat(z, 1)
+        box = z[:, :, :4]
+        conf = z[:, :, 4:5]
+        score = z[:, :, 5:]
+        score *= conf
+        convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]],
+                                           dtype=torch.float32,
+                                           device=z.device)
+        box @= convert_matrix                          
+        return (box, score)
+
+
+class IKeypoint(nn.Module):
+    stride = None  # strides computed during build
+    export = False  # onnx export
+
+    def __init__(self, nc=80, anchors=(), nkpt=17, ch=(), inplace=True, dw_conv_kpt=False):  # detection layer
+        super(IKeypoint, self).__init__()
+        self.nc = nc  # number of classes
+        self.nkpt = nkpt
+        self.dw_conv_kpt = dw_conv_kpt
+        self.no_det=(nc + 5)  # number of outputs per anchor for box and class
+        self.no_kpt = 3*self.nkpt ## number of outputs per anchor for keypoints
+        self.no = self.no_det+self.no_kpt
+        self.nl = len(anchors)  # number of detection layers
+        self.na = len(anchors[0]) // 2  # number of anchors
+        self.grid = [torch.zeros(1)] * self.nl  # init grid
+        self.flip_test = False
+        a = torch.tensor(anchors).float().view(self.nl, -1, 2)
+        self.register_buffer('anchors', a)  # shape(nl,na,2)
+        self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2))  # shape(nl,1,na,1,1,2)
+        self.m = nn.ModuleList(nn.Conv2d(x, self.no_det * self.na, 1) for x in ch)  # output conv
+        
+        self.ia = nn.ModuleList(ImplicitA(x) for x in ch)
+        self.im = nn.ModuleList(ImplicitM(self.no_det * self.na) for _ in ch)
+        
+        if self.nkpt is not None:
+            if self.dw_conv_kpt: #keypoint head is slightly more complex
+                self.m_kpt = nn.ModuleList(
+                            nn.Sequential(DWConv(x, x, k=3), Conv(x,x),
+                                          DWConv(x, x, k=3), Conv(x, x),
+                                          DWConv(x, x, k=3), Conv(x,x),
+                                          DWConv(x, x, k=3), Conv(x, x),
+                                          DWConv(x, x, k=3), Conv(x, x),
+                                          DWConv(x, x, k=3), nn.Conv2d(x, self.no_kpt * self.na, 1)) for x in ch)
+            else: #keypoint head is a single convolution
+                self.m_kpt = nn.ModuleList(nn.Conv2d(x, self.no_kpt * self.na, 1) for x in ch)
+
+        self.inplace = inplace  # use in-place ops (e.g. slice assignment)
+
+    def forward(self, x):
+        # x = x.copy()  # for profiling
+        z = []  # inference output
+        self.training |= self.export
+        for i in range(self.nl):
+            if self.nkpt is None or self.nkpt==0:
+                x[i] = self.im[i](self.m[i](self.ia[i](x[i])))  # conv
+            else :
+                x[i] = torch.cat((self.im[i](self.m[i](self.ia[i](x[i]))), self.m_kpt[i](x[i])), axis=1)
+
+            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
+            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+            x_det = x[i][..., :6]
+            x_kpt = x[i][..., 6:]
+
+            if not self.training:  # inference
+                if self.grid[i].shape[2:4] != x[i].shape[2:4]:
+                    self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
+                kpt_grid_x = self.grid[i][..., 0:1]
+                kpt_grid_y = self.grid[i][..., 1:2]
+
+                if self.nkpt == 0:
+                    y = x[i].sigmoid()
+                else:
+                    y = x_det.sigmoid()
+
+                if self.inplace:
+                    xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].view(1, self.na, 1, 1, 2) # wh
+                    if self.nkpt != 0:
+                        x_kpt[..., 0::3] = (x_kpt[..., ::3] * 2. - 0.5 + kpt_grid_x.repeat(1,1,1,1,17)) * self.stride[i]  # xy
+                        x_kpt[..., 1::3] = (x_kpt[..., 1::3] * 2. - 0.5 + kpt_grid_y.repeat(1,1,1,1,17)) * self.stride[i]  # xy
+                        #x_kpt[..., 0::3] = (x_kpt[..., ::3] + kpt_grid_x.repeat(1,1,1,1,17)) * self.stride[i]  # xy
+                        #x_kpt[..., 1::3] = (x_kpt[..., 1::3] + kpt_grid_y.repeat(1,1,1,1,17)) * self.stride[i]  # xy
+                        #print('=============')
+                        #print(self.anchor_grid[i].shape)
+                        #print(self.anchor_grid[i][...,0].unsqueeze(4).shape)
+                        #print(x_kpt[..., 0::3].shape)
+                        #x_kpt[..., 0::3] = ((x_kpt[..., 0::3].tanh() * 2.) ** 3 * self.anchor_grid[i][...,0].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_x.repeat(1,1,1,1,17) * self.stride[i]  # xy
+                        #x_kpt[..., 1::3] = ((x_kpt[..., 1::3].tanh() * 2.) ** 3 * self.anchor_grid[i][...,1].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_y.repeat(1,1,1,1,17) * self.stride[i]  # xy
+                        #x_kpt[..., 0::3] = (((x_kpt[..., 0::3].sigmoid() * 4.) ** 2 - 8.) * self.anchor_grid[i][...,0].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_x.repeat(1,1,1,1,17) * self.stride[i]  # xy
+                        #x_kpt[..., 1::3] = (((x_kpt[..., 1::3].sigmoid() * 4.) ** 2 - 8.) * self.anchor_grid[i][...,1].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_y.repeat(1,1,1,1,17) * self.stride[i]  # xy
+                        x_kpt[..., 2::3] = x_kpt[..., 2::3].sigmoid()
+
+                    y = torch.cat((xy, wh, y[..., 4:], x_kpt), dim = -1)
+
+                else:  # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953
+                    xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+                    if self.nkpt != 0:
+                        y[..., 6:] = (y[..., 6:] * 2. - 0.5 + self.grid[i].repeat((1,1,1,1,self.nkpt))) * self.stride[i]  # xy
+                    y = torch.cat((xy, wh, y[..., 4:]), -1)
+
+                z.append(y.view(bs, -1, self.no))
+
+        return x if self.training else (torch.cat(z, 1), x)
+
+    @staticmethod
+    def _make_grid(nx=20, ny=20):
+        yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
+        return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
+
+
+class IAuxDetect(nn.Module):
+    stride = None  # strides computed during build
+    export = False  # onnx export
+    end2end = False
+    include_nms = False
+    concat = False
+
+    def __init__(self, nc=80, anchors=(), ch=()):  # detection layer
+        super(IAuxDetect, self).__init__()
+        self.nc = nc  # number of classes
+        self.no = nc + 5  # number of outputs per anchor
+        self.nl = len(anchors)  # number of detection layers
+        self.na = len(anchors[0]) // 2  # number of anchors
+        self.grid = [torch.zeros(1)] * self.nl  # init grid
+        a = torch.tensor(anchors).float().view(self.nl, -1, 2)
+        self.register_buffer('anchors', a)  # shape(nl,na,2)
+        self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2))  # shape(nl,1,na,1,1,2)
+        self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch[:self.nl])  # output conv
+        self.m2 = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch[self.nl:])  # output conv
+        
+        self.ia = nn.ModuleList(ImplicitA(x) for x in ch[:self.nl])
+        self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch[:self.nl])
+
+    def forward(self, x):
+        # x = x.copy()  # for profiling
+        z = []  # inference output
+        self.training |= self.export
+        for i in range(self.nl):
+            x[i] = self.m[i](self.ia[i](x[i]))  # conv
+            x[i] = self.im[i](x[i])
+            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
+            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+            
+            x[i+self.nl] = self.m2[i](x[i+self.nl])
+            x[i+self.nl] = x[i+self.nl].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+
+            if not self.training:  # inference
+                if self.grid[i].shape[2:4] != x[i].shape[2:4]:
+                    self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
+
+                y = x[i].sigmoid()
+                if not torch.onnx.is_in_onnx_export():
+                    y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+                else:
+                    xy, wh, conf = y.split((2, 2, self.nc + 1), 4)  # y.tensor_split((2, 4, 5), 4)  # torch 1.8.0
+                    xy = xy * (2. * self.stride[i]) + (self.stride[i] * (self.grid[i] - 0.5))  # new xy
+                    wh = wh ** 2 * (4 * self.anchor_grid[i].data)  # new wh
+                    y = torch.cat((xy, wh, conf), 4)
+                z.append(y.view(bs, -1, self.no))
+
+        return x if self.training else (torch.cat(z, 1), x[:self.nl])
+
+    def fuseforward(self, x):
+        # x = x.copy()  # for profiling
+        z = []  # inference output
+        self.training |= self.export
+        for i in range(self.nl):
+            x[i] = self.m[i](x[i])  # conv
+            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
+            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+
+            if not self.training:  # inference
+                if self.grid[i].shape[2:4] != x[i].shape[2:4]:
+                    self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
+
+                y = x[i].sigmoid()
+                if not torch.onnx.is_in_onnx_export():
+                    y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+                else:
+                    xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].data  # wh
+                    y = torch.cat((xy, wh, y[..., 4:]), -1)
+                z.append(y.view(bs, -1, self.no))
+
+        if self.training:
+            out = x
+        elif self.end2end:
+            out = torch.cat(z, 1)
+        elif self.include_nms:
+            z = self.convert(z)
+            out = (z, )
+        elif self.concat:
+            out = torch.cat(z, 1)            
+        else:
+            out = (torch.cat(z, 1), x)
+
+        return out
+    
+    def fuse(self):
+        print("IAuxDetect.fuse")
+        # fuse ImplicitA and Convolution
+        for i in range(len(self.m)):
+            c1,c2,_,_ = self.m[i].weight.shape
+            c1_,c2_, _,_ = self.ia[i].implicit.shape
+            self.m[i].bias += torch.matmul(self.m[i].weight.reshape(c1,c2),self.ia[i].implicit.reshape(c2_,c1_)).squeeze(1)
+
+        # fuse ImplicitM and Convolution
+        for i in range(len(self.m)):
+            c1,c2, _,_ = self.im[i].implicit.shape
+            self.m[i].bias *= self.im[i].implicit.reshape(c2)
+            self.m[i].weight *= self.im[i].implicit.transpose(0,1)
+
+    @staticmethod
+    def _make_grid(nx=20, ny=20):
+        yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
+        return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
+
+    def convert(self, z):
+        z = torch.cat(z, 1)
+        box = z[:, :, :4]
+        conf = z[:, :, 4:5]
+        score = z[:, :, 5:]
+        score *= conf
+        convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]],
+                                           dtype=torch.float32,
+                                           device=z.device)
+        box @= convert_matrix                          
+        return (box, score)
+
+
+class IBin(nn.Module):
+    stride = None  # strides computed during build
+    export = False  # onnx export
+
+    def __init__(self, nc=80, anchors=(), ch=(), bin_count=21):  # detection layer
+        super(IBin, self).__init__()
+        self.nc = nc  # number of classes
+        self.bin_count = bin_count
+
+        self.w_bin_sigmoid = SigmoidBin(bin_count=self.bin_count, min=0.0, max=4.0)
+        self.h_bin_sigmoid = SigmoidBin(bin_count=self.bin_count, min=0.0, max=4.0)
+        # classes, x,y,obj
+        self.no = nc + 3 + \
+            self.w_bin_sigmoid.get_length() + self.h_bin_sigmoid.get_length()   # w-bce, h-bce
+            # + self.x_bin_sigmoid.get_length() + self.y_bin_sigmoid.get_length()
+        
+        self.nl = len(anchors)  # number of detection layers
+        self.na = len(anchors[0]) // 2  # number of anchors
+        self.grid = [torch.zeros(1)] * self.nl  # init grid
+        a = torch.tensor(anchors).float().view(self.nl, -1, 2)
+        self.register_buffer('anchors', a)  # shape(nl,na,2)
+        self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2))  # shape(nl,1,na,1,1,2)
+        self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch)  # output conv
+        
+        self.ia = nn.ModuleList(ImplicitA(x) for x in ch)
+        self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch)
+
+    def forward(self, x):
+
+        #self.x_bin_sigmoid.use_fw_regression = True
+        #self.y_bin_sigmoid.use_fw_regression = True
+        self.w_bin_sigmoid.use_fw_regression = True
+        self.h_bin_sigmoid.use_fw_regression = True
+        
+        # x = x.copy()  # for profiling
+        z = []  # inference output
+        self.training |= self.export
+        for i in range(self.nl):
+            x[i] = self.m[i](self.ia[i](x[i]))  # conv
+            x[i] = self.im[i](x[i])
+            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
+            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+
+            if not self.training:  # inference
+                if self.grid[i].shape[2:4] != x[i].shape[2:4]:
+                    self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
+
+                y = x[i].sigmoid()
+                y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                #y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+                
+
+                #px = (self.x_bin_sigmoid.forward(y[..., 0:12]) + self.grid[i][..., 0]) * self.stride[i]
+                #py = (self.y_bin_sigmoid.forward(y[..., 12:24]) + self.grid[i][..., 1]) * self.stride[i]
+
+                pw = self.w_bin_sigmoid.forward(y[..., 2:24]) * self.anchor_grid[i][..., 0]
+                ph = self.h_bin_sigmoid.forward(y[..., 24:46]) * self.anchor_grid[i][..., 1]
+
+                #y[..., 0] = px
+                #y[..., 1] = py
+                y[..., 2] = pw
+                y[..., 3] = ph
+                
+                y = torch.cat((y[..., 0:4], y[..., 46:]), dim=-1)
+                
+                z.append(y.view(bs, -1, y.shape[-1]))
+
+        return x if self.training else (torch.cat(z, 1), x)
+
+    @staticmethod
+    def _make_grid(nx=20, ny=20):
+        yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
+        return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
+
+
+class Model(nn.Module):
+    def __init__(self, cfg='yolor-csp-c.yaml', ch=3, nc=None, anchors=None):  # model, input channels, number of classes
+        super(Model, self).__init__()
+        self.traced = False
+        if isinstance(cfg, dict):
+            self.yaml = cfg  # model dict
+        else:  # is *.yaml
+            import yaml  # for torch hub
+            self.yaml_file = Path(cfg).name
+            with open(cfg) as f:
+                self.yaml = yaml.load(f, Loader=yaml.SafeLoader)  # model dict
+
+        # Define model
+        ch = self.yaml['ch'] = self.yaml.get('ch', ch)  # input channels
+        if nc and nc != self.yaml['nc']:
+            logger.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}")
+            self.yaml['nc'] = nc  # override yaml value
+        if anchors:
+            logger.info(f'Overriding model.yaml anchors with anchors={anchors}')
+            self.yaml['anchors'] = round(anchors)  # override yaml value
+        self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch])  # model, savelist
+        self.names = [str(i) for i in range(self.yaml['nc'])]  # default names
+        # print([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))])
+
+        # Build strides, anchors
+        m = self.model[-1]  # Detect()
+        if isinstance(m, Detect):
+            s = 256  # 2x min stride
+            m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward
+            check_anchor_order(m)
+            m.anchors /= m.stride.view(-1, 1, 1)
+            self.stride = m.stride
+            self._initialize_biases()  # only run once
+            # print('Strides: %s' % m.stride.tolist())
+        if isinstance(m, IDetect):
+            s = 256  # 2x min stride
+            m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward
+            check_anchor_order(m)
+            m.anchors /= m.stride.view(-1, 1, 1)
+            self.stride = m.stride
+            self._initialize_biases()  # only run once
+            # print('Strides: %s' % m.stride.tolist())
+        if isinstance(m, IAuxDetect):
+            s = 256  # 2x min stride
+            m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))[:4]])  # forward
+            #print(m.stride)
+            check_anchor_order(m)
+            m.anchors /= m.stride.view(-1, 1, 1)
+            self.stride = m.stride
+            self._initialize_aux_biases()  # only run once
+            # print('Strides: %s' % m.stride.tolist())
+        if isinstance(m, IBin):
+            s = 256  # 2x min stride
+            m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward
+            check_anchor_order(m)
+            m.anchors /= m.stride.view(-1, 1, 1)
+            self.stride = m.stride
+            self._initialize_biases_bin()  # only run once
+            # print('Strides: %s' % m.stride.tolist())
+        if isinstance(m, IKeypoint):
+            s = 256  # 2x min stride
+            m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward
+            check_anchor_order(m)
+            m.anchors /= m.stride.view(-1, 1, 1)
+            self.stride = m.stride
+            self._initialize_biases_kpt()  # only run once
+            # print('Strides: %s' % m.stride.tolist())
+
+        # Init weights, biases
+        initialize_weights(self)
+        self.info()
+        logger.info('')
+
+    def forward(self, x, augment=False, profile=False):
+        if augment:
+            img_size = x.shape[-2:]  # height, width
+            s = [1, 0.83, 0.67]  # scales
+            f = [None, 3, None]  # flips (2-ud, 3-lr)
+            y = []  # outputs
+            for si, fi in zip(s, f):
+                xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max()))
+                yi = self.forward_once(xi)[0]  # forward
+                # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1])  # save
+                yi[..., :4] /= si  # de-scale
+                if fi == 2:
+                    yi[..., 1] = img_size[0] - yi[..., 1]  # de-flip ud
+                elif fi == 3:
+                    yi[..., 0] = img_size[1] - yi[..., 0]  # de-flip lr
+                y.append(yi)
+            return torch.cat(y, 1), None  # augmented inference, train
+        else:
+            return self.forward_once(x, profile)  # single-scale inference, train
+
+    def forward_once(self, x, profile=False):
+        y, dt = [], []  # outputs
+        for m in self.model:
+            if m.f != -1:  # if not from previous layer
+                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
+
+            if not hasattr(self, 'traced'):
+                self.traced=False
+
+            if self.traced:
+                if isinstance(m, Detect) or isinstance(m, IDetect) or isinstance(m, IAuxDetect) or isinstance(m, IKeypoint):
+                    break
+
+            if profile:
+                c = isinstance(m, (Detect, IDetect, IAuxDetect, IBin))
+                o = thop.profile(m, inputs=(x.copy() if c else x,), verbose=False)[0] / 1E9 * 2 if thop else 0  # FLOPS
+                for _ in range(10):
+                    m(x.copy() if c else x)
+                t = time_synchronized()
+                for _ in range(10):
+                    m(x.copy() if c else x)
+                dt.append((time_synchronized() - t) * 100)
+                print('%10.1f%10.0f%10.1fms %-40s' % (o, m.np, dt[-1], m.type))
+
+            x = m(x)  # run
+            
+            y.append(x if m.i in self.save else None)  # save output
+
+        if profile:
+            print('%.1fms total' % sum(dt))
+        return x
+
+    def _initialize_biases(self, cf=None):  # initialize biases into Detect(), cf is class frequency
+        # https://arxiv.org/abs/1708.02002 section 3.3
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
+        m = self.model[-1]  # Detect() module
+        for mi, s in zip(m.m, m.stride):  # from
+            b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
+            b.data[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
+            b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
+            mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+
+    def _initialize_aux_biases(self, cf=None):  # initialize biases into Detect(), cf is class frequency
+        # https://arxiv.org/abs/1708.02002 section 3.3
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
+        m = self.model[-1]  # Detect() module
+        for mi, mi2, s in zip(m.m, m.m2, m.stride):  # from
+            b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
+            b.data[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
+            b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
+            mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+            b2 = mi2.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
+            b2.data[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
+            b2.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
+            mi2.bias = torch.nn.Parameter(b2.view(-1), requires_grad=True)
+
+    def _initialize_biases_bin(self, cf=None):  # initialize biases into Detect(), cf is class frequency
+        # https://arxiv.org/abs/1708.02002 section 3.3
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
+        m = self.model[-1]  # Bin() module
+        bc = m.bin_count
+        for mi, s in zip(m.m, m.stride):  # from
+            b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
+            old = b[:, (0,1,2,bc+3)].data
+            obj_idx = 2*bc+4
+            b[:, :obj_idx].data += math.log(0.6 / (bc + 1 - 0.99))
+            b[:, obj_idx].data += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
+            b[:, (obj_idx+1):].data += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
+            b[:, (0,1,2,bc+3)].data = old
+            mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+
+    def _initialize_biases_kpt(self, cf=None):  # initialize biases into Detect(), cf is class frequency
+        # https://arxiv.org/abs/1708.02002 section 3.3
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
+        m = self.model[-1]  # Detect() module
+        for mi, s in zip(m.m, m.stride):  # from
+            b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
+            b.data[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
+            b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
+            mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+
+    def _print_biases(self):
+        m = self.model[-1]  # Detect() module
+        for mi in m.m:  # from
+            b = mi.bias.detach().view(m.na, -1).T  # conv.bias(255) to (3,85)
+            print(('%6g Conv2d.bias:' + '%10.3g' * 6) % (mi.weight.shape[1], *b[:5].mean(1).tolist(), b[5:].mean()))
+
+    # def _print_weights(self):
+    #     for m in self.model.modules():
+    #         if type(m) is Bottleneck:
+    #             print('%10.3g' % (m.w.detach().sigmoid() * 2))  # shortcut weights
+
+    def fuse(self):  # fuse model Conv2d() + BatchNorm2d() layers
+        print('Fusing layers... ')
+        for m in self.model.modules():
+            if isinstance(m, RepConv):
+                #print(f" fuse_repvgg_block")
+                m.fuse_repvgg_block()
+            elif isinstance(m, RepConv_OREPA):
+                #print(f" switch_to_deploy")
+                m.switch_to_deploy()
+            elif type(m) is Conv and hasattr(m, 'bn'):
+                m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
+                delattr(m, 'bn')  # remove batchnorm
+                m.forward = m.fuseforward  # update forward
+            elif isinstance(m, (IDetect, IAuxDetect)):
+                m.fuse()
+                m.forward = m.fuseforward
+        self.info()
+        return self
+
+    def nms(self, mode=True):  # add or remove NMS module
+        present = type(self.model[-1]) is NMS  # last layer is NMS
+        if mode and not present:
+            print('Adding NMS... ')
+            m = NMS()  # module
+            m.f = -1  # from
+            m.i = self.model[-1].i + 1  # index
+            self.model.add_module(name='%s' % m.i, module=m)  # add
+            self.eval()
+        elif not mode and present:
+            print('Removing NMS... ')
+            self.model = self.model[:-1]  # remove
+        return self
+
+    def autoshape(self):  # add autoShape module
+        print('Adding autoShape... ')
+        m = autoShape(self)  # wrap model
+        copy_attr(m, self, include=('yaml', 'nc', 'hyp', 'names', 'stride'), exclude=())  # copy attributes
+        return m
+
+    def info(self, verbose=False, img_size=640):  # print model information
+        model_info(self, verbose, img_size)
+
+
+def parse_model(d, ch):  # model_dict, input_channels(3)
+    logger.info('\n%3s%18s%3s%10s  %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments'))
+    anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple']
+    na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors  # number of anchors
+    no = na * (nc + 5)  # number of outputs = anchors * (classes + 5)
+
+    layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out
+    for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):  # from, number, module, args
+        m = eval(m) if isinstance(m, str) else m  # eval strings
+        for j, a in enumerate(args):
+            try:
+                args[j] = eval(a) if isinstance(a, str) else a  # eval strings
+            except:
+                pass
+
+        n = max(round(n * gd), 1) if n > 1 else n  # depth gain
+        if m in [nn.Conv2d, Conv, RobustConv, RobustConv2, DWConv, GhostConv, RepConv, RepConv_OREPA, DownC, 
+                 SPP, SPPF, SPPCSPC, GhostSPPCSPC, MixConv2d, Focus, Stem, GhostStem, CrossConv, 
+                 Bottleneck, BottleneckCSPA, BottleneckCSPB, BottleneckCSPC, 
+                 RepBottleneck, RepBottleneckCSPA, RepBottleneckCSPB, RepBottleneckCSPC,  
+                 Res, ResCSPA, ResCSPB, ResCSPC, 
+                 RepRes, RepResCSPA, RepResCSPB, RepResCSPC, 
+                 ResX, ResXCSPA, ResXCSPB, ResXCSPC, 
+                 RepResX, RepResXCSPA, RepResXCSPB, RepResXCSPC, 
+                 Ghost, GhostCSPA, GhostCSPB, GhostCSPC,
+                 SwinTransformerBlock, STCSPA, STCSPB, STCSPC,
+                 SwinTransformer2Block, ST2CSPA, ST2CSPB, ST2CSPC]:
+            c1, c2 = ch[f], args[0]
+            if c2 != no:  # if not output
+                c2 = make_divisible(c2 * gw, 8)
+
+            args = [c1, c2, *args[1:]]
+            if m in [DownC, SPPCSPC, GhostSPPCSPC, 
+                     BottleneckCSPA, BottleneckCSPB, BottleneckCSPC, 
+                     RepBottleneckCSPA, RepBottleneckCSPB, RepBottleneckCSPC, 
+                     ResCSPA, ResCSPB, ResCSPC, 
+                     RepResCSPA, RepResCSPB, RepResCSPC, 
+                     ResXCSPA, ResXCSPB, ResXCSPC, 
+                     RepResXCSPA, RepResXCSPB, RepResXCSPC,
+                     GhostCSPA, GhostCSPB, GhostCSPC,
+                     STCSPA, STCSPB, STCSPC,
+                     ST2CSPA, ST2CSPB, ST2CSPC]:
+                args.insert(2, n)  # number of repeats
+                n = 1
+        elif m is nn.BatchNorm2d:
+            args = [ch[f]]
+        elif m is Concat:
+            c2 = sum([ch[x] for x in f])
+        elif m is Chuncat:
+            c2 = sum([ch[x] for x in f])
+        elif m is Shortcut:
+            c2 = ch[f[0]]
+        elif m is Foldcut:
+            c2 = ch[f] // 2
+        elif m in [Detect, IDetect, IAuxDetect, IBin, IKeypoint]:
+            args.append([ch[x] for x in f])
+            if isinstance(args[1], int):  # number of anchors
+                args[1] = [list(range(args[1] * 2))] * len(f)
+        elif m is ReOrg:
+            c2 = ch[f] * 4
+        elif m is Contract:
+            c2 = ch[f] * args[0] ** 2
+        elif m is Expand:
+            c2 = ch[f] // args[0] ** 2
+        else:
+            c2 = ch[f]
+
+        m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args)  # module
+        t = str(m)[8:-2].replace('__main__.', '')  # module type
+        np = sum([x.numel() for x in m_.parameters()])  # number params
+        m_.i, m_.f, m_.type, m_.np = i, f, t, np  # attach index, 'from' index, type, number params
+        logger.info('%3s%18s%3s%10.0f  %-40s%-30s' % (i, f, n, np, t, args))  # print
+        save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist
+        layers.append(m_)
+        if i == 0:
+            ch = []
+        ch.append(c2)
+    return nn.Sequential(*layers), sorted(save)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--cfg', type=str, default='yolor-csp-c.yaml', help='model.yaml')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--profile', action='store_true', help='profile model speed')
+    opt = parser.parse_args()
+    opt.cfg = check_file(opt.cfg)  # check file
+    set_logging()
+    device = select_device(opt.device)
+
+    # Create model
+    model = Model(opt.cfg).to(device)
+    model.train()
+    
+    if opt.profile:
+        img = torch.rand(1, 3, 640, 640).to(device)
+        y = model(img, profile=True)
+
+    # Profile
+    # img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 640, 640).to(device)
+    # y = model(img, profile=True)
+
+    # Tensorboard
+    # from torch.utils.tensorboard import SummaryWriter
+    # tb_writer = SummaryWriter()
+    # print("Run 'tensorboard --logdir=models/runs' to view tensorboard at http://localhost:6006/")
+    # tb_writer.add_graph(model.model, img)  # add model to tensorboard
+    # tb_writer.add_image('test', img[0], dataformats='CWH')  # add model to tensorboard

utils/__init__.py

@@ -0,0 +1 @@
+# init

utils/activations.py

@@ -0,0 +1,72 @@
+# Activation functions
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+# SiLU https://arxiv.org/pdf/1606.08415.pdf ----------------------------------------------------------------------------
+class SiLU(nn.Module):  # export-friendly version of nn.SiLU()
+    @staticmethod
+    def forward(x):
+        return x * torch.sigmoid(x)
+
+
+class Hardswish(nn.Module):  # export-friendly version of nn.Hardswish()
+    @staticmethod
+    def forward(x):
+        # return x * F.hardsigmoid(x)  # for torchscript and CoreML
+        return x * F.hardtanh(x + 3, 0., 6.) / 6.  # for torchscript, CoreML and ONNX
+
+
+class MemoryEfficientSwish(nn.Module):
+    class F(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x):
+            ctx.save_for_backward(x)
+            return x * torch.sigmoid(x)
+
+        @staticmethod
+        def backward(ctx, grad_output):
+            x = ctx.saved_tensors[0]
+            sx = torch.sigmoid(x)
+            return grad_output * (sx * (1 + x * (1 - sx)))
+
+    def forward(self, x):
+        return self.F.apply(x)
+
+
+# Mish https://github.com/digantamisra98/Mish --------------------------------------------------------------------------
+class Mish(nn.Module):
+    @staticmethod
+    def forward(x):
+        return x * F.softplus(x).tanh()
+
+
+class MemoryEfficientMish(nn.Module):
+    class F(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x):
+            ctx.save_for_backward(x)
+            return x.mul(torch.tanh(F.softplus(x)))  # x * tanh(ln(1 + exp(x)))
+
+        @staticmethod
+        def backward(ctx, grad_output):
+            x = ctx.saved_tensors[0]
+            sx = torch.sigmoid(x)
+            fx = F.softplus(x).tanh()
+            return grad_output * (fx + x * sx * (1 - fx * fx))
+
+    def forward(self, x):
+        return self.F.apply(x)
+
+
+# FReLU https://arxiv.org/abs/2007.11824 -------------------------------------------------------------------------------
+class FReLU(nn.Module):
+    def __init__(self, c1, k=3):  # ch_in, kernel
+        super().__init__()
+        self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False)
+        self.bn = nn.BatchNorm2d(c1)
+
+    def forward(self, x):
+        return torch.max(x, self.bn(self.conv(x)))

utils/add_nms.py

@@ -0,0 +1,155 @@
+import numpy as np
+import onnx
+from onnx import shape_inference
+try:
+    import onnx_graphsurgeon as gs
+except Exception as e:
+    print('Import onnx_graphsurgeon failure: %s' % e)
+
+import logging
+
+LOGGER = logging.getLogger(__name__)
+
+class RegisterNMS(object):
+    def __init__(
+        self,
+        onnx_model_path: str,
+        precision: str = "fp32",
+    ):
+
+        self.graph = gs.import_onnx(onnx.load(onnx_model_path))
+        assert self.graph
+        LOGGER.info("ONNX graph created successfully")
+        # Fold constants via ONNX-GS that PyTorch2ONNX may have missed
+        self.graph.fold_constants()
+        self.precision = precision
+        self.batch_size = 1
+    def infer(self):
+        """
+        Sanitize the graph by cleaning any unconnected nodes, do a topological resort,
+        and fold constant inputs values. When possible, run shape inference on the
+        ONNX graph to determine tensor shapes.
+        """
+        for _ in range(3):
+            count_before = len(self.graph.nodes)
+
+            self.graph.cleanup().toposort()
+            try:
+                for node in self.graph.nodes:
+                    for o in node.outputs:
+                        o.shape = None
+                model = gs.export_onnx(self.graph)
+                model = shape_inference.infer_shapes(model)
+                self.graph = gs.import_onnx(model)
+            except Exception as e:
+                LOGGER.info(f"Shape inference could not be performed at this time:\n{e}")
+            try:
+                self.graph.fold_constants(fold_shapes=True)
+            except TypeError as e:
+                LOGGER.error(
+                    "This version of ONNX GraphSurgeon does not support folding shapes, "
+                    f"please upgrade your onnx_graphsurgeon module. Error:\n{e}"
+                )
+                raise
+
+            count_after = len(self.graph.nodes)
+            if count_before == count_after:
+                # No new folding occurred in this iteration, so we can stop for now.
+                break
+
+    def save(self, output_path):
+        """
+        Save the ONNX model to the given location.
+        Args:
+            output_path: Path pointing to the location where to write
+                out the updated ONNX model.
+        """
+        self.graph.cleanup().toposort()
+        model = gs.export_onnx(self.graph)
+        onnx.save(model, output_path)
+        LOGGER.info(f"Saved ONNX model to {output_path}")
+
+    def register_nms(
+        self,
+        *,
+        score_thresh: float = 0.25,
+        nms_thresh: float = 0.45,
+        detections_per_img: int = 100,
+    ):
+        """
+        Register the ``EfficientNMS_TRT`` plugin node.
+        NMS expects these shapes for its input tensors:
+            - box_net: [batch_size, number_boxes, 4]
+            - class_net: [batch_size, number_boxes, number_labels]
+        Args:
+            score_thresh (float): The scalar threshold for score (low scoring boxes are removed).
+            nms_thresh (float): The scalar threshold for IOU (new boxes that have high IOU
+                overlap with previously selected boxes are removed).
+            detections_per_img (int): Number of best detections to keep after NMS.
+        """
+
+        self.infer()
+        # Find the concat node at the end of the network
+        op_inputs = self.graph.outputs
+        op = "EfficientNMS_TRT"
+        attrs = {
+            "plugin_version": "1",
+            "background_class": -1,  # no background class
+            "max_output_boxes": detections_per_img,
+            "score_threshold": score_thresh,
+            "iou_threshold": nms_thresh,
+            "score_activation": False,
+            "box_coding": 0,
+        }
+
+        if self.precision == "fp32":
+            dtype_output = np.float32
+        elif self.precision == "fp16":
+            dtype_output = np.float16
+        else:
+            raise NotImplementedError(f"Currently not supports precision: {self.precision}")
+
+        # NMS Outputs
+        output_num_detections = gs.Variable(
+            name="num_dets",
+            dtype=np.int32,
+            shape=[self.batch_size, 1],
+        )  # A scalar indicating the number of valid detections per batch image.
+        output_boxes = gs.Variable(
+            name="det_boxes",
+            dtype=dtype_output,
+            shape=[self.batch_size, detections_per_img, 4],
+        )
+        output_scores = gs.Variable(
+            name="det_scores",
+            dtype=dtype_output,
+            shape=[self.batch_size, detections_per_img],
+        )
+        output_labels = gs.Variable(
+            name="det_classes",
+            dtype=np.int32,
+            shape=[self.batch_size, detections_per_img],
+        )
+
+        op_outputs = [output_num_detections, output_boxes, output_scores, output_labels]
+
+        # Create the NMS Plugin node with the selected inputs. The outputs of the node will also
+        # become the final outputs of the graph.
+        self.graph.layer(op=op, name="batched_nms", inputs=op_inputs, outputs=op_outputs, attrs=attrs)
+        LOGGER.info(f"Created NMS plugin '{op}' with attributes: {attrs}")
+
+        self.graph.outputs = op_outputs
+
+        self.infer()
+
+    def save(self, output_path):
+        """
+        Save the ONNX model to the given location.
+        Args:
+            output_path: Path pointing to the location where to write
+                out the updated ONNX model.
+        """
+        self.graph.cleanup().toposort()
+        model = gs.export_onnx(self.graph)
+        onnx.save(model, output_path)
+        LOGGER.info(f"Saved ONNX model to {output_path}")

utils/autoanchor.py

@@ -0,0 +1,160 @@
+# Auto-anchor utils
+
+import numpy as np
+import torch
+import yaml
+from scipy.cluster.vq import kmeans
+from tqdm import tqdm
+
+from utils.general import colorstr
+
+
+def check_anchor_order(m):
+    # Check anchor order against stride order for YOLO Detect() module m, and correct if necessary
+    a = m.anchor_grid.prod(-1).view(-1)  # anchor area
+    da = a[-1] - a[0]  # delta a
+    ds = m.stride[-1] - m.stride[0]  # delta s
+    if da.sign() != ds.sign():  # same order
+        print('Reversing anchor order')
+        m.anchors[:] = m.anchors.flip(0)
+        m.anchor_grid[:] = m.anchor_grid.flip(0)
+
+
+def check_anchors(dataset, model, thr=4.0, imgsz=640):
+    # Check anchor fit to data, recompute if necessary
+    prefix = colorstr('autoanchor: ')
+    print(f'\n{prefix}Analyzing anchors... ', end='')
+    m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1]  # Detect()
+    shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
+    scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1))  # augment scale
+    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float()  # wh
+
+    def metric(k):  # compute metric
+        r = wh[:, None] / k[None]
+        x = torch.min(r, 1. / r).min(2)[0]  # ratio metric
+        best = x.max(1)[0]  # best_x
+        aat = (x > 1. / thr).float().sum(1).mean()  # anchors above threshold
+        bpr = (best > 1. / thr).float().mean()  # best possible recall
+        return bpr, aat
+
+    anchors = m.anchor_grid.clone().cpu().view(-1, 2)  # current anchors
+    bpr, aat = metric(anchors)
+    print(f'anchors/target = {aat:.2f}, Best Possible Recall (BPR) = {bpr:.4f}', end='')
+    if bpr < 0.98:  # threshold to recompute
+        print('. Attempting to improve anchors, please wait...')
+        na = m.anchor_grid.numel() // 2  # number of anchors
+        try:
+            anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False)
+        except Exception as e:
+            print(f'{prefix}ERROR: {e}')
+        new_bpr = metric(anchors)[0]
+        if new_bpr > bpr:  # replace anchors
+            anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)
+            m.anchor_grid[:] = anchors.clone().view_as(m.anchor_grid)  # for inference
+            check_anchor_order(m)
+            m.anchors[:] = anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1)  # loss
+            print(f'{prefix}New anchors saved to model. Update model *.yaml to use these anchors in the future.')
+        else:
+            print(f'{prefix}Original anchors better than new anchors. Proceeding with original anchors.')
+    print('')  # newline
+
+
+def kmean_anchors(path='./data/coco.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True):
+    """ Creates kmeans-evolved anchors from training dataset
+
+        Arguments:
+            path: path to dataset *.yaml, or a loaded dataset
+            n: number of anchors
+            img_size: image size used for training
+            thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0
+            gen: generations to evolve anchors using genetic algorithm
+            verbose: print all results
+
+        Return:
+            k: kmeans evolved anchors
+
+        Usage:
+            from utils.autoanchor import *; _ = kmean_anchors()
+    """
+    thr = 1. / thr
+    prefix = colorstr('autoanchor: ')
+
+    def metric(k, wh):  # compute metrics
+        r = wh[:, None] / k[None]
+        x = torch.min(r, 1. / r).min(2)[0]  # ratio metric
+        # x = wh_iou(wh, torch.tensor(k))  # iou metric
+        return x, x.max(1)[0]  # x, best_x
+
+    def anchor_fitness(k):  # mutation fitness
+        _, best = metric(torch.tensor(k, dtype=torch.float32), wh)
+        return (best * (best > thr).float()).mean()  # fitness
+
+    def print_results(k):
+        k = k[np.argsort(k.prod(1))]  # sort small to large
+        x, best = metric(k, wh0)
+        bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n  # best possible recall, anch > thr
+        print(f'{prefix}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr')
+        print(f'{prefix}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, '
+              f'past_thr={x[x > thr].mean():.3f}-mean: ', end='')
+        for i, x in enumerate(k):
+            print('%i,%i' % (round(x[0]), round(x[1])), end=',  ' if i < len(k) - 1 else '\n')  # use in *.cfg
+        return k
+
+    if isinstance(path, str):  # *.yaml file
+        with open(path) as f:
+            data_dict = yaml.load(f, Loader=yaml.SafeLoader)  # model dict
+        from utils.datasets import LoadImagesAndLabels
+        dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True)
+    else:
+        dataset = path  # dataset
+
+    # Get label wh
+    shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True)
+    wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])  # wh
+
+    # Filter
+    i = (wh0 < 3.0).any(1).sum()
+    if i:
+        print(f'{prefix}WARNING: Extremely small objects found. {i} of {len(wh0)} labels are < 3 pixels in size.')
+    wh = wh0[(wh0 >= 2.0).any(1)]  # filter > 2 pixels
+    # wh = wh * (np.random.rand(wh.shape[0], 1) * 0.9 + 0.1)  # multiply by random scale 0-1
+
+    # Kmeans calculation
+    print(f'{prefix}Running kmeans for {n} anchors on {len(wh)} points...')
+    s = wh.std(0)  # sigmas for whitening
+    k, dist = kmeans(wh / s, n, iter=30)  # points, mean distance
+    assert len(k) == n, print(f'{prefix}ERROR: scipy.cluster.vq.kmeans requested {n} points but returned only {len(k)}')
+    k *= s
+    wh = torch.tensor(wh, dtype=torch.float32)  # filtered
+    wh0 = torch.tensor(wh0, dtype=torch.float32)  # unfiltered
+    k = print_results(k)
+
+    # Plot
+    # k, d = [None] * 20, [None] * 20
+    # for i in tqdm(range(1, 21)):
+    #     k[i-1], d[i-1] = kmeans(wh / s, i)  # points, mean distance
+    # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True)
+    # ax = ax.ravel()
+    # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.')
+    # fig, ax = plt.subplots(1, 2, figsize=(14, 7))  # plot wh
+    # ax[0].hist(wh[wh[:, 0]<100, 0],400)
+    # ax[1].hist(wh[wh[:, 1]<100, 1],400)
+    # fig.savefig('wh.png', dpi=200)
+
+    # Evolve
+    npr = np.random
+    f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1  # fitness, generations, mutation prob, sigma
+    pbar = tqdm(range(gen), desc=f'{prefix}Evolving anchors with Genetic Algorithm:')  # progress bar
+    for _ in pbar:
+        v = np.ones(sh)
+        while (v == 1).all():  # mutate until a change occurs (prevent duplicates)
+            v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0)
+        kg = (k.copy() * v).clip(min=2.0)
+        fg = anchor_fitness(kg)
+        if fg > f:
+            f, k = fg, kg.copy()
+            pbar.desc = f'{prefix}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}'
+            if verbose:
+                print_results(k)
+
+    return print_results(k)

utils/aws/__init__.py

@@ -0,0 +1 @@
+#init

utils/aws/mime.sh

@@ -0,0 +1,26 @@
+# AWS EC2 instance startup 'MIME' script https://aws.amazon.com/premiumsupport/knowledge-center/execute-user-data-ec2/
+# This script will run on every instance restart, not only on first start
+# --- DO NOT COPY ABOVE COMMENTS WHEN PASTING INTO USERDATA ---
+
+Content-Type: multipart/mixed; boundary="//"
+MIME-Version: 1.0
+
+--//
+Content-Type: text/cloud-config; charset="us-ascii"
+MIME-Version: 1.0
+Content-Transfer-Encoding: 7bit
+Content-Disposition: attachment; filename="cloud-config.txt"
+
+#cloud-config
+cloud_final_modules:
+- [scripts-user, always]
+
+--//
+Content-Type: text/x-shellscript; charset="us-ascii"
+MIME-Version: 1.0
+Content-Transfer-Encoding: 7bit
+Content-Disposition: attachment; filename="userdata.txt"
+
+#!/bin/bash
+# --- paste contents of userdata.sh here ---
+--//

utils/aws/resume.py

@@ -0,0 +1,37 @@
+# Resume all interrupted trainings in yolor/ dir including DDP trainings
+# Usage: $ python utils/aws/resume.py
+
+import os
+import sys
+from pathlib import Path
+
+import torch
+import yaml
+
+sys.path.append('./')  # to run '$ python *.py' files in subdirectories
+
+port = 0  # --master_port
+path = Path('').resolve()
+for last in path.rglob('*/**/last.pt'):
+    ckpt = torch.load(last)
+    if ckpt['optimizer'] is None:
+        continue
+
+    # Load opt.yaml
+    with open(last.parent.parent / 'opt.yaml') as f:
+        opt = yaml.load(f, Loader=yaml.SafeLoader)
+
+    # Get device count
+    d = opt['device'].split(',')  # devices
+    nd = len(d)  # number of devices
+    ddp = nd > 1 or (nd == 0 and torch.cuda.device_count() > 1)  # distributed data parallel
+
+    if ddp:  # multi-GPU
+        port += 1
+        cmd = f'python -m torch.distributed.launch --nproc_per_node {nd} --master_port {port} train.py --resume {last}'
+    else:  # single-GPU
+        cmd = f'python train.py --resume {last}'
+
+    cmd += ' > /dev/null 2>&1 &'  # redirect output to dev/null and run in daemon thread
+    print(cmd)
+    os.system(cmd)

utils/aws/userdata.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+# AWS EC2 instance startup script https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/user-data.html
+# This script will run only once on first instance start (for a re-start script see mime.sh)
+# /home/ubuntu (ubuntu) or /home/ec2-user (amazon-linux) is working dir
+# Use >300 GB SSD
+
+cd home/ubuntu
+if [ ! -d yolor ]; then
+  echo "Running first-time script." # install dependencies, download COCO, pull Docker
+  git clone -b main https://github.com/WongKinYiu/yolov7 && sudo chmod -R 777 yolov7
+  cd yolov7
+  bash data/scripts/get_coco.sh && echo "Data done." &
+  sudo docker pull nvcr.io/nvidia/pytorch:21.08-py3 && echo "Docker done." &
+  python -m pip install --upgrade pip && pip install -r requirements.txt && python detect.py && echo "Requirements done." &
+  wait && echo "All tasks done." # finish background tasks
+else
+  echo "Running re-start script." # resume interrupted runs
+  i=0
+  list=$(sudo docker ps -qa) # container list i.e. $'one\ntwo\nthree\nfour'
+  while IFS= read -r id; do
+    ((i++))
+    echo "restarting container $i: $id"
+    sudo docker start $id
+    # sudo docker exec -it $id python train.py --resume # single-GPU
+    sudo docker exec -d $id python utils/aws/resume.py # multi-scenario
+  done <<<"$list"
+fi

utils/datasets.py

@@ -0,0 +1,1320 @@
+# Dataset utils and dataloaders
+
+import glob
+import logging
+import math
+import os
+import random
+import shutil
+import time
+from itertools import repeat
+from multiprocessing.pool import ThreadPool
+from pathlib import Path
+from threading import Thread
+
+import cv2
+import numpy as np
+import torch
+import torch.nn.functional as F
+from PIL import Image, ExifTags
+from torch.utils.data import Dataset
+from tqdm import tqdm
+
+import pickle
+from copy import deepcopy
+#from pycocotools import mask as maskUtils
+from torchvision.utils import save_image
+from torchvision.ops import roi_pool, roi_align, ps_roi_pool, ps_roi_align
+
+from utils.general import check_requirements, xyxy2xywh, xywh2xyxy, xywhn2xyxy, xyn2xy, segment2box, segments2boxes, \
+    resample_segments, clean_str
+from utils.torch_utils import torch_distributed_zero_first
+
+# Parameters
+help_url = 'https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data'
+img_formats = ['bmp', 'jpg', 'jpeg', 'png', 'tif', 'tiff', 'dng', 'webp', 'mpo']  # acceptable image suffixes
+vid_formats = ['mov', 'avi', 'mp4', 'mpg', 'mpeg', 'm4v', 'wmv', 'mkv']  # acceptable video suffixes
+logger = logging.getLogger(__name__)
+
+# Get orientation exif tag
+for orientation in ExifTags.TAGS.keys():
+    if ExifTags.TAGS[orientation] == 'Orientation':
+        break
+
+
+def get_hash(files):
+    # Returns a single hash value of a list of files
+    return sum(os.path.getsize(f) for f in files if os.path.isfile(f))
+
+
+def exif_size(img):
+    # Returns exif-corrected PIL size
+    s = img.size  # (width, height)
+    try:
+        rotation = dict(img._getexif().items())[orientation]
+        if rotation == 6:  # rotation 270
+            s = (s[1], s[0])
+        elif rotation == 8:  # rotation 90
+            s = (s[1], s[0])
+    except:
+        pass
+
+    return s
+
+
+def create_dataloader(path, imgsz, batch_size, stride, opt, hyp=None, augment=False, cache=False, pad=0.0, rect=False,
+                      rank=-1, world_size=1, workers=8, image_weights=False, quad=False, prefix=''):
+    # Make sure only the first process in DDP process the dataset first, and the following others can use the cache
+    with torch_distributed_zero_first(rank):
+        dataset = LoadImagesAndLabels(path, imgsz, batch_size,
+                                      augment=augment,  # augment images
+                                      hyp=hyp,  # augmentation hyperparameters
+                                      rect=rect,  # rectangular training
+                                      cache_images=cache,
+                                      single_cls=opt.single_cls,
+                                      stride=int(stride),
+                                      pad=pad,
+                                      image_weights=image_weights,
+                                      prefix=prefix)
+
+    batch_size = min(batch_size, len(dataset))
+    nw = min([os.cpu_count() // world_size, batch_size if batch_size > 1 else 0, workers])  # number of workers
+    sampler = torch.utils.data.distributed.DistributedSampler(dataset) if rank != -1 else None
+    loader = torch.utils.data.DataLoader if image_weights else InfiniteDataLoader
+    # Use torch.utils.data.DataLoader() if dataset.properties will update during training else InfiniteDataLoader()
+    dataloader = loader(dataset,
+                        batch_size=batch_size,
+                        num_workers=nw,
+                        sampler=sampler,
+                        pin_memory=True,
+                        collate_fn=LoadImagesAndLabels.collate_fn4 if quad else LoadImagesAndLabels.collate_fn)
+    return dataloader, dataset
+
+
+class InfiniteDataLoader(torch.utils.data.dataloader.DataLoader):
+    """ Dataloader that reuses workers
+
+    Uses same syntax as vanilla DataLoader
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        object.__setattr__(self, 'batch_sampler', _RepeatSampler(self.batch_sampler))
+        self.iterator = super().__iter__()
+
+    def __len__(self):
+        return len(self.batch_sampler.sampler)
+
+    def __iter__(self):
+        for i in range(len(self)):
+            yield next(self.iterator)
+
+
+class _RepeatSampler(object):
+    """ Sampler that repeats forever
+
+    Args:
+        sampler (Sampler)
+    """
+
+    def __init__(self, sampler):
+        self.sampler = sampler
+
+    def __iter__(self):
+        while True:
+            yield from iter(self.sampler)
+
+
+class LoadImages:  # for inference
+    def __init__(self, path, img_size=640, stride=32):
+        p = str(Path(path).absolute())  # os-agnostic absolute path
+        if '*' in p:
+            files = sorted(glob.glob(p, recursive=True))  # glob
+        elif os.path.isdir(p):
+            files = sorted(glob.glob(os.path.join(p, '*.*')))  # dir
+        elif os.path.isfile(p):
+            files = [p]  # files
+        else:
+            raise Exception(f'ERROR: {p} does not exist')
+
+        images = [x for x in files if x.split('.')[-1].lower() in img_formats]
+        videos = [x for x in files if x.split('.')[-1].lower() in vid_formats]
+        ni, nv = len(images), len(videos)
+
+        self.img_size = img_size
+        self.stride = stride
+        self.files = images + videos
+        self.nf = ni + nv  # number of files
+        self.video_flag = [False] * ni + [True] * nv
+        self.mode = 'image'
+        if any(videos):
+            self.new_video(videos[0])  # new video
+        else:
+            self.cap = None
+        assert self.nf > 0, f'No images or videos found in {p}. ' \
+                            f'Supported formats are:\nimages: {img_formats}\nvideos: {vid_formats}'
+
+    def __iter__(self):
+        self.count = 0
+        return self
+
+    def __next__(self):
+        if self.count == self.nf:
+            raise StopIteration
+        path = self.files[self.count]
+
+        if self.video_flag[self.count]:
+            # Read video
+            self.mode = 'video'
+            ret_val, img0 = self.cap.read()
+            if not ret_val:
+                self.count += 1
+                self.cap.release()
+                if self.count == self.nf:  # last video
+                    raise StopIteration
+                else:
+                    path = self.files[self.count]
+                    self.new_video(path)
+                    ret_val, img0 = self.cap.read()
+
+            self.frame += 1
+            print(f'video {self.count + 1}/{self.nf} ({self.frame}/{self.nframes}) {path}: ', end='')
+
+        else:
+            # Read image
+            self.count += 1
+            img0 = cv2.imread(path)  # BGR
+            assert img0 is not None, 'Image Not Found ' + path
+            #print(f'image {self.count}/{self.nf} {path}: ', end='')
+
+        # Padded resize
+        img = letterbox(img0, self.img_size, stride=self.stride)[0]
+
+        # Convert
+        img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
+        img = np.ascontiguousarray(img)
+
+        return path, img, img0, self.cap
+
+    def new_video(self, path):
+        self.frame = 0
+        self.cap = cv2.VideoCapture(path)
+        self.nframes = int(self.cap.get(cv2.CAP_PROP_FRAME_COUNT))
+
+    def __len__(self):
+        return self.nf  # number of files
+
+
+class LoadWebcam:  # for inference
+    def __init__(self, pipe='0', img_size=640, stride=32):
+        self.img_size = img_size
+        self.stride = stride
+
+        if pipe.isnumeric():
+            pipe = eval(pipe)  # local camera
+        # pipe = 'rtsp://192.168.1.64/1'  # IP camera
+        # pipe = 'rtsp://username:password@192.168.1.64/1'  # IP camera with login
+        # pipe = 'http://wmccpinetop.axiscam.net/mjpg/video.mjpg'  # IP golf camera
+
+        self.pipe = pipe
+        self.cap = cv2.VideoCapture(pipe)  # video capture object
+        self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 3)  # set buffer size
+
+    def __iter__(self):
+        self.count = -1
+        return self
+
+    def __next__(self):
+        self.count += 1
+        if cv2.waitKey(1) == ord('q'):  # q to quit
+            self.cap.release()
+            cv2.destroyAllWindows()
+            raise StopIteration
+
+        # Read frame
+        if self.pipe == 0:  # local camera
+            ret_val, img0 = self.cap.read()
+            img0 = cv2.flip(img0, 1)  # flip left-right
+        else:  # IP camera
+            n = 0
+            while True:
+                n += 1
+                self.cap.grab()
+                if n % 30 == 0:  # skip frames
+                    ret_val, img0 = self.cap.retrieve()
+                    if ret_val:
+                        break
+
+        # Print
+        assert ret_val, f'Camera Error {self.pipe}'
+        img_path = 'webcam.jpg'
+        print(f'webcam {self.count}: ', end='')
+
+        # Padded resize
+        img = letterbox(img0, self.img_size, stride=self.stride)[0]
+
+        # Convert
+        img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
+        img = np.ascontiguousarray(img)
+
+        return img_path, img, img0, None
+
+    def __len__(self):
+        return 0
+
+
+class LoadStreams:  # multiple IP or RTSP cameras
+    def __init__(self, sources='streams.txt', img_size=640, stride=32):
+        self.mode = 'stream'
+        self.img_size = img_size
+        self.stride = stride
+
+        if os.path.isfile(sources):
+            with open(sources, 'r') as f:
+                sources = [x.strip() for x in f.read().strip().splitlines() if len(x.strip())]
+        else:
+            sources = [sources]
+
+        n = len(sources)
+        self.imgs = [None] * n
+        self.sources = [clean_str(x) for x in sources]  # clean source names for later
+        for i, s in enumerate(sources):
+            # Start the thread to read frames from the video stream
+            print(f'{i + 1}/{n}: {s}... ', end='')
+            url = eval(s) if s.isnumeric() else s
+            if 'youtube.com/' in str(url) or 'youtu.be/' in str(url):  # if source is YouTube video
+                check_requirements(('pafy', 'youtube_dl'))
+                import pafy
+                url = pafy.new(url).getbest(preftype="mp4").url
+            cap = cv2.VideoCapture(url)
+            assert cap.isOpened(), f'Failed to open {s}'
+            w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+            h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+            self.fps = cap.get(cv2.CAP_PROP_FPS) % 100
+
+            _, self.imgs[i] = cap.read()  # guarantee first frame
+            thread = Thread(target=self.update, args=([i, cap]), daemon=True)
+            print(f' success ({w}x{h} at {self.fps:.2f} FPS).')
+            thread.start()
+        print('')  # newline
+
+        # check for common shapes
+        s = np.stack([letterbox(x, self.img_size, stride=self.stride)[0].shape for x in self.imgs], 0)  # shapes
+        self.rect = np.unique(s, axis=0).shape[0] == 1  # rect inference if all shapes equal
+        if not self.rect:
+            print('WARNING: Different stream shapes detected. For optimal performance supply similarly-shaped streams.')
+
+    def update(self, index, cap):
+        # Read next stream frame in a daemon thread
+        n = 0
+        while cap.isOpened():
+            n += 1
+            # _, self.imgs[index] = cap.read()
+            cap.grab()
+            if n == 4:  # read every 4th frame
+                success, im = cap.retrieve()
+                self.imgs[index] = im if success else self.imgs[index] * 0
+                n = 0
+            time.sleep(1 / self.fps)  # wait time
+
+    def __iter__(self):
+        self.count = -1
+        return self
+
+    def __next__(self):
+        self.count += 1
+        img0 = self.imgs.copy()
+        if cv2.waitKey(1) == ord('q'):  # q to quit
+            cv2.destroyAllWindows()
+            raise StopIteration
+
+        # Letterbox
+        img = [letterbox(x, self.img_size, auto=self.rect, stride=self.stride)[0] for x in img0]
+
+        # Stack
+        img = np.stack(img, 0)
+
+        # Convert
+        img = img[:, :, :, ::-1].transpose(0, 3, 1, 2)  # BGR to RGB, to bsx3x416x416
+        img = np.ascontiguousarray(img)
+
+        return self.sources, img, img0, None
+
+    def __len__(self):
+        return 0  # 1E12 frames = 32 streams at 30 FPS for 30 years
+
+
+def img2label_paths(img_paths):
+    # Define label paths as a function of image paths
+    sa, sb = os.sep + 'images' + os.sep, os.sep + 'labels' + os.sep  # /images/, /labels/ substrings
+    return ['txt'.join(x.replace(sa, sb, 1).rsplit(x.split('.')[-1], 1)) for x in img_paths]
+
+
+class LoadImagesAndLabels(Dataset):  # for training/testing
+    def __init__(self, path, img_size=640, batch_size=16, augment=False, hyp=None, rect=False, image_weights=False,
+                 cache_images=False, single_cls=False, stride=32, pad=0.0, prefix=''):
+        self.img_size = img_size
+        self.augment = augment
+        self.hyp = hyp
+        self.image_weights = image_weights
+        self.rect = False if image_weights else rect
+        self.mosaic = self.augment and not self.rect  # load 4 images at a time into a mosaic (only during training)
+        self.mosaic_border = [-img_size // 2, -img_size // 2]
+        self.stride = stride
+        self.path = path        
+        #self.albumentations = Albumentations() if augment else None
+
+        try:
+            f = []  # image files
+            for p in path if isinstance(path, list) else [path]:
+                p = Path(p)  # os-agnostic
+                if p.is_dir():  # dir
+                    f += glob.glob(str(p / '**' / '*.*'), recursive=True)
+                    # f = list(p.rglob('**/*.*'))  # pathlib
+                elif p.is_file():  # file
+                    with open(p, 'r') as t:
+                        t = t.read().strip().splitlines()
+                        parent = str(p.parent) + os.sep
+                        f += [x.replace('./', parent) if x.startswith('./') else x for x in t]  # local to global path
+                        # f += [p.parent / x.lstrip(os.sep) for x in t]  # local to global path (pathlib)
+                else:
+                    raise Exception(f'{prefix}{p} does not exist')
+            self.img_files = sorted([x.replace('/', os.sep) for x in f if x.split('.')[-1].lower() in img_formats])
+            # self.img_files = sorted([x for x in f if x.suffix[1:].lower() in img_formats])  # pathlib
+            assert self.img_files, f'{prefix}No images found'
+        except Exception as e:
+            raise Exception(f'{prefix}Error loading data from {path}: {e}\nSee {help_url}')
+
+        # Check cache
+        self.label_files = img2label_paths(self.img_files)  # labels
+        cache_path = (p if p.is_file() else Path(self.label_files[0]).parent).with_suffix('.cache')  # cached labels
+        if cache_path.is_file():
+            cache, exists = torch.load(cache_path), True  # load
+            #if cache['hash'] != get_hash(self.label_files + self.img_files) or 'version' not in cache:  # changed
+            #    cache, exists = self.cache_labels(cache_path, prefix), False  # re-cache
+        else:
+            cache, exists = self.cache_labels(cache_path, prefix), False  # cache
+
+        # Display cache
+        nf, nm, ne, nc, n = cache.pop('results')  # found, missing, empty, corrupted, total
+        if exists:
+            d = f"Scanning '{cache_path}' images and labels... {nf} found, {nm} missing, {ne} empty, {nc} corrupted"
+            tqdm(None, desc=prefix + d, total=n, initial=n)  # display cache results
+        assert nf > 0 or not augment, f'{prefix}No labels in {cache_path}. Can not train without labels. See {help_url}'
+
+        # Read cache
+        cache.pop('hash')  # remove hash
+        cache.pop('version')  # remove version
+        labels, shapes, self.segments = zip(*cache.values())
+        self.labels = list(labels)
+        self.shapes = np.array(shapes, dtype=np.float64)
+        self.img_files = list(cache.keys())  # update
+        self.label_files = img2label_paths(cache.keys())  # update
+        if single_cls:
+            for x in self.labels:
+                x[:, 0] = 0
+
+        n = len(shapes)  # number of images
+        bi = np.floor(np.arange(n) / batch_size).astype(int)  # batch index
+        nb = bi[-1] + 1  # number of batches
+        self.batch = bi  # batch index of image
+        self.n = n
+        self.indices = range(n)
+
+        # Rectangular Training
+        if self.rect:
+            # Sort by aspect ratio
+            s = self.shapes  # wh
+            ar = s[:, 1] / s[:, 0]  # aspect ratio
+            irect = ar.argsort()
+            self.img_files = [self.img_files[i] for i in irect]
+            self.label_files = [self.label_files[i] for i in irect]
+            self.labels = [self.labels[i] for i in irect]
+            self.shapes = s[irect]  # wh
+            ar = ar[irect]
+
+            # Set training image shapes
+            shapes = [[1, 1]] * nb
+            for i in range(nb):
+                ari = ar[bi == i]
+                mini, maxi = ari.min(), ari.max()
+                if maxi < 1:
+                    shapes[i] = [maxi, 1]
+                elif mini > 1:
+                    shapes[i] = [1, 1 / mini]
+
+            self.batch_shapes = np.ceil(np.array(shapes) * img_size / stride + pad).astype(int) * stride
+
+        # Cache images into memory for faster training (WARNING: large datasets may exceed system RAM)
+        self.imgs = [None] * n
+        if cache_images:
+            if cache_images == 'disk':
+                self.im_cache_dir = Path(Path(self.img_files[0]).parent.as_posix() + '_npy')
+                self.img_npy = [self.im_cache_dir / Path(f).with_suffix('.npy').name for f in self.img_files]
+                self.im_cache_dir.mkdir(parents=True, exist_ok=True)
+            gb = 0  # Gigabytes of cached images
+            self.img_hw0, self.img_hw = [None] * n, [None] * n
+            results = ThreadPool(8).imap(lambda x: load_image(*x), zip(repeat(self), range(n)))
+            pbar = tqdm(enumerate(results), total=n)
+            for i, x in pbar:
+                if cache_images == 'disk':
+                    if not self.img_npy[i].exists():
+                        np.save(self.img_npy[i].as_posix(), x[0])
+                    gb += self.img_npy[i].stat().st_size
+                else:
+                    self.imgs[i], self.img_hw0[i], self.img_hw[i] = x
+                    gb += self.imgs[i].nbytes
+                pbar.desc = f'{prefix}Caching images ({gb / 1E9:.1f}GB)'
+            pbar.close()
+
+    def cache_labels(self, path=Path('./labels.cache'), prefix=''):
+        # Cache dataset labels, check images and read shapes
+        x = {}  # dict
+        nm, nf, ne, nc = 0, 0, 0, 0  # number missing, found, empty, duplicate
+        pbar = tqdm(zip(self.img_files, self.label_files), desc='Scanning images', total=len(self.img_files))
+        for i, (im_file, lb_file) in enumerate(pbar):
+            try:
+                # verify images
+                im = Image.open(im_file)
+                im.verify()  # PIL verify
+                shape = exif_size(im)  # image size
+                segments = []  # instance segments
+                assert (shape[0] > 9) & (shape[1] > 9), f'image size {shape} <10 pixels'
+                assert im.format.lower() in img_formats, f'invalid image format {im.format}'
+
+                # verify labels
+                if os.path.isfile(lb_file):
+                    nf += 1  # label found
+                    with open(lb_file, 'r') as f:
+                        l = [x.split() for x in f.read().strip().splitlines()]
+                        if any([len(x) > 8 for x in l]):  # is segment
+                            classes = np.array([x[0] for x in l], dtype=np.float32)
+                            segments = [np.array(x[1:], dtype=np.float32).reshape(-1, 2) for x in l]  # (cls, xy1...)
+                            l = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1)  # (cls, xywh)
+                        l = np.array(l, dtype=np.float32)
+                    if len(l):
+                        assert l.shape[1] == 5, 'labels require 5 columns each'
+                        assert (l >= 0).all(), 'negative labels'
+                        assert (l[:, 1:] <= 1).all(), 'non-normalized or out of bounds coordinate labels'
+                        assert np.unique(l, axis=0).shape[0] == l.shape[0], 'duplicate labels'
+                    else:
+                        ne += 1  # label empty
+                        l = np.zeros((0, 5), dtype=np.float32)
+                else:
+                    nm += 1  # label missing
+                    l = np.zeros((0, 5), dtype=np.float32)
+                x[im_file] = [l, shape, segments]
+            except Exception as e:
+                nc += 1
+                print(f'{prefix}WARNING: Ignoring corrupted image and/or label {im_file}: {e}')
+
+            pbar.desc = f"{prefix}Scanning '{path.parent / path.stem}' images and labels... " \
+                        f"{nf} found, {nm} missing, {ne} empty, {nc} corrupted"
+        pbar.close()
+
+        if nf == 0:
+            print(f'{prefix}WARNING: No labels found in {path}. See {help_url}')
+
+        x['hash'] = get_hash(self.label_files + self.img_files)
+        x['results'] = nf, nm, ne, nc, i + 1
+        x['version'] = 0.1  # cache version
+        torch.save(x, path)  # save for next time
+        logging.info(f'{prefix}New cache created: {path}')
+        return x
+
+    def __len__(self):
+        return len(self.img_files)
+
+    # def __iter__(self):
+    #     self.count = -1
+    #     print('ran dataset iter')
+    #     #self.shuffled_vector = np.random.permutation(self.nF) if self.augment else np.arange(self.nF)
+    #     return self
+
+    def __getitem__(self, index):
+        index = self.indices[index]  # linear, shuffled, or image_weights
+
+        hyp = self.hyp
+        mosaic = self.mosaic and random.random() < hyp['mosaic']
+        if mosaic:
+            # Load mosaic
+            if random.random() < 0.8:
+                img, labels = load_mosaic(self, index)
+            else:
+                img, labels = load_mosaic9(self, index)
+            shapes = None
+
+            # MixUp https://arxiv.org/pdf/1710.09412.pdf
+            if random.random() < hyp['mixup']:
+                if random.random() < 0.8:
+                    img2, labels2 = load_mosaic(self, random.randint(0, len(self.labels) - 1))
+                else:
+                    img2, labels2 = load_mosaic9(self, random.randint(0, len(self.labels) - 1))
+                r = np.random.beta(8.0, 8.0)  # mixup ratio, alpha=beta=8.0
+                img = (img * r + img2 * (1 - r)).astype(np.uint8)
+                labels = np.concatenate((labels, labels2), 0)
+
+        else:
+            # Load image
+            img, (h0, w0), (h, w) = load_image(self, index)
+
+            # Letterbox
+            shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size  # final letterboxed shape
+            img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment)
+            shapes = (h0, w0), ((h / h0, w / w0), pad)  # for COCO mAP rescaling
+
+            labels = self.labels[index].copy()
+            if labels.size:  # normalized xywh to pixel xyxy format
+                labels[:, 1:] = xywhn2xyxy(labels[:, 1:], ratio[0] * w, ratio[1] * h, padw=pad[0], padh=pad[1])
+
+        if self.augment:
+            # Augment imagespace
+            if not mosaic:
+                img, labels = random_perspective(img, labels,
+                                                 degrees=hyp['degrees'],
+                                                 translate=hyp['translate'],
+                                                 scale=hyp['scale'],
+                                                 shear=hyp['shear'],
+                                                 perspective=hyp['perspective'])
+            
+            
+            #img, labels = self.albumentations(img, labels)
+
+            # Augment colorspace
+            augment_hsv(img, hgain=hyp['hsv_h'], sgain=hyp['hsv_s'], vgain=hyp['hsv_v'])
+
+            # Apply cutouts
+            # if random.random() < 0.9:
+            #     labels = cutout(img, labels)
+            
+            if random.random() < hyp['paste_in']:
+                sample_labels, sample_images, sample_masks = [], [], [] 
+                while len(sample_labels) < 30:
+                    sample_labels_, sample_images_, sample_masks_ = load_samples(self, random.randint(0, len(self.labels) - 1))
+                    sample_labels += sample_labels_
+                    sample_images += sample_images_
+                    sample_masks += sample_masks_
+                    #print(len(sample_labels))
+                    if len(sample_labels) == 0:
+                        break
+                labels = pastein(img, labels, sample_labels, sample_images, sample_masks)
+
+        nL = len(labels)  # number of labels
+        if nL:
+            labels[:, 1:5] = xyxy2xywh(labels[:, 1:5])  # convert xyxy to xywh
+            labels[:, [2, 4]] /= img.shape[0]  # normalized height 0-1
+            labels[:, [1, 3]] /= img.shape[1]  # normalized width 0-1
+
+        if self.augment:
+            # flip up-down
+            if random.random() < hyp['flipud']:
+                img = np.flipud(img)
+                if nL:
+                    labels[:, 2] = 1 - labels[:, 2]
+
+            # flip left-right
+            if random.random() < hyp['fliplr']:
+                img = np.fliplr(img)
+                if nL:
+                    labels[:, 1] = 1 - labels[:, 1]
+
+        labels_out = torch.zeros((nL, 6))
+        if nL:
+            labels_out[:, 1:] = torch.from_numpy(labels)
+
+        # Convert
+        img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
+        img = np.ascontiguousarray(img)
+
+        return torch.from_numpy(img), labels_out, self.img_files[index], shapes
+
+    @staticmethod
+    def collate_fn(batch):
+        img, label, path, shapes = zip(*batch)  # transposed
+        for i, l in enumerate(label):
+            l[:, 0] = i  # add target image index for build_targets()
+        return torch.stack(img, 0), torch.cat(label, 0), path, shapes
+
+    @staticmethod
+    def collate_fn4(batch):
+        img, label, path, shapes = zip(*batch)  # transposed
+        n = len(shapes) // 4
+        img4, label4, path4, shapes4 = [], [], path[:n], shapes[:n]
+
+        ho = torch.tensor([[0., 0, 0, 1, 0, 0]])
+        wo = torch.tensor([[0., 0, 1, 0, 0, 0]])
+        s = torch.tensor([[1, 1, .5, .5, .5, .5]])  # scale
+        for i in range(n):  # zidane torch.zeros(16,3,720,1280)  # BCHW
+            i *= 4
+            if random.random() < 0.5:
+                im = F.interpolate(img[i].unsqueeze(0).float(), scale_factor=2., mode='bilinear', align_corners=False)[
+                    0].type(img[i].type())
+                l = label[i]
+            else:
+                im = torch.cat((torch.cat((img[i], img[i + 1]), 1), torch.cat((img[i + 2], img[i + 3]), 1)), 2)
+                l = torch.cat((label[i], label[i + 1] + ho, label[i + 2] + wo, label[i + 3] + ho + wo), 0) * s
+            img4.append(im)
+            label4.append(l)
+
+        for i, l in enumerate(label4):
+            l[:, 0] = i  # add target image index for build_targets()
+
+        return torch.stack(img4, 0), torch.cat(label4, 0), path4, shapes4
+
+
+# Ancillary functions --------------------------------------------------------------------------------------------------
+def load_image(self, index):
+    # loads 1 image from dataset, returns img, original hw, resized hw
+    img = self.imgs[index]
+    if img is None:  # not cached
+        path = self.img_files[index]
+        img = cv2.imread(path)  # BGR
+        assert img is not None, 'Image Not Found ' + path
+        h0, w0 = img.shape[:2]  # orig hw
+        r = self.img_size / max(h0, w0)  # resize image to img_size
+        if r != 1:  # always resize down, only resize up if training with augmentation
+            interp = cv2.INTER_AREA if r < 1 and not self.augment else cv2.INTER_LINEAR
+            img = cv2.resize(img, (int(w0 * r), int(h0 * r)), interpolation=interp)
+        return img, (h0, w0), img.shape[:2]  # img, hw_original, hw_resized
+    else:
+        return self.imgs[index], self.img_hw0[index], self.img_hw[index]  # img, hw_original, hw_resized
+
+
+def augment_hsv(img, hgain=0.5, sgain=0.5, vgain=0.5):
+    r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1  # random gains
+    hue, sat, val = cv2.split(cv2.cvtColor(img, cv2.COLOR_BGR2HSV))
+    dtype = img.dtype  # uint8
+
+    x = np.arange(0, 256, dtype=np.int16)
+    lut_hue = ((x * r[0]) % 180).astype(dtype)
+    lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
+    lut_val = np.clip(x * r[2], 0, 255).astype(dtype)
+
+    img_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val))).astype(dtype)
+    cv2.cvtColor(img_hsv, cv2.COLOR_HSV2BGR, dst=img)  # no return needed
+
+
+def hist_equalize(img, clahe=True, bgr=False):
+    # Equalize histogram on BGR image 'img' with img.shape(n,m,3) and range 0-255
+    yuv = cv2.cvtColor(img, cv2.COLOR_BGR2YUV if bgr else cv2.COLOR_RGB2YUV)
+    if clahe:
+        c = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
+        yuv[:, :, 0] = c.apply(yuv[:, :, 0])
+    else:
+        yuv[:, :, 0] = cv2.equalizeHist(yuv[:, :, 0])  # equalize Y channel histogram
+    return cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR if bgr else cv2.COLOR_YUV2RGB)  # convert YUV image to RGB
+
+
+def load_mosaic(self, index):
+    # loads images in a 4-mosaic
+
+    labels4, segments4 = [], []
+    s = self.img_size
+    yc, xc = [int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border]  # mosaic center x, y
+    indices = [index] + random.choices(self.indices, k=3)  # 3 additional image indices
+    for i, index in enumerate(indices):
+        # Load image
+        img, _, (h, w) = load_image(self, index)
+
+        # place img in img4
+        if i == 0:  # top left
+            img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
+            x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc  # xmin, ymin, xmax, ymax (large image)
+            x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h  # xmin, ymin, xmax, ymax (small image)
+        elif i == 1:  # top right
+            x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
+            x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
+        elif i == 2:  # bottom left
+            x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
+            x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
+        elif i == 3:  # bottom right
+            x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
+            x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)
+
+        img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b]  # img4[ymin:ymax, xmin:xmax]
+        padw = x1a - x1b
+        padh = y1a - y1b
+
+        # Labels
+        labels, segments = self.labels[index].copy(), self.segments[index].copy()
+        if labels.size:
+            labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh)  # normalized xywh to pixel xyxy format
+            segments = [xyn2xy(x, w, h, padw, padh) for x in segments]
+        labels4.append(labels)
+        segments4.extend(segments)
+
+    # Concat/clip labels
+    labels4 = np.concatenate(labels4, 0)
+    for x in (labels4[:, 1:], *segments4):
+        np.clip(x, 0, 2 * s, out=x)  # clip when using random_perspective()
+    # img4, labels4 = replicate(img4, labels4)  # replicate
+
+    # Augment
+    #img4, labels4, segments4 = remove_background(img4, labels4, segments4)
+    #sample_segments(img4, labels4, segments4, probability=self.hyp['copy_paste'])
+    img4, labels4, segments4 = copy_paste(img4, labels4, segments4, probability=self.hyp['copy_paste'])
+    img4, labels4 = random_perspective(img4, labels4, segments4,
+                                       degrees=self.hyp['degrees'],
+                                       translate=self.hyp['translate'],
+                                       scale=self.hyp['scale'],
+                                       shear=self.hyp['shear'],
+                                       perspective=self.hyp['perspective'],
+                                       border=self.mosaic_border)  # border to remove
+
+    return img4, labels4
+
+
+def load_mosaic9(self, index):
+    # loads images in a 9-mosaic
+
+    labels9, segments9 = [], []
+    s = self.img_size
+    indices = [index] + random.choices(self.indices, k=8)  # 8 additional image indices
+    for i, index in enumerate(indices):
+        # Load image
+        img, _, (h, w) = load_image(self, index)
+
+        # place img in img9
+        if i == 0:  # center
+            img9 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
+            h0, w0 = h, w
+            c = s, s, s + w, s + h  # xmin, ymin, xmax, ymax (base) coordinates
+        elif i == 1:  # top
+            c = s, s - h, s + w, s
+        elif i == 2:  # top right
+            c = s + wp, s - h, s + wp + w, s
+        elif i == 3:  # right
+            c = s + w0, s, s + w0 + w, s + h
+        elif i == 4:  # bottom right
+            c = s + w0, s + hp, s + w0 + w, s + hp + h
+        elif i == 5:  # bottom
+            c = s + w0 - w, s + h0, s + w0, s + h0 + h
+        elif i == 6:  # bottom left
+            c = s + w0 - wp - w, s + h0, s + w0 - wp, s + h0 + h
+        elif i == 7:  # left
+            c = s - w, s + h0 - h, s, s + h0
+        elif i == 8:  # top left
+            c = s - w, s + h0 - hp - h, s, s + h0 - hp
+
+        padx, pady = c[:2]
+        x1, y1, x2, y2 = [max(x, 0) for x in c]  # allocate coords
+
+        # Labels
+        labels, segments = self.labels[index].copy(), self.segments[index].copy()
+        if labels.size:
+            labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padx, pady)  # normalized xywh to pixel xyxy format
+            segments = [xyn2xy(x, w, h, padx, pady) for x in segments]
+        labels9.append(labels)
+        segments9.extend(segments)
+
+        # Image
+        img9[y1:y2, x1:x2] = img[y1 - pady:, x1 - padx:]  # img9[ymin:ymax, xmin:xmax]
+        hp, wp = h, w  # height, width previous
+
+    # Offset
+    yc, xc = [int(random.uniform(0, s)) for _ in self.mosaic_border]  # mosaic center x, y
+    img9 = img9[yc:yc + 2 * s, xc:xc + 2 * s]
+
+    # Concat/clip labels
+    labels9 = np.concatenate(labels9, 0)
+    labels9[:, [1, 3]] -= xc
+    labels9[:, [2, 4]] -= yc
+    c = np.array([xc, yc])  # centers
+    segments9 = [x - c for x in segments9]
+
+    for x in (labels9[:, 1:], *segments9):
+        np.clip(x, 0, 2 * s, out=x)  # clip when using random_perspective()
+    # img9, labels9 = replicate(img9, labels9)  # replicate
+
+    # Augment
+    #img9, labels9, segments9 = remove_background(img9, labels9, segments9)
+    img9, labels9, segments9 = copy_paste(img9, labels9, segments9, probability=self.hyp['copy_paste'])
+    img9, labels9 = random_perspective(img9, labels9, segments9,
+                                       degrees=self.hyp['degrees'],
+                                       translate=self.hyp['translate'],
+                                       scale=self.hyp['scale'],
+                                       shear=self.hyp['shear'],
+                                       perspective=self.hyp['perspective'],
+                                       border=self.mosaic_border)  # border to remove
+
+    return img9, labels9
+
+
+def load_samples(self, index):
+    # loads images in a 4-mosaic
+
+    labels4, segments4 = [], []
+    s = self.img_size
+    yc, xc = [int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border]  # mosaic center x, y
+    indices = [index] + random.choices(self.indices, k=3)  # 3 additional image indices
+    for i, index in enumerate(indices):
+        # Load image
+        img, _, (h, w) = load_image(self, index)
+
+        # place img in img4
+        if i == 0:  # top left
+            img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
+            x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc  # xmin, ymin, xmax, ymax (large image)
+            x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h  # xmin, ymin, xmax, ymax (small image)
+        elif i == 1:  # top right
+            x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
+            x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
+        elif i == 2:  # bottom left
+            x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
+            x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
+        elif i == 3:  # bottom right
+            x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
+            x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)
+
+        img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b]  # img4[ymin:ymax, xmin:xmax]
+        padw = x1a - x1b
+        padh = y1a - y1b
+
+        # Labels
+        labels, segments = self.labels[index].copy(), self.segments[index].copy()
+        if labels.size:
+            labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh)  # normalized xywh to pixel xyxy format
+            segments = [xyn2xy(x, w, h, padw, padh) for x in segments]
+        labels4.append(labels)
+        segments4.extend(segments)
+
+    # Concat/clip labels
+    labels4 = np.concatenate(labels4, 0)
+    for x in (labels4[:, 1:], *segments4):
+        np.clip(x, 0, 2 * s, out=x)  # clip when using random_perspective()
+    # img4, labels4 = replicate(img4, labels4)  # replicate
+
+    # Augment
+    #img4, labels4, segments4 = remove_background(img4, labels4, segments4)
+    sample_labels, sample_images, sample_masks = sample_segments(img4, labels4, segments4, probability=0.5)
+
+    return sample_labels, sample_images, sample_masks
+
+
+def copy_paste(img, labels, segments, probability=0.5):
+    # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)
+    n = len(segments)
+    if probability and n:
+        h, w, c = img.shape  # height, width, channels
+        im_new = np.zeros(img.shape, np.uint8)
+        for j in random.sample(range(n), k=round(probability * n)):
+            l, s = labels[j], segments[j]
+            box = w - l[3], l[2], w - l[1], l[4]
+            ioa = bbox_ioa(box, labels[:, 1:5])  # intersection over area
+            if (ioa < 0.30).all():  # allow 30% obscuration of existing labels
+                labels = np.concatenate((labels, [[l[0], *box]]), 0)
+                segments.append(np.concatenate((w - s[:, 0:1], s[:, 1:2]), 1))
+                cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (255, 255, 255), cv2.FILLED)
+
+        result = cv2.bitwise_and(src1=img, src2=im_new)
+        result = cv2.flip(result, 1)  # augment segments (flip left-right)
+        i = result > 0  # pixels to replace
+        # i[:, :] = result.max(2).reshape(h, w, 1)  # act over ch
+        img[i] = result[i]  # cv2.imwrite('debug.jpg', img)  # debug
+
+    return img, labels, segments
+
+
+def remove_background(img, labels, segments):
+    # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)
+    n = len(segments)
+    h, w, c = img.shape  # height, width, channels
+    im_new = np.zeros(img.shape, np.uint8)
+    img_new = np.ones(img.shape, np.uint8) * 114
+    for j in range(n):
+        cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (255, 255, 255), cv2.FILLED)
+
+        result = cv2.bitwise_and(src1=img, src2=im_new)
+        
+        i = result > 0  # pixels to replace
+        img_new[i] = result[i]  # cv2.imwrite('debug.jpg', img)  # debug
+
+    return img_new, labels, segments
+
+
+def sample_segments(img, labels, segments, probability=0.5):
+    # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)
+    n = len(segments)
+    sample_labels = []
+    sample_images = []
+    sample_masks = []
+    if probability and n:
+        h, w, c = img.shape  # height, width, channels
+        for j in random.sample(range(n), k=round(probability * n)):
+            l, s = labels[j], segments[j]
+            box = l[1].astype(int).clip(0,w-1), l[2].astype(int).clip(0,h-1), l[3].astype(int).clip(0,w-1), l[4].astype(int).clip(0,h-1) 
+            
+            #print(box)
+            if (box[2] <= box[0]) or (box[3] <= box[1]):
+                continue
+            
+            sample_labels.append(l[0])
+            
+            mask = np.zeros(img.shape, np.uint8)
+            
+            cv2.drawContours(mask, [segments[j].astype(np.int32)], -1, (255, 255, 255), cv2.FILLED)
+            sample_masks.append(mask[box[1]:box[3],box[0]:box[2],:])
+            
+            result = cv2.bitwise_and(src1=img, src2=mask)
+            i = result > 0  # pixels to replace
+            mask[i] = result[i]  # cv2.imwrite('debug.jpg', img)  # debug
+            #print(box)
+            sample_images.append(mask[box[1]:box[3],box[0]:box[2],:])
+
+    return sample_labels, sample_images, sample_masks
+
+
+def replicate(img, labels):
+    # Replicate labels
+    h, w = img.shape[:2]
+    boxes = labels[:, 1:].astype(int)
+    x1, y1, x2, y2 = boxes.T
+    s = ((x2 - x1) + (y2 - y1)) / 2  # side length (pixels)
+    for i in s.argsort()[:round(s.size * 0.5)]:  # smallest indices
+        x1b, y1b, x2b, y2b = boxes[i]
+        bh, bw = y2b - y1b, x2b - x1b
+        yc, xc = int(random.uniform(0, h - bh)), int(random.uniform(0, w - bw))  # offset x, y
+        x1a, y1a, x2a, y2a = [xc, yc, xc + bw, yc + bh]
+        img[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b]  # img4[ymin:ymax, xmin:xmax]
+        labels = np.append(labels, [[labels[i, 0], x1a, y1a, x2a, y2a]], axis=0)
+
+    return img, labels
+
+
+def letterbox(img, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):
+    # Resize and pad image while meeting stride-multiple constraints
+    shape = img.shape[:2]  # current shape [height, width]
+    if isinstance(new_shape, int):
+        new_shape = (new_shape, new_shape)
+
+    # Scale ratio (new / old)
+    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
+    if not scaleup:  # only scale down, do not scale up (for better test mAP)
+        r = min(r, 1.0)
+
+    # Compute padding
+    ratio = r, r  # width, height ratios
+    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
+    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
+    if auto:  # minimum rectangle
+        dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # wh padding
+    elif scaleFill:  # stretch
+        dw, dh = 0.0, 0.0
+        new_unpad = (new_shape[1], new_shape[0])
+        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios
+
+    dw /= 2  # divide padding into 2 sides
+    dh /= 2
+
+    if shape[::-1] != new_unpad:  # resize
+        img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
+    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
+    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
+    img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
+    return img, ratio, (dw, dh)
+
+
+def random_perspective(img, targets=(), segments=(), degrees=10, translate=.1, scale=.1, shear=10, perspective=0.0,
+                       border=(0, 0)):
+    # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))
+    # targets = [cls, xyxy]
+
+    height = img.shape[0] + border[0] * 2  # shape(h,w,c)
+    width = img.shape[1] + border[1] * 2
+
+    # Center
+    C = np.eye(3)
+    C[0, 2] = -img.shape[1] / 2  # x translation (pixels)
+    C[1, 2] = -img.shape[0] / 2  # y translation (pixels)
+
+    # Perspective
+    P = np.eye(3)
+    P[2, 0] = random.uniform(-perspective, perspective)  # x perspective (about y)
+    P[2, 1] = random.uniform(-perspective, perspective)  # y perspective (about x)
+
+    # Rotation and Scale
+    R = np.eye(3)
+    a = random.uniform(-degrees, degrees)
+    # a += random.choice([-180, -90, 0, 90])  # add 90deg rotations to small rotations
+    s = random.uniform(1 - scale, 1.1 + scale)
+    # s = 2 ** random.uniform(-scale, scale)
+    R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)
+
+    # Shear
+    S = np.eye(3)
+    S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # x shear (deg)
+    S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # y shear (deg)
+
+    # Translation
+    T = np.eye(3)
+    T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width  # x translation (pixels)
+    T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height  # y translation (pixels)
+
+    # Combined rotation matrix
+    M = T @ S @ R @ P @ C  # order of operations (right to left) is IMPORTANT
+    if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any():  # image changed
+        if perspective:
+            img = cv2.warpPerspective(img, M, dsize=(width, height), borderValue=(114, 114, 114))
+        else:  # affine
+            img = cv2.warpAffine(img, M[:2], dsize=(width, height), borderValue=(114, 114, 114))
+
+    # Visualize
+    # import matplotlib.pyplot as plt
+    # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()
+    # ax[0].imshow(img[:, :, ::-1])  # base
+    # ax[1].imshow(img2[:, :, ::-1])  # warped
+
+    # Transform label coordinates
+    n = len(targets)
+    if n:
+        use_segments = any(x.any() for x in segments)
+        new = np.zeros((n, 4))
+        if use_segments:  # warp segments
+            segments = resample_segments(segments)  # upsample
+            for i, segment in enumerate(segments):
+                xy = np.ones((len(segment), 3))
+                xy[:, :2] = segment
+                xy = xy @ M.T  # transform
+                xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]  # perspective rescale or affine
+
+                # clip
+                new[i] = segment2box(xy, width, height)
+
+        else:  # warp boxes
+            xy = np.ones((n * 4, 3))
+            xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2)  # x1y1, x2y2, x1y2, x2y1
+            xy = xy @ M.T  # transform
+            xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8)  # perspective rescale or affine
+
+            # create new boxes
+            x = xy[:, [0, 2, 4, 6]]
+            y = xy[:, [1, 3, 5, 7]]
+            new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T
+
+            # clip
+            new[:, [0, 2]] = new[:, [0, 2]].clip(0, width)
+            new[:, [1, 3]] = new[:, [1, 3]].clip(0, height)
+
+        # filter candidates
+        i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10)
+        targets = targets[i]
+        targets[:, 1:5] = new[i]
+
+    return img, targets
+
+
+def box_candidates(box1, box2, wh_thr=2, ar_thr=20, area_thr=0.1, eps=1e-16):  # box1(4,n), box2(4,n)
+    # Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio
+    w1, h1 = box1[2] - box1[0], box1[3] - box1[1]
+    w2, h2 = box2[2] - box2[0], box2[3] - box2[1]
+    ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps))  # aspect ratio
+    return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr)  # candidates
+
+
+def bbox_ioa(box1, box2):
+    # Returns the intersection over box2 area given box1, box2. box1 is 4, box2 is nx4. boxes are x1y1x2y2
+    box2 = box2.transpose()
+
+    # Get the coordinates of bounding boxes
+    b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]
+    b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]
+
+    # Intersection area
+    inter_area = (np.minimum(b1_x2, b2_x2) - np.maximum(b1_x1, b2_x1)).clip(0) * \
+                 (np.minimum(b1_y2, b2_y2) - np.maximum(b1_y1, b2_y1)).clip(0)
+
+    # box2 area
+    box2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1) + 1e-16
+
+    # Intersection over box2 area
+    return inter_area / box2_area
+    
+
+def cutout(image, labels):
+    # Applies image cutout augmentation https://arxiv.org/abs/1708.04552
+    h, w = image.shape[:2]
+
+    # create random masks
+    scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16  # image size fraction
+    for s in scales:
+        mask_h = random.randint(1, int(h * s))
+        mask_w = random.randint(1, int(w * s))
+
+        # box
+        xmin = max(0, random.randint(0, w) - mask_w // 2)
+        ymin = max(0, random.randint(0, h) - mask_h // 2)
+        xmax = min(w, xmin + mask_w)
+        ymax = min(h, ymin + mask_h)
+
+        # apply random color mask
+        image[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)]
+
+        # return unobscured labels
+        if len(labels) and s > 0.03:
+            box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32)
+            ioa = bbox_ioa(box, labels[:, 1:5])  # intersection over area
+            labels = labels[ioa < 0.60]  # remove >60% obscured labels
+
+    return labels
+    
+
+def pastein(image, labels, sample_labels, sample_images, sample_masks):
+    # Applies image cutout augmentation https://arxiv.org/abs/1708.04552
+    h, w = image.shape[:2]
+
+    # create random masks
+    scales = [0.75] * 2 + [0.5] * 4 + [0.25] * 4 + [0.125] * 4 + [0.0625] * 6  # image size fraction
+    for s in scales:
+        if random.random() < 0.2:
+            continue
+        mask_h = random.randint(1, int(h * s))
+        mask_w = random.randint(1, int(w * s))
+
+        # box
+        xmin = max(0, random.randint(0, w) - mask_w // 2)
+        ymin = max(0, random.randint(0, h) - mask_h // 2)
+        xmax = min(w, xmin + mask_w)
+        ymax = min(h, ymin + mask_h)   
+        
+        box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32)
+        if len(labels):
+            ioa = bbox_ioa(box, labels[:, 1:5])  # intersection over area     
+        else:
+            ioa = np.zeros(1)
+        
+        if (ioa < 0.30).all() and len(sample_labels) and (xmax > xmin+20) and (ymax > ymin+20):  # allow 30% obscuration of existing labels
+            sel_ind = random.randint(0, len(sample_labels)-1)
+            #print(len(sample_labels))
+            #print(sel_ind)
+            #print((xmax-xmin, ymax-ymin))
+            #print(image[ymin:ymax, xmin:xmax].shape)
+            #print([[sample_labels[sel_ind], *box]])
+            #print(labels.shape)
+            hs, ws, cs = sample_images[sel_ind].shape
+            r_scale = min((ymax-ymin)/hs, (xmax-xmin)/ws)
+            r_w = int(ws*r_scale)
+            r_h = int(hs*r_scale)
+            
+            if (r_w > 10) and (r_h > 10):
+                r_mask = cv2.resize(sample_masks[sel_ind], (r_w, r_h))
+                r_image = cv2.resize(sample_images[sel_ind], (r_w, r_h))
+                temp_crop = image[ymin:ymin+r_h, xmin:xmin+r_w]
+                m_ind = r_mask > 0
+                if m_ind.astype(np.int32).sum() > 60:
+                    temp_crop[m_ind] = r_image[m_ind]
+                    #print(sample_labels[sel_ind])
+                    #print(sample_images[sel_ind].shape)
+                    #print(temp_crop.shape)
+                    box = np.array([xmin, ymin, xmin+r_w, ymin+r_h], dtype=np.float32)
+                    if len(labels):
+                        labels = np.concatenate((labels, [[sample_labels[sel_ind], *box]]), 0)
+                    else:
+                        labels = np.array([[sample_labels[sel_ind], *box]])
+                              
+                    image[ymin:ymin+r_h, xmin:xmin+r_w] = temp_crop
+
+    return labels
+
+class Albumentations:
+    # YOLOv5 Albumentations class (optional, only used if package is installed)
+    def __init__(self):
+        self.transform = None
+        import albumentations as A
+
+        self.transform = A.Compose([
+            A.CLAHE(p=0.01),
+            A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.01),
+            A.RandomGamma(gamma_limit=[80, 120], p=0.01),
+            A.Blur(p=0.01),
+            A.MedianBlur(p=0.01),
+            A.ToGray(p=0.01),
+            A.ImageCompression(quality_lower=75, p=0.01),],
+            bbox_params=A.BboxParams(format='pascal_voc', label_fields=['class_labels']))
+
+            #logging.info(colorstr('albumentations: ') + ', '.join(f'{x}' for x in self.transform.transforms if x.p))
+
+    def __call__(self, im, labels, p=1.0):
+        if self.transform and random.random() < p:
+            new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0])  # transformed
+            im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])])
+        return im, labels
+
+
+def create_folder(path='./new'):
+    # Create folder
+    if os.path.exists(path):
+        shutil.rmtree(path)  # delete output folder
+    os.makedirs(path)  # make new output folder
+
+
+def flatten_recursive(path='../coco'):
+    # Flatten a recursive directory by bringing all files to top level
+    new_path = Path(path + '_flat')
+    create_folder(new_path)
+    for file in tqdm(glob.glob(str(Path(path)) + '/**/*.*', recursive=True)):
+        shutil.copyfile(file, new_path / Path(file).name)
+
+
+def extract_boxes(path='../coco/'):  # from utils.datasets import *; extract_boxes('../coco128')
+    # Convert detection dataset into classification dataset, with one directory per class
+
+    path = Path(path)  # images dir
+    shutil.rmtree(path / 'classifier') if (path / 'classifier').is_dir() else None  # remove existing
+    files = list(path.rglob('*.*'))
+    n = len(files)  # number of files
+    for im_file in tqdm(files, total=n):
+        if im_file.suffix[1:] in img_formats:
+            # image
+            im = cv2.imread(str(im_file))[..., ::-1]  # BGR to RGB
+            h, w = im.shape[:2]
+
+            # labels
+            lb_file = Path(img2label_paths([str(im_file)])[0])
+            if Path(lb_file).exists():
+                with open(lb_file, 'r') as f:
+                    lb = np.array([x.split() for x in f.read().strip().splitlines()], dtype=np.float32)  # labels
+
+                for j, x in enumerate(lb):
+                    c = int(x[0])  # class
+                    f = (path / 'classifier') / f'{c}' / f'{path.stem}_{im_file.stem}_{j}.jpg'  # new filename
+                    if not f.parent.is_dir():
+                        f.parent.mkdir(parents=True)
+
+                    b = x[1:] * [w, h, w, h]  # box
+                    # b[2:] = b[2:].max()  # rectangle to square
+                    b[2:] = b[2:] * 1.2 + 3  # pad
+                    b = xywh2xyxy(b.reshape(-1, 4)).ravel().astype(np.int)
+
+                    b[[0, 2]] = np.clip(b[[0, 2]], 0, w)  # clip boxes outside of image
+                    b[[1, 3]] = np.clip(b[[1, 3]], 0, h)
+                    assert cv2.imwrite(str(f), im[b[1]:b[3], b[0]:b[2]]), f'box failure in {f}'
+
+
+def autosplit(path='../coco', weights=(0.9, 0.1, 0.0), annotated_only=False):
+    """ Autosplit a dataset into train/val/test splits and save path/autosplit_*.txt files
+    Usage: from utils.datasets import *; autosplit('../coco')
+    Arguments
+        path:           Path to images directory
+        weights:        Train, val, test weights (list)
+        annotated_only: Only use images with an annotated txt file
+    """
+    path = Path(path)  # images dir
+    files = sum([list(path.rglob(f"*.{img_ext}")) for img_ext in img_formats], [])  # image files only
+    n = len(files)  # number of files
+    indices = random.choices([0, 1, 2], weights=weights, k=n)  # assign each image to a split
+
+    txt = ['autosplit_train.txt', 'autosplit_val.txt', 'autosplit_test.txt']  # 3 txt files
+    [(path / x).unlink() for x in txt if (path / x).exists()]  # remove existing
+
+    print(f'Autosplitting images from {path}' + ', using *.txt labeled images only' * annotated_only)
+    for i, img in tqdm(zip(indices, files), total=n):
+        if not annotated_only or Path(img2label_paths([str(img)])[0]).exists():  # check label
+            with open(path / txt[i], 'a') as f:
+                f.write(str(img) + '\n')  # add image to txt file
+    
+    
+def load_segmentations(self, index):
+    key = '/work/handsomejw66/coco17/' + self.img_files[index]
+    #print(key)
+    # /work/handsomejw66/coco17/
+    return self.segs[key]

utils/general.py

@@ -0,0 +1,892 @@
+# YOLOR general utils
+
+import glob
+import logging
+import math
+import os
+import platform
+import random
+import re
+import subprocess
+import time
+from pathlib import Path
+
+import cv2
+import numpy as np
+import pandas as pd
+import torch
+import torchvision
+import yaml
+
+from utils.google_utils import gsutil_getsize
+from utils.metrics import fitness
+from utils.torch_utils import init_torch_seeds
+
+# Settings
+torch.set_printoptions(linewidth=320, precision=5, profile='long')
+np.set_printoptions(linewidth=320, formatter={'float_kind': '{:11.5g}'.format})  # format short g, %precision=5
+pd.options.display.max_columns = 10
+cv2.setNumThreads(0)  # prevent OpenCV from multithreading (incompatible with PyTorch DataLoader)
+os.environ['NUMEXPR_MAX_THREADS'] = str(min(os.cpu_count(), 8))  # NumExpr max threads
+
+
+def set_logging(rank=-1):
+    logging.basicConfig(
+        format="%(message)s",
+        level=logging.INFO if rank in [-1, 0] else logging.WARN)
+
+
+def init_seeds(seed=0):
+    # Initialize random number generator (RNG) seeds
+    random.seed(seed)
+    np.random.seed(seed)
+    init_torch_seeds(seed)
+
+
+def get_latest_run(search_dir='.'):
+    # Return path to most recent 'last.pt' in /runs (i.e. to --resume from)
+    last_list = glob.glob(f'{search_dir}/**/last*.pt', recursive=True)
+    return max(last_list, key=os.path.getctime) if last_list else ''
+
+
+def isdocker():
+    # Is environment a Docker container
+    return Path('/workspace').exists()  # or Path('/.dockerenv').exists()
+
+
+def emojis(str=''):
+    # Return platform-dependent emoji-safe version of string
+    return str.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else str
+
+
+def check_online():
+    # Check internet connectivity
+    import socket
+    try:
+        socket.create_connection(("1.1.1.1", 443), 5)  # check host accesability
+        return True
+    except OSError:
+        return False
+
+
+def check_git_status():
+    # Recommend 'git pull' if code is out of date
+    print(colorstr('github: '), end='')
+    try:
+        assert Path('.git').exists(), 'skipping check (not a git repository)'
+        assert not isdocker(), 'skipping check (Docker image)'
+        assert check_online(), 'skipping check (offline)'
+
+        cmd = 'git fetch && git config --get remote.origin.url'
+        url = subprocess.check_output(cmd, shell=True).decode().strip().rstrip('.git')  # github repo url
+        branch = subprocess.check_output('git rev-parse --abbrev-ref HEAD', shell=True).decode().strip()  # checked out
+        n = int(subprocess.check_output(f'git rev-list {branch}..origin/master --count', shell=True))  # commits behind
+        if n > 0:
+            s = f"⚠️ WARNING: code is out of date by {n} commit{'s' * (n > 1)}. " \
+                f"Use 'git pull' to update or 'git clone {url}' to download latest."
+        else:
+            s = f'up to date with {url} ✅'
+        print(emojis(s))  # emoji-safe
+    except Exception as e:
+        print(e)
+
+
+def check_requirements(requirements='requirements.txt', exclude=()):
+    # Check installed dependencies meet requirements (pass *.txt file or list of packages)
+    import pkg_resources as pkg
+    prefix = colorstr('red', 'bold', 'requirements:')
+    if isinstance(requirements, (str, Path)):  # requirements.txt file
+        file = Path(requirements)
+        if not file.exists():
+            print(f"{prefix} {file.resolve()} not found, check failed.")
+            return
+        requirements = [f'{x.name}{x.specifier}' for x in pkg.parse_requirements(file.open()) if x.name not in exclude]
+    else:  # list or tuple of packages
+        requirements = [x for x in requirements if x not in exclude]
+
+    n = 0  # number of packages updates
+    for r in requirements:
+        try:
+            pkg.require(r)
+        except Exception as e:  # DistributionNotFound or VersionConflict if requirements not met
+            n += 1
+            print(f"{prefix} {e.req} not found and is required by YOLOR, attempting auto-update...")
+            print(subprocess.check_output(f"pip install '{e.req}'", shell=True).decode())
+
+    if n:  # if packages updated
+        source = file.resolve() if 'file' in locals() else requirements
+        s = f"{prefix} {n} package{'s' * (n > 1)} updated per {source}\n" \
+            f"{prefix} ⚠️ {colorstr('bold', 'Restart runtime or rerun command for updates to take effect')}\n"
+        print(emojis(s))  # emoji-safe
+
+
+def check_img_size(img_size, s=32):
+    # Verify img_size is a multiple of stride s
+    new_size = make_divisible(img_size, int(s))  # ceil gs-multiple
+    if new_size != img_size:
+        print('WARNING: --img-size %g must be multiple of max stride %g, updating to %g' % (img_size, s, new_size))
+    return new_size
+
+
+def check_imshow():
+    # Check if environment supports image displays
+    try:
+        assert not isdocker(), 'cv2.imshow() is disabled in Docker environments'
+        cv2.imshow('test', np.zeros((1, 1, 3)))
+        cv2.waitKey(1)
+        cv2.destroyAllWindows()
+        cv2.waitKey(1)
+        return True
+    except Exception as e:
+        print(f'WARNING: Environment does not support cv2.imshow() or PIL Image.show() image displays\n{e}')
+        return False
+
+
+def check_file(file):
+    # Search for file if not found
+    if Path(file).is_file() or file == '':
+        return file
+    else:
+        files = glob.glob('./**/' + file, recursive=True)  # find file
+        assert len(files), f'File Not Found: {file}'  # assert file was found
+        assert len(files) == 1, f"Multiple files match '{file}', specify exact path: {files}"  # assert unique
+        return files[0]  # return file
+
+
+def check_dataset(dict):
+    # Download dataset if not found locally
+    val, s = dict.get('val'), dict.get('download')
+    if val and len(val):
+        val = [Path(x).resolve() for x in (val if isinstance(val, list) else [val])]  # val path
+        if not all(x.exists() for x in val):
+            print('\nWARNING: Dataset not found, nonexistent paths: %s' % [str(x) for x in val if not x.exists()])
+            if s and len(s):  # download script
+                print('Downloading %s ...' % s)
+                if s.startswith('http') and s.endswith('.zip'):  # URL
+                    f = Path(s).name  # filename
+                    torch.hub.download_url_to_file(s, f)
+                    r = os.system('unzip -q %s -d ../ && rm %s' % (f, f))  # unzip
+                else:  # bash script
+                    r = os.system(s)
+                print('Dataset autodownload %s\n' % ('success' if r == 0 else 'failure'))  # analyze return value
+            else:
+                raise Exception('Dataset not found.')
+
+
+def make_divisible(x, divisor):
+    # Returns x evenly divisible by divisor
+    return math.ceil(x / divisor) * divisor
+
+
+def clean_str(s):
+    # Cleans a string by replacing special characters with underscore _
+    return re.sub(pattern="[|@#!¡·$€%&()=?¿^*;:,¨´><+]", repl="_", string=s)
+
+
+def one_cycle(y1=0.0, y2=1.0, steps=100):
+    # lambda function for sinusoidal ramp from y1 to y2
+    return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1
+
+
+def colorstr(*input):
+    # Colors a string https://en.wikipedia.org/wiki/ANSI_escape_code, i.e.  colorstr('blue', 'hello world')
+    *args, string = input if len(input) > 1 else ('blue', 'bold', input[0])  # color arguments, string
+    colors = {'black': '\033[30m',  # basic colors
+              'red': '\033[31m',
+              'green': '\033[32m',
+              'yellow': '\033[33m',
+              'blue': '\033[34m',
+              'magenta': '\033[35m',
+              'cyan': '\033[36m',
+              'white': '\033[37m',
+              'bright_black': '\033[90m',  # bright colors
+              'bright_red': '\033[91m',
+              'bright_green': '\033[92m',
+              'bright_yellow': '\033[93m',
+              'bright_blue': '\033[94m',
+              'bright_magenta': '\033[95m',
+              'bright_cyan': '\033[96m',
+              'bright_white': '\033[97m',
+              'end': '\033[0m',  # misc
+              'bold': '\033[1m',
+              'underline': '\033[4m'}
+    return ''.join(colors[x] for x in args) + f'{string}' + colors['end']
+
+
+def labels_to_class_weights(labels, nc=80):
+    # Get class weights (inverse frequency) from training labels
+    if labels[0] is None:  # no labels loaded
+        return torch.Tensor()
+
+    labels = np.concatenate(labels, 0)  # labels.shape = (866643, 5) for COCO
+    classes = labels[:, 0].astype(np.int32)  # labels = [class xywh]
+    weights = np.bincount(classes, minlength=nc)  # occurrences per class
+
+    # Prepend gridpoint count (for uCE training)
+    # gpi = ((320 / 32 * np.array([1, 2, 4])) ** 2 * 3).sum()  # gridpoints per image
+    # weights = np.hstack([gpi * len(labels)  - weights.sum() * 9, weights * 9]) ** 0.5  # prepend gridpoints to start
+
+    weights[weights == 0] = 1  # replace empty bins with 1
+    weights = 1 / weights  # number of targets per class
+    weights /= weights.sum()  # normalize
+    return torch.from_numpy(weights)
+
+
+def labels_to_image_weights(labels, nc=80, class_weights=np.ones(80)):
+    # Produces image weights based on class_weights and image contents
+    class_counts = np.array([np.bincount(x[:, 0].astype(np.int32), minlength=nc) for x in labels])
+    image_weights = (class_weights.reshape(1, nc) * class_counts).sum(1)
+    # index = random.choices(range(n), weights=image_weights, k=1)  # weight image sample
+    return image_weights
+
+
+def coco80_to_coco91_class():  # converts 80-index (val2014) to 91-index (paper)
+    # https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/
+    # a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n')
+    # b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n')
+    # x1 = [list(a[i] == b).index(True) + 1 for i in range(80)]  # darknet to coco
+    # x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)]  # coco to darknet
+    x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34,
+         35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
+         64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90]
+    return x
+
+
+def xyxy2xywh(x):
+    # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = (x[:, 0] + x[:, 2]) / 2  # x center
+    y[:, 1] = (x[:, 1] + x[:, 3]) / 2  # y center
+    y[:, 2] = x[:, 2] - x[:, 0]  # width
+    y[:, 3] = x[:, 3] - x[:, 1]  # height
+    return y
+
+
+def xywh2xyxy(x):
+    # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = x[:, 0] - x[:, 2] / 2  # top left x
+    y[:, 1] = x[:, 1] - x[:, 3] / 2  # top left y
+    y[:, 2] = x[:, 0] + x[:, 2] / 2  # bottom right x
+    y[:, 3] = x[:, 1] + x[:, 3] / 2  # bottom right y
+    return y
+
+
+def xywhn2xyxy(x, w=640, h=640, padw=0, padh=0):
+    # Convert nx4 boxes from [x, y, w, h] normalized to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = w * (x[:, 0] - x[:, 2] / 2) + padw  # top left x
+    y[:, 1] = h * (x[:, 1] - x[:, 3] / 2) + padh  # top left y
+    y[:, 2] = w * (x[:, 0] + x[:, 2] / 2) + padw  # bottom right x
+    y[:, 3] = h * (x[:, 1] + x[:, 3] / 2) + padh  # bottom right y
+    return y
+
+
+def xyn2xy(x, w=640, h=640, padw=0, padh=0):
+    # Convert normalized segments into pixel segments, shape (n,2)
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = w * x[:, 0] + padw  # top left x
+    y[:, 1] = h * x[:, 1] + padh  # top left y
+    return y
+
+
+def segment2box(segment, width=640, height=640):
+    # Convert 1 segment label to 1 box label, applying inside-image constraint, i.e. (xy1, xy2, ...) to (xyxy)
+    x, y = segment.T  # segment xy
+    inside = (x >= 0) & (y >= 0) & (x <= width) & (y <= height)
+    x, y, = x[inside], y[inside]
+    return np.array([x.min(), y.min(), x.max(), y.max()]) if any(x) else np.zeros((1, 4))  # xyxy
+
+
+def segments2boxes(segments):
+    # Convert segment labels to box labels, i.e. (cls, xy1, xy2, ...) to (cls, xywh)
+    boxes = []
+    for s in segments:
+        x, y = s.T  # segment xy
+        boxes.append([x.min(), y.min(), x.max(), y.max()])  # cls, xyxy
+    return xyxy2xywh(np.array(boxes))  # cls, xywh
+
+
+def resample_segments(segments, n=1000):
+    # Up-sample an (n,2) segment
+    for i, s in enumerate(segments):
+        s = np.concatenate((s, s[0:1, :]), axis=0)
+        x = np.linspace(0, len(s) - 1, n)
+        xp = np.arange(len(s))
+        segments[i] = np.concatenate([np.interp(x, xp, s[:, i]) for i in range(2)]).reshape(2, -1).T  # segment xy
+    return segments
+
+
+def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None):
+    # Rescale coords (xyxy) from img1_shape to img0_shape
+    if ratio_pad is None:  # calculate from img0_shape
+        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
+        pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2  # wh padding
+    else:
+        gain = ratio_pad[0][0]
+        pad = ratio_pad[1]
+
+    coords[:, [0, 2]] -= pad[0]  # x padding
+    coords[:, [1, 3]] -= pad[1]  # y padding
+    coords[:, :4] /= gain
+    clip_coords(coords, img0_shape)
+    return coords
+
+
+def clip_coords(boxes, img_shape):
+    # Clip bounding xyxy bounding boxes to image shape (height, width)
+    boxes[:, 0].clamp_(0, img_shape[1])  # x1
+    boxes[:, 1].clamp_(0, img_shape[0])  # y1
+    boxes[:, 2].clamp_(0, img_shape[1])  # x2
+    boxes[:, 3].clamp_(0, img_shape[0])  # y2
+
+
+def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7):
+    # Returns the IoU of box1 to box2. box1 is 4, box2 is nx4
+    box2 = box2.T
+
+    # Get the coordinates of bounding boxes
+    if x1y1x2y2:  # x1, y1, x2, y2 = box1
+        b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]
+        b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]
+    else:  # transform from xywh to xyxy
+        b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2
+        b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2
+        b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2
+        b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2
+
+    # Intersection area
+    inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \
+            (torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)
+
+    # Union Area
+    w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
+    w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps
+    union = w1 * h1 + w2 * h2 - inter + eps
+
+    iou = inter / union
+
+    if GIoU or DIoU or CIoU:
+        cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1)  # convex (smallest enclosing box) width
+        ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1)  # convex height
+        if CIoU or DIoU:  # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
+            c2 = cw ** 2 + ch ** 2 + eps  # convex diagonal squared
+            rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 +
+                    (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4  # center distance squared
+            if DIoU:
+                return iou - rho2 / c2  # DIoU
+            elif CIoU:  # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
+                v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / (h2 + eps)) - torch.atan(w1 / (h1 + eps)), 2)
+                with torch.no_grad():
+                    alpha = v / (v - iou + (1 + eps))
+                return iou - (rho2 / c2 + v * alpha)  # CIoU
+        else:  # GIoU https://arxiv.org/pdf/1902.09630.pdf
+            c_area = cw * ch + eps  # convex area
+            return iou - (c_area - union) / c_area  # GIoU
+    else:
+        return iou  # IoU
+
+
+
+
+def bbox_alpha_iou(box1, box2, x1y1x2y2=False, GIoU=False, DIoU=False, CIoU=False, alpha=2, eps=1e-9):
+    # Returns tsqrt_he IoU of box1 to box2. box1 is 4, box2 is nx4
+    box2 = box2.T
+
+    # Get the coordinates of bounding boxes
+    if x1y1x2y2:  # x1, y1, x2, y2 = box1
+        b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]
+        b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]
+    else:  # transform from xywh to xyxy
+        b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2
+        b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2
+        b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2
+        b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2
+
+    # Intersection area
+    inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \
+            (torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)
+
+    # Union Area
+    w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
+    w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps
+    union = w1 * h1 + w2 * h2 - inter + eps
+
+    # change iou into pow(iou+eps)
+    # iou = inter / union
+    iou = torch.pow(inter/union + eps, alpha)
+    # beta = 2 * alpha
+    if GIoU or DIoU or CIoU:
+        cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1)  # convex (smallest enclosing box) width
+        ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1)  # convex height
+        if CIoU or DIoU:  # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
+            c2 = (cw ** 2 + ch ** 2) ** alpha + eps  # convex diagonal
+            rho_x = torch.abs(b2_x1 + b2_x2 - b1_x1 - b1_x2)
+            rho_y = torch.abs(b2_y1 + b2_y2 - b1_y1 - b1_y2)
+            rho2 = ((rho_x ** 2 + rho_y ** 2) / 4) ** alpha  # center distance
+            if DIoU:
+                return iou - rho2 / c2  # DIoU
+            elif CIoU:  # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
+                v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2)
+                with torch.no_grad():
+                    alpha_ciou = v / ((1 + eps) - inter / union + v)
+                # return iou - (rho2 / c2 + v * alpha_ciou)  # CIoU
+                return iou - (rho2 / c2 + torch.pow(v * alpha_ciou + eps, alpha))  # CIoU
+        else:  # GIoU https://arxiv.org/pdf/1902.09630.pdf
+            # c_area = cw * ch + eps  # convex area
+            # return iou - (c_area - union) / c_area  # GIoU
+            c_area = torch.max(cw * ch + eps, union) # convex area
+            return iou - torch.pow((c_area - union) / c_area + eps, alpha)  # GIoU
+    else:
+        return iou # torch.log(iou+eps) or iou
+
+
+def box_iou(box1, box2):
+    # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
+    """
+    Return intersection-over-union (Jaccard index) of boxes.
+    Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
+    Arguments:
+        box1 (Tensor[N, 4])
+        box2 (Tensor[M, 4])
+    Returns:
+        iou (Tensor[N, M]): the NxM matrix containing the pairwise
+            IoU values for every element in boxes1 and boxes2
+    """
+
+    def box_area(box):
+        # box = 4xn
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    area1 = box_area(box1.T)
+    area2 = box_area(box2.T)
+
+    # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
+    inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
+    return inter / (area1[:, None] + area2 - inter)  # iou = inter / (area1 + area2 - inter)
+
+
+def wh_iou(wh1, wh2):
+    # Returns the nxm IoU matrix. wh1 is nx2, wh2 is mx2
+    wh1 = wh1[:, None]  # [N,1,2]
+    wh2 = wh2[None]  # [1,M,2]
+    inter = torch.min(wh1, wh2).prod(2)  # [N,M]
+    return inter / (wh1.prod(2) + wh2.prod(2) - inter)  # iou = inter / (area1 + area2 - inter)
+
+
+def box_giou(box1, box2):
+    """
+    Return generalized intersection-over-union (Jaccard index) between two sets of boxes.
+    Both sets of boxes are expected to be in ``(x1, y1, x2, y2)`` format with
+    ``0 <= x1 < x2`` and ``0 <= y1 < y2``.
+    Args:
+        boxes1 (Tensor[N, 4]): first set of boxes
+        boxes2 (Tensor[M, 4]): second set of boxes
+    Returns:
+        Tensor[N, M]: the NxM matrix containing the pairwise generalized IoU values
+        for every element in boxes1 and boxes2
+    """
+
+    def box_area(box):
+        # box = 4xn
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    area1 = box_area(box1.T)
+    area2 = box_area(box2.T)
+    
+    inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
+    union = (area1[:, None] + area2 - inter)
+
+    iou = inter / union
+
+    lti = torch.min(box1[:, None, :2], box2[:, :2])
+    rbi = torch.max(box1[:, None, 2:], box2[:, 2:])
+
+    whi = (rbi - lti).clamp(min=0)  # [N,M,2]
+    areai = whi[:, :, 0] * whi[:, :, 1]
+
+    return iou - (areai - union) / areai
+
+
+def box_ciou(box1, box2, eps: float = 1e-7):
+    """
+    Return complete intersection-over-union (Jaccard index) between two sets of boxes.
+    Both sets of boxes are expected to be in ``(x1, y1, x2, y2)`` format with
+    ``0 <= x1 < x2`` and ``0 <= y1 < y2``.
+    Args:
+        boxes1 (Tensor[N, 4]): first set of boxes
+        boxes2 (Tensor[M, 4]): second set of boxes
+        eps (float, optional): small number to prevent division by zero. Default: 1e-7
+    Returns:
+        Tensor[N, M]: the NxM matrix containing the pairwise complete IoU values
+        for every element in boxes1 and boxes2
+    """
+
+    def box_area(box):
+        # box = 4xn
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    area1 = box_area(box1.T)
+    area2 = box_area(box2.T)
+    
+    inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
+    union = (area1[:, None] + area2 - inter)
+
+    iou = inter / union
+
+    lti = torch.min(box1[:, None, :2], box2[:, :2])
+    rbi = torch.max(box1[:, None, 2:], box2[:, 2:])
+
+    whi = (rbi - lti).clamp(min=0)  # [N,M,2]
+    diagonal_distance_squared = (whi[:, :, 0] ** 2) + (whi[:, :, 1] ** 2) + eps
+
+    # centers of boxes
+    x_p = (box1[:, None, 0] + box1[:, None, 2]) / 2
+    y_p = (box1[:, None, 1] + box1[:, None, 3]) / 2
+    x_g = (box2[:, 0] + box2[:, 2]) / 2
+    y_g = (box2[:, 1] + box2[:, 3]) / 2
+    # The distance between boxes' centers squared.
+    centers_distance_squared = (x_p - x_g) ** 2 + (y_p - y_g) ** 2
+
+    w_pred = box1[:, None, 2] - box1[:, None, 0]
+    h_pred = box1[:, None, 3] - box1[:, None, 1]
+
+    w_gt = box2[:, 2] - box2[:, 0]
+    h_gt = box2[:, 3] - box2[:, 1]
+
+    v = (4 / (torch.pi ** 2)) * torch.pow((torch.atan(w_gt / h_gt) - torch.atan(w_pred / h_pred)), 2)
+    with torch.no_grad():
+        alpha = v / (1 - iou + v + eps)
+    return iou - (centers_distance_squared / diagonal_distance_squared) - alpha * v
+
+
+def box_diou(box1, box2, eps: float = 1e-7):
+    """
+    Return distance intersection-over-union (Jaccard index) between two sets of boxes.
+    Both sets of boxes are expected to be in ``(x1, y1, x2, y2)`` format with
+    ``0 <= x1 < x2`` and ``0 <= y1 < y2``.
+    Args:
+        boxes1 (Tensor[N, 4]): first set of boxes
+        boxes2 (Tensor[M, 4]): second set of boxes
+        eps (float, optional): small number to prevent division by zero. Default: 1e-7
+    Returns:
+        Tensor[N, M]: the NxM matrix containing the pairwise distance IoU values
+        for every element in boxes1 and boxes2
+    """
+
+    def box_area(box):
+        # box = 4xn
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    area1 = box_area(box1.T)
+    area2 = box_area(box2.T)
+    
+    inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
+    union = (area1[:, None] + area2 - inter)
+
+    iou = inter / union
+
+    lti = torch.min(box1[:, None, :2], box2[:, :2])
+    rbi = torch.max(box1[:, None, 2:], box2[:, 2:])
+
+    whi = (rbi - lti).clamp(min=0)  # [N,M,2]
+    diagonal_distance_squared = (whi[:, :, 0] ** 2) + (whi[:, :, 1] ** 2) + eps
+
+    # centers of boxes
+    x_p = (box1[:, None, 0] + box1[:, None, 2]) / 2
+    y_p = (box1[:, None, 1] + box1[:, None, 3]) / 2
+    x_g = (box2[:, 0] + box2[:, 2]) / 2
+    y_g = (box2[:, 1] + box2[:, 3]) / 2
+    # The distance between boxes' centers squared.
+    centers_distance_squared = (x_p - x_g) ** 2 + (y_p - y_g) ** 2
+
+    # The distance IoU is the IoU penalized by a normalized
+    # distance between boxes' centers squared.
+    return iou - (centers_distance_squared / diagonal_distance_squared)
+
+
+def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False,
+                        labels=()):
+    """Runs Non-Maximum Suppression (NMS) on inference results
+
+    Returns:
+         list of detections, on (n,6) tensor per image [xyxy, conf, cls]
+    """
+
+    nc = prediction.shape[2] - 5  # number of classes
+    xc = prediction[..., 4] > conf_thres  # candidates
+
+    # Settings
+    min_wh, max_wh = 2, 4096  # (pixels) minimum and maximum box width and height
+    max_det = 300  # maximum number of detections per image
+    max_nms = 30000  # maximum number of boxes into torchvision.ops.nms()
+    time_limit = 10.0  # seconds to quit after
+    redundant = True  # require redundant detections
+    multi_label &= nc > 1  # multiple labels per box (adds 0.5ms/img)
+    merge = False  # use merge-NMS
+
+    t = time.time()
+    output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0]
+    for xi, x in enumerate(prediction):  # image index, image inference
+        # Apply constraints
+        # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0  # width-height
+        x = x[xc[xi]]  # confidence
+
+        # Cat apriori labels if autolabelling
+        if labels and len(labels[xi]):
+            l = labels[xi]
+            v = torch.zeros((len(l), nc + 5), device=x.device)
+            v[:, :4] = l[:, 1:5]  # box
+            v[:, 4] = 1.0  # conf
+            v[range(len(l)), l[:, 0].long() + 5] = 1.0  # cls
+            x = torch.cat((x, v), 0)
+
+        # If none remain process next image
+        if not x.shape[0]:
+            continue
+
+        # Compute conf
+        if nc == 1:
+            x[:, 5:] = x[:, 4:5] # for models with one class, cls_loss is 0 and cls_conf is always 0.5,
+                                 # so there is no need to multiplicate.
+        else:
+            x[:, 5:] *= x[:, 4:5]  # conf = obj_conf * cls_conf
+
+        # Box (center x, center y, width, height) to (x1, y1, x2, y2)
+        box = xywh2xyxy(x[:, :4])
+
+        # Detections matrix nx6 (xyxy, conf, cls)
+        if multi_label:
+            i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
+            x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
+        else:  # best class only
+            conf, j = x[:, 5:].max(1, keepdim=True)
+            x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
+
+        # Filter by class
+        if classes is not None:
+            x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
+
+        # Apply finite constraint
+        # if not torch.isfinite(x).all():
+        #     x = x[torch.isfinite(x).all(1)]
+
+        # Check shape
+        n = x.shape[0]  # number of boxes
+        if not n:  # no boxes
+            continue
+        elif n > max_nms:  # excess boxes
+            x = x[x[:, 4].argsort(descending=True)[:max_nms]]  # sort by confidence
+
+        # Batched NMS
+        c = x[:, 5:6] * (0 if agnostic else max_wh)  # classes
+        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
+        i = torchvision.ops.nms(boxes, scores, iou_thres)  # NMS
+        if i.shape[0] > max_det:  # limit detections
+            i = i[:max_det]
+        if merge and (1 < n < 3E3):  # Merge NMS (boxes merged using weighted mean)
+            # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
+            iou = box_iou(boxes[i], boxes) > iou_thres  # iou matrix
+            weights = iou * scores[None]  # box weights
+            x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True)  # merged boxes
+            if redundant:
+                i = i[iou.sum(1) > 1]  # require redundancy
+
+        output[xi] = x[i]
+        if (time.time() - t) > time_limit:
+            print(f'WARNING: NMS time limit {time_limit}s exceeded')
+            break  # time limit exceeded
+
+    return output
+
+
+def non_max_suppression_kpt(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False,
+                        labels=(), kpt_label=False, nc=None, nkpt=None):
+    """Runs Non-Maximum Suppression (NMS) on inference results
+
+    Returns:
+         list of detections, on (n,6) tensor per image [xyxy, conf, cls]
+    """
+    if nc is None:
+        nc = prediction.shape[2] - 5  if not kpt_label else prediction.shape[2] - 56 # number of classes
+    xc = prediction[..., 4] > conf_thres  # candidates
+
+    # Settings
+    min_wh, max_wh = 2, 4096  # (pixels) minimum and maximum box width and height
+    max_det = 300  # maximum number of detections per image
+    max_nms = 30000  # maximum number of boxes into torchvision.ops.nms()
+    time_limit = 10.0  # seconds to quit after
+    redundant = True  # require redundant detections
+    multi_label &= nc > 1  # multiple labels per box (adds 0.5ms/img)
+    merge = False  # use merge-NMS
+
+    t = time.time()
+    output = [torch.zeros((0,6), device=prediction.device)] * prediction.shape[0]
+    for xi, x in enumerate(prediction):  # image index, image inference
+        # Apply constraints
+        # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0  # width-height
+        x = x[xc[xi]]  # confidence
+
+        # Cat apriori labels if autolabelling
+        if labels and len(labels[xi]):
+            l = labels[xi]
+            v = torch.zeros((len(l), nc + 5), device=x.device)
+            v[:, :4] = l[:, 1:5]  # box
+            v[:, 4] = 1.0  # conf
+            v[range(len(l)), l[:, 0].long() + 5] = 1.0  # cls
+            x = torch.cat((x, v), 0)
+
+        # If none remain process next image
+        if not x.shape[0]:
+            continue
+
+        # Compute conf
+        x[:, 5:5+nc] *= x[:, 4:5]  # conf = obj_conf * cls_conf
+
+        # Box (center x, center y, width, height) to (x1, y1, x2, y2)
+        box = xywh2xyxy(x[:, :4])
+
+        # Detections matrix nx6 (xyxy, conf, cls)
+        if multi_label:
+            i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
+            x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
+        else:  # best class only
+            if not kpt_label:
+                conf, j = x[:, 5:].max(1, keepdim=True)
+                x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
+            else:
+                kpts = x[:, 6:]
+                conf, j = x[:, 5:6].max(1, keepdim=True)
+                x = torch.cat((box, conf, j.float(), kpts), 1)[conf.view(-1) > conf_thres]
+
+
+        # Filter by class
+        if classes is not None:
+            x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
+
+        # Apply finite constraint
+        # if not torch.isfinite(x).all():
+        #     x = x[torch.isfinite(x).all(1)]
+
+        # Check shape
+        n = x.shape[0]  # number of boxes
+        if not n:  # no boxes
+            continue
+        elif n > max_nms:  # excess boxes
+            x = x[x[:, 4].argsort(descending=True)[:max_nms]]  # sort by confidence
+
+        # Batched NMS
+        c = x[:, 5:6] * (0 if agnostic else max_wh)  # classes
+        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
+        i = torchvision.ops.nms(boxes, scores, iou_thres)  # NMS
+        if i.shape[0] > max_det:  # limit detections
+            i = i[:max_det]
+        if merge and (1 < n < 3E3):  # Merge NMS (boxes merged using weighted mean)
+            # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
+            iou = box_iou(boxes[i], boxes) > iou_thres  # iou matrix
+            weights = iou * scores[None]  # box weights
+            x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True)  # merged boxes
+            if redundant:
+                i = i[iou.sum(1) > 1]  # require redundancy
+
+        output[xi] = x[i]
+        if (time.time() - t) > time_limit:
+            print(f'WARNING: NMS time limit {time_limit}s exceeded')
+            break  # time limit exceeded
+
+    return output
+
+
+def strip_optimizer(f='best.pt', s=''):  # from utils.general import *; strip_optimizer()
+    # Strip optimizer from 'f' to finalize training, optionally save as 's'
+    x = torch.load(f, map_location=torch.device('cpu'))
+    if x.get('ema'):
+        x['model'] = x['ema']  # replace model with ema
+    for k in 'optimizer', 'training_results', 'wandb_id', 'ema', 'updates':  # keys
+        x[k] = None
+    x['epoch'] = -1
+    x['model'].half()  # to FP16
+    for p in x['model'].parameters():
+        p.requires_grad = False
+    torch.save(x, s or f)
+    mb = os.path.getsize(s or f) / 1E6  # filesize
+    print(f"Optimizer stripped from {f},{(' saved as %s,' % s) if s else ''} {mb:.1f}MB")
+
+
+def print_mutation(hyp, results, yaml_file='hyp_evolved.yaml', bucket=''):
+    # Print mutation results to evolve.txt (for use with train.py --evolve)
+    a = '%10s' * len(hyp) % tuple(hyp.keys())  # hyperparam keys
+    b = '%10.3g' * len(hyp) % tuple(hyp.values())  # hyperparam values
+    c = '%10.4g' * len(results) % results  # results (P, R, mAP@0.5, mAP@0.5:0.95, val_losses x 3)
+    print('\n%s\n%s\nEvolved fitness: %s\n' % (a, b, c))
+
+    if bucket:
+        url = 'gs://%s/evolve.txt' % bucket
+        if gsutil_getsize(url) > (os.path.getsize('evolve.txt') if os.path.exists('evolve.txt') else 0):
+            os.system('gsutil cp %s .' % url)  # download evolve.txt if larger than local
+
+    with open('evolve.txt', 'a') as f:  # append result
+        f.write(c + b + '\n')
+    x = np.unique(np.loadtxt('evolve.txt', ndmin=2), axis=0)  # load unique rows
+    x = x[np.argsort(-fitness(x))]  # sort
+    np.savetxt('evolve.txt', x, '%10.3g')  # save sort by fitness
+
+    # Save yaml
+    for i, k in enumerate(hyp.keys()):
+        hyp[k] = float(x[0, i + 7])
+    with open(yaml_file, 'w') as f:
+        results = tuple(x[0, :7])
+        c = '%10.4g' * len(results) % results  # results (P, R, mAP@0.5, mAP@0.5:0.95, val_losses x 3)
+        f.write('# Hyperparameter Evolution Results\n# Generations: %g\n# Metrics: ' % len(x) + c + '\n\n')
+        yaml.dump(hyp, f, sort_keys=False)
+
+    if bucket:
+        os.system('gsutil cp evolve.txt %s gs://%s' % (yaml_file, bucket))  # upload
+
+
+def apply_classifier(x, model, img, im0):
+    # applies a second stage classifier to yolo outputs
+    im0 = [im0] if isinstance(im0, np.ndarray) else im0
+    for i, d in enumerate(x):  # per image
+        if d is not None and len(d):
+            d = d.clone()
+
+            # Reshape and pad cutouts
+            b = xyxy2xywh(d[:, :4])  # boxes
+            b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1)  # rectangle to square
+            b[:, 2:] = b[:, 2:] * 1.3 + 30  # pad
+            d[:, :4] = xywh2xyxy(b).long()
+
+            # Rescale boxes from img_size to im0 size
+            scale_coords(img.shape[2:], d[:, :4], im0[i].shape)
+
+            # Classes
+            pred_cls1 = d[:, 5].long()
+            ims = []
+            for j, a in enumerate(d):  # per item
+                cutout = im0[i][int(a[1]):int(a[3]), int(a[0]):int(a[2])]
+                im = cv2.resize(cutout, (224, 224))  # BGR
+                # cv2.imwrite('test%i.jpg' % j, cutout)
+
+                im = im[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
+                im = np.ascontiguousarray(im, dtype=np.float32)  # uint8 to float32
+                im /= 255.0  # 0 - 255 to 0.0 - 1.0
+                ims.append(im)
+
+            pred_cls2 = model(torch.Tensor(ims).to(d.device)).argmax(1)  # classifier prediction
+            x[i] = x[i][pred_cls1 == pred_cls2]  # retain matching class detections
+
+    return x
+
+
+def increment_path(path, exist_ok=True, sep=''):
+    # Increment path, i.e. runs/exp --> runs/exp{sep}0, runs/exp{sep}1 etc.
+    path = Path(path)  # os-agnostic
+    if (path.exists() and exist_ok) or (not path.exists()):
+        return str(path)
+    else:
+        dirs = glob.glob(f"{path}{sep}*")  # similar paths
+        matches = [re.search(rf"%s{sep}(\d+)" % path.stem, d) for d in dirs]
+        i = [int(m.groups()[0]) for m in matches if m]  # indices
+        n = max(i) + 1 if i else 2  # increment number
+        return f"{path}{sep}{n}"  # update path

utils/google_app_engine/Dockerfile

@@ -0,0 +1,25 @@
+FROM gcr.io/google-appengine/python
+
+# Create a virtualenv for dependencies. This isolates these packages from
+# system-level packages.
+# Use -p python3 or -p python3.7 to select python version. Default is version 2.
+RUN virtualenv /env -p python3
+
+# Setting these environment variables are the same as running
+# source /env/bin/activate.
+ENV VIRTUAL_ENV /env
+ENV PATH /env/bin:$PATH
+
+RUN apt-get update && apt-get install -y python-opencv
+
+# Copy the application's requirements.txt and run pip to install all
+# dependencies into the virtualenv.
+ADD requirements.txt /app/requirements.txt
+RUN pip install -r /app/requirements.txt
+
+# Add the application source code.
+ADD . /app
+
+# Run a WSGI server to serve the application. gunicorn must be declared as
+# a dependency in requirements.txt.
+CMD gunicorn -b :$PORT main:app

utils/google_app_engine/additional_requirements.txt

@@ -0,0 +1,4 @@
+# add these requirements in your app on top of the existing ones
+pip==18.1
+Flask==1.0.2
+gunicorn==19.9.0

utils/google_app_engine/app.yaml

@@ -0,0 +1,14 @@
+runtime: custom
+env: flex
+
+service: yolorapp
+
+liveness_check:
+  initial_delay_sec: 600
+
+manual_scaling:
+  instances: 1
+resources:
+  cpu: 1
+  memory_gb: 4
+  disk_size_gb: 20

utils/google_utils.py

@@ -0,0 +1,123 @@
+# Google utils: https://cloud.google.com/storage/docs/reference/libraries
+
+import os
+import platform
+import subprocess
+import time
+from pathlib import Path
+
+import requests
+import torch
+
+
+def gsutil_getsize(url=''):
+    # gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du
+    s = subprocess.check_output(f'gsutil du {url}', shell=True).decode('utf-8')
+    return eval(s.split(' ')[0]) if len(s) else 0  # bytes
+
+
+def attempt_download(file, repo='WongKinYiu/yolov7'):
+    # Attempt file download if does not exist
+    file = Path(str(file).strip().replace("'", '').lower())
+
+    if not file.exists():
+        try:
+            response = requests.get(f'https://api.github.com/repos/{repo}/releases/latest').json()  # github api
+            assets = [x['name'] for x in response['assets']]  # release assets
+            tag = response['tag_name']  # i.e. 'v1.0'
+        except:  # fallback plan
+            assets = ['yolov7.pt', 'yolov7-tiny.pt', 'yolov7x.pt', 'yolov7-d6.pt', 'yolov7-e6.pt', 
+                      'yolov7-e6e.pt', 'yolov7-w6.pt']
+            tag = subprocess.check_output('git tag', shell=True).decode().split()[-1]
+
+        name = file.name
+        if name in assets:
+            msg = f'{file} missing, try downloading from https://github.com/{repo}/releases/'
+            redundant = False  # second download option
+            try:  # GitHub
+                url = f'https://github.com/{repo}/releases/download/{tag}/{name}'
+                print(f'Downloading {url} to {file}...')
+                torch.hub.download_url_to_file(url, file)
+                assert file.exists() and file.stat().st_size > 1E6  # check
+            except Exception as e:  # GCP
+                print(f'Download error: {e}')
+                assert redundant, 'No secondary mirror'
+                url = f'https://storage.googleapis.com/{repo}/ckpt/{name}'
+                print(f'Downloading {url} to {file}...')
+                os.system(f'curl -L {url} -o {file}')  # torch.hub.download_url_to_file(url, weights)
+            finally:
+                if not file.exists() or file.stat().st_size < 1E6:  # check
+                    file.unlink(missing_ok=True)  # remove partial downloads
+                    print(f'ERROR: Download failure: {msg}')
+                print('')
+                return
+
+
+def gdrive_download(id='', file='tmp.zip'):
+    # Downloads a file from Google Drive. from yolov7.utils.google_utils import *; gdrive_download()
+    t = time.time()
+    file = Path(file)
+    cookie = Path('cookie')  # gdrive cookie
+    print(f'Downloading https://drive.google.com/uc?export=download&id={id} as {file}... ', end='')
+    file.unlink(missing_ok=True)  # remove existing file
+    cookie.unlink(missing_ok=True)  # remove existing cookie
+
+    # Attempt file download
+    out = "NUL" if platform.system() == "Windows" else "/dev/null"
+    os.system(f'curl -c ./cookie -s -L "drive.google.com/uc?export=download&id={id}" > {out}')
+    if os.path.exists('cookie'):  # large file
+        s = f'curl -Lb ./cookie "drive.google.com/uc?export=download&confirm={get_token()}&id={id}" -o {file}'
+    else:  # small file
+        s = f'curl -s -L -o {file} "drive.google.com/uc?export=download&id={id}"'
+    r = os.system(s)  # execute, capture return
+    cookie.unlink(missing_ok=True)  # remove existing cookie
+
+    # Error check
+    if r != 0:
+        file.unlink(missing_ok=True)  # remove partial
+        print('Download error ')  # raise Exception('Download error')
+        return r
+
+    # Unzip if archive
+    if file.suffix == '.zip':
+        print('unzipping... ', end='')
+        os.system(f'unzip -q {file}')  # unzip
+        file.unlink()  # remove zip to free space
+
+    print(f'Done ({time.time() - t:.1f}s)')
+    return r
+
+
+def get_token(cookie="./cookie"):
+    with open(cookie) as f:
+        for line in f:
+            if "download" in line:
+                return line.split()[-1]
+    return ""
+
+# def upload_blob(bucket_name, source_file_name, destination_blob_name):
+#     # Uploads a file to a bucket
+#     # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python
+#
+#     storage_client = storage.Client()
+#     bucket = storage_client.get_bucket(bucket_name)
+#     blob = bucket.blob(destination_blob_name)
+#
+#     blob.upload_from_filename(source_file_name)
+#
+#     print('File {} uploaded to {}.'.format(
+#         source_file_name,
+#         destination_blob_name))
+#
+#
+# def download_blob(bucket_name, source_blob_name, destination_file_name):
+#     # Uploads a blob from a bucket
+#     storage_client = storage.Client()
+#     bucket = storage_client.get_bucket(bucket_name)
+#     blob = bucket.blob(source_blob_name)
+#
+#     blob.download_to_filename(destination_file_name)
+#
+#     print('Blob {} downloaded to {}.'.format(
+#         source_blob_name,
+#         destination_file_name))

utils/loss.py

@@ -0,0 +1,1697 @@
+# Loss functions
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from utils.general import bbox_iou, bbox_alpha_iou, box_iou, box_giou, box_diou, box_ciou, xywh2xyxy
+from utils.torch_utils import is_parallel
+
+
+def smooth_BCE(eps=0.1):  # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441
+    # return positive, negative label smoothing BCE targets
+    return 1.0 - 0.5 * eps, 0.5 * eps
+
+
+class BCEBlurWithLogitsLoss(nn.Module):
+    # BCEwithLogitLoss() with reduced missing label effects.
+    def __init__(self, alpha=0.05):
+        super(BCEBlurWithLogitsLoss, self).__init__()
+        self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none')  # must be nn.BCEWithLogitsLoss()
+        self.alpha = alpha
+
+    def forward(self, pred, true):
+        loss = self.loss_fcn(pred, true)
+        pred = torch.sigmoid(pred)  # prob from logits
+        dx = pred - true  # reduce only missing label effects
+        # dx = (pred - true).abs()  # reduce missing label and false label effects
+        alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4))
+        loss *= alpha_factor
+        return loss.mean()
+
+
+class SigmoidBin(nn.Module):
+    stride = None  # strides computed during build
+    export = False  # onnx export
+
+    def __init__(self, bin_count=10, min=0.0, max=1.0, reg_scale = 2.0, use_loss_regression=True, use_fw_regression=True, BCE_weight=1.0, smooth_eps=0.0):
+        super(SigmoidBin, self).__init__()
+        
+        self.bin_count = bin_count
+        self.length = bin_count + 1
+        self.min = min
+        self.max = max
+        self.scale = float(max - min)
+        self.shift = self.scale / 2.0
+
+        self.use_loss_regression = use_loss_regression
+        self.use_fw_regression = use_fw_regression
+        self.reg_scale = reg_scale
+        self.BCE_weight = BCE_weight
+
+        start = min + (self.scale/2.0) / self.bin_count
+        end = max - (self.scale/2.0) / self.bin_count
+        step = self.scale / self.bin_count
+        self.step = step
+        #print(f" start = {start}, end = {end}, step = {step} ")
+
+        bins = torch.range(start, end + 0.0001, step).float() 
+        self.register_buffer('bins', bins) 
+               
+
+        self.cp = 1.0 - 0.5 * smooth_eps
+        self.cn = 0.5 * smooth_eps
+
+        self.BCEbins = nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([BCE_weight]))
+        self.MSELoss = nn.MSELoss()
+
+    def get_length(self):
+        return self.length
+
+    def forward(self, pred):
+        assert pred.shape[-1] == self.length, 'pred.shape[-1]=%d is not equal to self.length=%d' % (pred.shape[-1], self.length)
+
+        pred_reg = (pred[..., 0] * self.reg_scale - self.reg_scale/2.0) * self.step
+        pred_bin = pred[..., 1:(1+self.bin_count)]
+
+        _, bin_idx = torch.max(pred_bin, dim=-1)
+        bin_bias = self.bins[bin_idx]
+
+        if self.use_fw_regression:
+            result = pred_reg + bin_bias
+        else:
+            result = bin_bias
+        result = result.clamp(min=self.min, max=self.max)
+
+        return result
+
+
+    def training_loss(self, pred, target):
+        assert pred.shape[-1] == self.length, 'pred.shape[-1]=%d is not equal to self.length=%d' % (pred.shape[-1], self.length)
+        assert pred.shape[0] == target.shape[0], 'pred.shape=%d is not equal to the target.shape=%d' % (pred.shape[0], target.shape[0])
+        device = pred.device
+
+        pred_reg = (pred[..., 0].sigmoid() * self.reg_scale - self.reg_scale/2.0) * self.step
+        pred_bin = pred[..., 1:(1+self.bin_count)]
+
+        diff_bin_target = torch.abs(target[..., None] - self.bins)
+        _, bin_idx = torch.min(diff_bin_target, dim=-1)
+    
+        bin_bias = self.bins[bin_idx]
+        bin_bias.requires_grad = False
+        result = pred_reg + bin_bias
+
+        target_bins = torch.full_like(pred_bin, self.cn, device=device)  # targets
+        n = pred.shape[0] 
+        target_bins[range(n), bin_idx] = self.cp
+
+        loss_bin = self.BCEbins(pred_bin, target_bins) # BCE
+
+        if self.use_loss_regression:
+            loss_regression = self.MSELoss(result, target)  # MSE        
+            loss = loss_bin + loss_regression
+        else:
+            loss = loss_bin
+
+        out_result = result.clamp(min=self.min, max=self.max)
+
+        return loss, out_result
+
+
+class FocalLoss(nn.Module):
+    # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
+    def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
+        super(FocalLoss, self).__init__()
+        self.loss_fcn = loss_fcn  # must be nn.BCEWithLogitsLoss()
+        self.gamma = gamma
+        self.alpha = alpha
+        self.reduction = loss_fcn.reduction
+        self.loss_fcn.reduction = 'none'  # required to apply FL to each element
+
+    def forward(self, pred, true):
+        loss = self.loss_fcn(pred, true)
+        # p_t = torch.exp(-loss)
+        # loss *= self.alpha * (1.000001 - p_t) ** self.gamma  # non-zero power for gradient stability
+
+        # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py
+        pred_prob = torch.sigmoid(pred)  # prob from logits
+        p_t = true * pred_prob + (1 - true) * (1 - pred_prob)
+        alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
+        modulating_factor = (1.0 - p_t) ** self.gamma
+        loss *= alpha_factor * modulating_factor
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        else:  # 'none'
+            return loss
+
+
+class QFocalLoss(nn.Module):
+    # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
+    def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
+        super(QFocalLoss, self).__init__()
+        self.loss_fcn = loss_fcn  # must be nn.BCEWithLogitsLoss()
+        self.gamma = gamma
+        self.alpha = alpha
+        self.reduction = loss_fcn.reduction
+        self.loss_fcn.reduction = 'none'  # required to apply FL to each element
+
+    def forward(self, pred, true):
+        loss = self.loss_fcn(pred, true)
+
+        pred_prob = torch.sigmoid(pred)  # prob from logits
+        alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
+        modulating_factor = torch.abs(true - pred_prob) ** self.gamma
+        loss *= alpha_factor * modulating_factor
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        else:  # 'none'
+            return loss
+
+class RankSort(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, logits, targets, delta_RS=0.50, eps=1e-10): 
+
+        classification_grads=torch.zeros(logits.shape).cuda()
+        
+        #Filter fg logits
+        fg_labels = (targets > 0.)
+        fg_logits = logits[fg_labels]
+        fg_targets = targets[fg_labels]
+        fg_num = len(fg_logits)
+
+        #Do not use bg with scores less than minimum fg logit
+        #since changing its score does not have an effect on precision
+        threshold_logit = torch.min(fg_logits)-delta_RS
+        relevant_bg_labels=((targets==0) & (logits>=threshold_logit))
+        
+        relevant_bg_logits = logits[relevant_bg_labels] 
+        relevant_bg_grad=torch.zeros(len(relevant_bg_logits)).cuda()
+        sorting_error=torch.zeros(fg_num).cuda()
+        ranking_error=torch.zeros(fg_num).cuda()
+        fg_grad=torch.zeros(fg_num).cuda()
+        
+        #sort the fg logits
+        order=torch.argsort(fg_logits)
+        #Loops over each positive following the order
+        for ii in order:
+            # Difference Transforms (x_ij)
+            fg_relations=fg_logits-fg_logits[ii] 
+            bg_relations=relevant_bg_logits-fg_logits[ii]
+
+            if delta_RS > 0:
+                fg_relations=torch.clamp(fg_relations/(2*delta_RS)+0.5,min=0,max=1)
+                bg_relations=torch.clamp(bg_relations/(2*delta_RS)+0.5,min=0,max=1)
+            else:
+                fg_relations = (fg_relations >= 0).float()
+                bg_relations = (bg_relations >= 0).float()
+
+            # Rank of ii among pos and false positive number (bg with larger scores)
+            rank_pos=torch.sum(fg_relations)
+            FP_num=torch.sum(bg_relations)
+
+            # Rank of ii among all examples
+            rank=rank_pos+FP_num
+                            
+            # Ranking error of example ii. target_ranking_error is always 0. (Eq. 7)
+            ranking_error[ii]=FP_num/rank      
+
+            # Current sorting error of example ii. (Eq. 7)
+            current_sorting_error = torch.sum(fg_relations*(1-fg_targets))/rank_pos
+
+            #Find examples in the target sorted order for example ii         
+            iou_relations = (fg_targets >= fg_targets[ii])
+            target_sorted_order = iou_relations * fg_relations
+
+            #The rank of ii among positives in sorted order
+            rank_pos_target = torch.sum(target_sorted_order)
+
+            #Compute target sorting error. (Eq. 8)
+            #Since target ranking error is 0, this is also total target error 
+            target_sorting_error= torch.sum(target_sorted_order*(1-fg_targets))/rank_pos_target
+
+            #Compute sorting error on example ii
+            sorting_error[ii] = current_sorting_error - target_sorting_error
+  
+            #Identity Update for Ranking Error 
+            if FP_num > eps:
+                #For ii the update is the ranking error
+                fg_grad[ii] -= ranking_error[ii]
+                #For negatives, distribute error via ranking pmf (i.e. bg_relations/FP_num)
+                relevant_bg_grad += (bg_relations*(ranking_error[ii]/FP_num))
+
+            #Find the positives that are misranked (the cause of the error)
+            #These are the ones with smaller IoU but larger logits
+            missorted_examples = (~ iou_relations) * fg_relations
+
+            #Denominotor of sorting pmf 
+            sorting_pmf_denom = torch.sum(missorted_examples)
+
+            #Identity Update for Sorting Error 
+            if sorting_pmf_denom > eps:
+                #For ii the update is the sorting error
+                fg_grad[ii] -= sorting_error[ii]
+                #For positives, distribute error via sorting pmf (i.e. missorted_examples/sorting_pmf_denom)
+                fg_grad += (missorted_examples*(sorting_error[ii]/sorting_pmf_denom))
+
+        #Normalize gradients by number of positives 
+        classification_grads[fg_labels]= (fg_grad/fg_num)
+        classification_grads[relevant_bg_labels]= (relevant_bg_grad/fg_num)
+
+        ctx.save_for_backward(classification_grads)
+
+        return ranking_error.mean(), sorting_error.mean()
+
+    @staticmethod
+    def backward(ctx, out_grad1, out_grad2):
+        g1, =ctx.saved_tensors
+        return g1*out_grad1, None, None, None
+
+class aLRPLoss(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, logits, targets, regression_losses, delta=1., eps=1e-5): 
+        classification_grads=torch.zeros(logits.shape).cuda()
+        
+        #Filter fg logits
+        fg_labels = (targets == 1)
+        fg_logits = logits[fg_labels]
+        fg_num = len(fg_logits)
+
+        #Do not use bg with scores less than minimum fg logit
+        #since changing its score does not have an effect on precision
+        threshold_logit = torch.min(fg_logits)-delta
+
+        #Get valid bg logits
+        relevant_bg_labels=((targets==0)&(logits>=threshold_logit))
+        relevant_bg_logits=logits[relevant_bg_labels] 
+        relevant_bg_grad=torch.zeros(len(relevant_bg_logits)).cuda()
+        rank=torch.zeros(fg_num).cuda()
+        prec=torch.zeros(fg_num).cuda()
+        fg_grad=torch.zeros(fg_num).cuda()
+        
+        max_prec=0                                           
+        #sort the fg logits
+        order=torch.argsort(fg_logits)
+        #Loops over each positive following the order
+        for ii in order:
+            #x_ij s as score differences with fgs
+            fg_relations=fg_logits-fg_logits[ii] 
+            #Apply piecewise linear function and determine relations with fgs
+            fg_relations=torch.clamp(fg_relations/(2*delta)+0.5,min=0,max=1)
+            #Discard i=j in the summation in rank_pos
+            fg_relations[ii]=0
+
+            #x_ij s as score differences with bgs
+            bg_relations=relevant_bg_logits-fg_logits[ii]
+            #Apply piecewise linear function and determine relations with bgs
+            bg_relations=torch.clamp(bg_relations/(2*delta)+0.5,min=0,max=1)
+
+            #Compute the rank of the example within fgs and number of bgs with larger scores
+            rank_pos=1+torch.sum(fg_relations)
+            FP_num=torch.sum(bg_relations)
+            #Store the total since it is normalizer also for aLRP Regression error
+            rank[ii]=rank_pos+FP_num
+                            
+            #Compute precision for this example to compute classification loss 
+            prec[ii]=rank_pos/rank[ii]                
+            #For stability, set eps to a infinitesmall value (e.g. 1e-6), then compute grads
+            if FP_num > eps:   
+                fg_grad[ii] = -(torch.sum(fg_relations*regression_losses)+FP_num)/rank[ii]
+                relevant_bg_grad += (bg_relations*(-fg_grad[ii]/FP_num))   
+                    
+        #aLRP with grad formulation fg gradient
+        classification_grads[fg_labels]= fg_grad
+        #aLRP with grad formulation bg gradient
+        classification_grads[relevant_bg_labels]= relevant_bg_grad 
+ 
+        classification_grads /= (fg_num)
+    
+        cls_loss=1-prec.mean()
+        ctx.save_for_backward(classification_grads)
+
+        return cls_loss, rank, order
+
+    @staticmethod
+    def backward(ctx, out_grad1, out_grad2, out_grad3):
+        g1, =ctx.saved_tensors
+        return g1*out_grad1, None, None, None, None
+    
+    
+class APLoss(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, logits, targets, delta=1.): 
+        classification_grads=torch.zeros(logits.shape).cuda()
+        
+        #Filter fg logits
+        fg_labels = (targets == 1)
+        fg_logits = logits[fg_labels]
+        fg_num = len(fg_logits)
+
+        #Do not use bg with scores less than minimum fg logit
+        #since changing its score does not have an effect on precision
+        threshold_logit = torch.min(fg_logits)-delta
+
+        #Get valid bg logits
+        relevant_bg_labels=((targets==0)&(logits>=threshold_logit))
+        relevant_bg_logits=logits[relevant_bg_labels] 
+        relevant_bg_grad=torch.zeros(len(relevant_bg_logits)).cuda()
+        rank=torch.zeros(fg_num).cuda()
+        prec=torch.zeros(fg_num).cuda()
+        fg_grad=torch.zeros(fg_num).cuda()
+        
+        max_prec=0                                           
+        #sort the fg logits
+        order=torch.argsort(fg_logits)
+        #Loops over each positive following the order
+        for ii in order:
+            #x_ij s as score differences with fgs
+            fg_relations=fg_logits-fg_logits[ii] 
+            #Apply piecewise linear function and determine relations with fgs
+            fg_relations=torch.clamp(fg_relations/(2*delta)+0.5,min=0,max=1)
+            #Discard i=j in the summation in rank_pos
+            fg_relations[ii]=0
+
+            #x_ij s as score differences with bgs
+            bg_relations=relevant_bg_logits-fg_logits[ii]
+            #Apply piecewise linear function and determine relations with bgs
+            bg_relations=torch.clamp(bg_relations/(2*delta)+0.5,min=0,max=1)
+
+            #Compute the rank of the example within fgs and number of bgs with larger scores
+            rank_pos=1+torch.sum(fg_relations)
+            FP_num=torch.sum(bg_relations)
+            #Store the total since it is normalizer also for aLRP Regression error
+            rank[ii]=rank_pos+FP_num
+                            
+            #Compute precision for this example 
+            current_prec=rank_pos/rank[ii]
+            
+            #Compute interpolated AP and store gradients for relevant bg examples
+            if (max_prec<=current_prec):
+                max_prec=current_prec
+                relevant_bg_grad += (bg_relations/rank[ii])
+            else:
+                relevant_bg_grad += (bg_relations/rank[ii])*(((1-max_prec)/(1-current_prec)))
+            
+            #Store fg gradients
+            fg_grad[ii]=-(1-max_prec)
+            prec[ii]=max_prec 
+
+        #aLRP with grad formulation fg gradient
+        classification_grads[fg_labels]= fg_grad
+        #aLRP with grad formulation bg gradient
+        classification_grads[relevant_bg_labels]= relevant_bg_grad 
+ 
+        classification_grads /= fg_num
+    
+        cls_loss=1-prec.mean()
+        ctx.save_for_backward(classification_grads)
+
+        return cls_loss
+
+    @staticmethod
+    def backward(ctx, out_grad1):
+        g1, =ctx.saved_tensors
+        return g1*out_grad1, None, None
+
+
+class ComputeLoss:
+    # Compute losses
+    def __init__(self, model, autobalance=False):
+        super(ComputeLoss, self).__init__()
+        device = next(model.parameters()).device  # get model device
+        h = model.hyp  # hyperparameters
+
+        # Define criteria
+        BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
+        BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
+
+        # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
+        self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0))  # positive, negative BCE targets
+
+        # Focal loss
+        g = h['fl_gamma']  # focal loss gamma
+        if g > 0:
+            BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
+
+        det = model.module.model[-1] if is_parallel(model) else model.model[-1]  # Detect() module
+        self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02])  # P3-P7
+        #self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.1, .05])  # P3-P7
+        #self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.5, 0.4, .1])  # P3-P7
+        self.ssi = list(det.stride).index(16) if autobalance else 0  # stride 16 index
+        self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
+        for k in 'na', 'nc', 'nl', 'anchors':
+            setattr(self, k, getattr(det, k))
+
+    def __call__(self, p, targets):  # predictions, targets, model
+        device = targets.device
+        lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
+        tcls, tbox, indices, anchors = self.build_targets(p, targets)  # targets
+
+        # Losses
+        for i, pi in enumerate(p):  # layer index, layer predictions
+            b, a, gj, gi = indices[i]  # image, anchor, gridy, gridx
+            tobj = torch.zeros_like(pi[..., 0], device=device)  # target obj
+
+            n = b.shape[0]  # number of targets
+            if n:
+                ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets
+
+                # Regression
+                pxy = ps[:, :2].sigmoid() * 2. - 0.5
+                pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
+                pbox = torch.cat((pxy, pwh), 1)  # predicted box
+                iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True)  # iou(prediction, target)
+                lbox += (1.0 - iou).mean()  # iou loss
+
+                # Objectness
+                tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype)  # iou ratio
+
+                # Classification
+                if self.nc > 1:  # cls loss (only if multiple classes)
+                    t = torch.full_like(ps[:, 5:], self.cn, device=device)  # targets
+                    t[range(n), tcls[i]] = self.cp
+                    #t[t==self.cp] = iou.detach().clamp(0).type(t.dtype)
+                    lcls += self.BCEcls(ps[:, 5:], t)  # BCE
+
+                # Append targets to text file
+                # with open('targets.txt', 'a') as file:
+                #     [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
+
+            obji = self.BCEobj(pi[..., 4], tobj)
+            lobj += obji * self.balance[i]  # obj loss
+            if self.autobalance:
+                self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
+
+        if self.autobalance:
+            self.balance = [x / self.balance[self.ssi] for x in self.balance]
+        lbox *= self.hyp['box']
+        lobj *= self.hyp['obj']
+        lcls *= self.hyp['cls']
+        bs = tobj.shape[0]  # batch size
+
+        loss = lbox + lobj + lcls
+        return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
+
+    def build_targets(self, p, targets):
+        # Build targets for compute_loss(), input targets(image,class,x,y,w,h)
+        na, nt = self.na, targets.shape[0]  # number of anchors, targets
+        tcls, tbox, indices, anch = [], [], [], []
+        gain = torch.ones(7, device=targets.device).long()  # normalized to gridspace gain
+        ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt)  # same as .repeat_interleave(nt)
+        targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2)  # append anchor indices
+
+        g = 0.5  # bias
+        off = torch.tensor([[0, 0],
+                            [1, 0], [0, 1], [-1, 0], [0, -1],  # j,k,l,m
+                            # [1, 1], [1, -1], [-1, 1], [-1, -1],  # jk,jm,lk,lm
+                            ], device=targets.device).float() * g  # offsets
+
+        for i in range(self.nl):
+            anchors = self.anchors[i]
+            gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]]  # xyxy gain
+
+            # Match targets to anchors
+            t = targets * gain
+            if nt:
+                # Matches
+                r = t[:, :, 4:6] / anchors[:, None]  # wh ratio
+                j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t']  # compare
+                # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t']  # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
+                t = t[j]  # filter
+
+                # Offsets
+                gxy = t[:, 2:4]  # grid xy
+                gxi = gain[[2, 3]] - gxy  # inverse
+                j, k = ((gxy % 1. < g) & (gxy > 1.)).T
+                l, m = ((gxi % 1. < g) & (gxi > 1.)).T
+                j = torch.stack((torch.ones_like(j), j, k, l, m))
+                t = t.repeat((5, 1, 1))[j]
+                offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
+            else:
+                t = targets[0]
+                offsets = 0
+
+            # Define
+            b, c = t[:, :2].long().T  # image, class
+            gxy = t[:, 2:4]  # grid xy
+            gwh = t[:, 4:6]  # grid wh
+            gij = (gxy - offsets).long()
+            gi, gj = gij.T  # grid xy indices
+
+            # Append
+            a = t[:, 6].long()  # anchor indices
+            indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1)))  # image, anchor, grid indices
+            tbox.append(torch.cat((gxy - gij, gwh), 1))  # box
+            anch.append(anchors[a])  # anchors
+            tcls.append(c)  # class
+
+        return tcls, tbox, indices, anch
+
+
+class ComputeLossOTA:
+    # Compute losses
+    def __init__(self, model, autobalance=False):
+        super(ComputeLossOTA, self).__init__()
+        device = next(model.parameters()).device  # get model device
+        h = model.hyp  # hyperparameters
+
+        # Define criteria
+        BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
+        BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
+
+        # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
+        self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0))  # positive, negative BCE targets
+
+        # Focal loss
+        g = h['fl_gamma']  # focal loss gamma
+        if g > 0:
+            BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
+
+        det = model.module.model[-1] if is_parallel(model) else model.model[-1]  # Detect() module
+        self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02])  # P3-P7
+        self.ssi = list(det.stride).index(16) if autobalance else 0  # stride 16 index
+        self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
+        for k in 'na', 'nc', 'nl', 'anchors', 'stride':
+            setattr(self, k, getattr(det, k))
+
+    def __call__(self, p, targets, imgs):  # predictions, targets, model   
+        device = targets.device
+        lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
+        bs, as_, gjs, gis, targets, anchors = self.build_targets(p, targets, imgs)
+        pre_gen_gains = [torch.tensor(pp.shape, device=device)[[3, 2, 3, 2]] for pp in p] 
+    
+
+        # Losses
+        for i, pi in enumerate(p):  # layer index, layer predictions
+            b, a, gj, gi = bs[i], as_[i], gjs[i], gis[i]  # image, anchor, gridy, gridx
+            tobj = torch.zeros_like(pi[..., 0], device=device)  # target obj
+
+            n = b.shape[0]  # number of targets
+            if n:
+                ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets
+
+                # Regression
+                grid = torch.stack([gi, gj], dim=1)
+                pxy = ps[:, :2].sigmoid() * 2. - 0.5
+                #pxy = ps[:, :2].sigmoid() * 3. - 1.
+                pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
+                pbox = torch.cat((pxy, pwh), 1)  # predicted box
+                selected_tbox = targets[i][:, 2:6] * pre_gen_gains[i]
+                selected_tbox[:, :2] -= grid
+                iou = bbox_iou(pbox.T, selected_tbox, x1y1x2y2=False, CIoU=True)  # iou(prediction, target)
+                lbox += (1.0 - iou).mean()  # iou loss
+
+                # Objectness
+                tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype)  # iou ratio
+
+                # Classification
+                selected_tcls = targets[i][:, 1].long()
+                if self.nc > 1:  # cls loss (only if multiple classes)
+                    t = torch.full_like(ps[:, 5:], self.cn, device=device)  # targets
+                    t[range(n), selected_tcls] = self.cp
+                    lcls += self.BCEcls(ps[:, 5:], t)  # BCE
+
+                # Append targets to text file
+                # with open('targets.txt', 'a') as file:
+                #     [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
+
+            obji = self.BCEobj(pi[..., 4], tobj)
+            lobj += obji * self.balance[i]  # obj loss
+            if self.autobalance:
+                self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
+
+        if self.autobalance:
+            self.balance = [x / self.balance[self.ssi] for x in self.balance]
+        lbox *= self.hyp['box']
+        lobj *= self.hyp['obj']
+        lcls *= self.hyp['cls']
+        bs = tobj.shape[0]  # batch size
+
+        loss = lbox + lobj + lcls
+        return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
+
+    def build_targets(self, p, targets, imgs):
+        
+        #indices, anch = self.find_positive(p, targets)
+        indices, anch = self.find_3_positive(p, targets)
+        #indices, anch = self.find_4_positive(p, targets)
+        #indices, anch = self.find_5_positive(p, targets)
+        #indices, anch = self.find_9_positive(p, targets)
+        device = torch.device(targets.device)
+        matching_bs = [[] for pp in p]
+        matching_as = [[] for pp in p]
+        matching_gjs = [[] for pp in p]
+        matching_gis = [[] for pp in p]
+        matching_targets = [[] for pp in p]
+        matching_anchs = [[] for pp in p]
+        
+        nl = len(p)    
+    
+        for batch_idx in range(p[0].shape[0]):
+        
+            b_idx = targets[:, 0]==batch_idx
+            this_target = targets[b_idx]
+            if this_target.shape[0] == 0:
+                continue
+                
+            txywh = this_target[:, 2:6] * imgs[batch_idx].shape[1]
+            txyxy = xywh2xyxy(txywh)
+
+            pxyxys = []
+            p_cls = []
+            p_obj = []
+            from_which_layer = []
+            all_b = []
+            all_a = []
+            all_gj = []
+            all_gi = []
+            all_anch = []
+            
+            for i, pi in enumerate(p):
+                
+                b, a, gj, gi = indices[i]
+                idx = (b == batch_idx)
+                b, a, gj, gi = b[idx], a[idx], gj[idx], gi[idx]                
+                all_b.append(b)
+                all_a.append(a)
+                all_gj.append(gj)
+                all_gi.append(gi)
+                all_anch.append(anch[i][idx])
+                from_which_layer.append((torch.ones(size=(len(b),)) * i).to(device))
+                
+                fg_pred = pi[b, a, gj, gi]                
+                p_obj.append(fg_pred[:, 4:5])
+                p_cls.append(fg_pred[:, 5:])
+                
+                grid = torch.stack([gi, gj], dim=1)
+                pxy = (fg_pred[:, :2].sigmoid() * 2. - 0.5 + grid) * self.stride[i] #/ 8.
+                #pxy = (fg_pred[:, :2].sigmoid() * 3. - 1. + grid) * self.stride[i]
+                pwh = (fg_pred[:, 2:4].sigmoid() * 2) ** 2 * anch[i][idx] * self.stride[i] #/ 8.
+                pxywh = torch.cat([pxy, pwh], dim=-1)
+                pxyxy = xywh2xyxy(pxywh)
+                pxyxys.append(pxyxy)
+            
+            pxyxys = torch.cat(pxyxys, dim=0)
+            if pxyxys.shape[0] == 0:
+                continue
+            p_obj = torch.cat(p_obj, dim=0)
+            p_cls = torch.cat(p_cls, dim=0)
+            from_which_layer = torch.cat(from_which_layer, dim=0)
+            all_b = torch.cat(all_b, dim=0)
+            all_a = torch.cat(all_a, dim=0)
+            all_gj = torch.cat(all_gj, dim=0)
+            all_gi = torch.cat(all_gi, dim=0)
+            all_anch = torch.cat(all_anch, dim=0)
+        
+            pair_wise_iou = box_iou(txyxy, pxyxys)
+
+            pair_wise_iou_loss = -torch.log(pair_wise_iou + 1e-8)
+
+            top_k, _ = torch.topk(pair_wise_iou, min(10, pair_wise_iou.shape[1]), dim=1)
+            dynamic_ks = torch.clamp(top_k.sum(1).int(), min=1)
+
+            gt_cls_per_image = (
+                F.one_hot(this_target[:, 1].to(torch.int64), self.nc)
+                .float()
+                .unsqueeze(1)
+                .repeat(1, pxyxys.shape[0], 1)
+            )
+
+            num_gt = this_target.shape[0]
+            cls_preds_ = (
+                p_cls.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
+                * p_obj.unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
+            )
+
+            y = cls_preds_.sqrt_()
+            pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
+               torch.log(y/(1-y)) , gt_cls_per_image, reduction="none"
+            ).sum(-1)
+            del cls_preds_
+        
+            cost = (
+                pair_wise_cls_loss
+                + 3.0 * pair_wise_iou_loss
+            )
+
+            matching_matrix = torch.zeros_like(cost, device=device)
+
+            for gt_idx in range(num_gt):
+                _, pos_idx = torch.topk(
+                    cost[gt_idx], k=dynamic_ks[gt_idx].item(), largest=False
+                )
+                matching_matrix[gt_idx][pos_idx] = 1.0
+
+            del top_k, dynamic_ks
+            anchor_matching_gt = matching_matrix.sum(0)
+            if (anchor_matching_gt > 1).sum() > 0:
+                _, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0)
+                matching_matrix[:, anchor_matching_gt > 1] *= 0.0
+                matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1.0
+            fg_mask_inboxes = (matching_matrix.sum(0) > 0.0).to(device)
+            matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)
+        
+            from_which_layer = from_which_layer[fg_mask_inboxes]
+            all_b = all_b[fg_mask_inboxes]
+            all_a = all_a[fg_mask_inboxes]
+            all_gj = all_gj[fg_mask_inboxes]
+            all_gi = all_gi[fg_mask_inboxes]
+            all_anch = all_anch[fg_mask_inboxes]
+        
+            this_target = this_target[matched_gt_inds]
+        
+            for i in range(nl):
+                layer_idx = from_which_layer == i
+                matching_bs[i].append(all_b[layer_idx])
+                matching_as[i].append(all_a[layer_idx])
+                matching_gjs[i].append(all_gj[layer_idx])
+                matching_gis[i].append(all_gi[layer_idx])
+                matching_targets[i].append(this_target[layer_idx])
+                matching_anchs[i].append(all_anch[layer_idx])
+
+        for i in range(nl):
+            if matching_targets[i] != []:
+                matching_bs[i] = torch.cat(matching_bs[i], dim=0)
+                matching_as[i] = torch.cat(matching_as[i], dim=0)
+                matching_gjs[i] = torch.cat(matching_gjs[i], dim=0)
+                matching_gis[i] = torch.cat(matching_gis[i], dim=0)
+                matching_targets[i] = torch.cat(matching_targets[i], dim=0)
+                matching_anchs[i] = torch.cat(matching_anchs[i], dim=0)
+            else:
+                matching_bs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_as[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_gjs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_gis[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_targets[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_anchs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+
+        return matching_bs, matching_as, matching_gjs, matching_gis, matching_targets, matching_anchs           
+
+    def find_3_positive(self, p, targets):
+        # Build targets for compute_loss(), input targets(image,class,x,y,w,h)
+        na, nt = self.na, targets.shape[0]  # number of anchors, targets
+        indices, anch = [], []
+        gain = torch.ones(7, device=targets.device).long()  # normalized to gridspace gain
+        ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt)  # same as .repeat_interleave(nt)
+        targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2)  # append anchor indices
+
+        g = 0.5  # bias
+        off = torch.tensor([[0, 0],
+                            [1, 0], [0, 1], [-1, 0], [0, -1],  # j,k,l,m
+                            # [1, 1], [1, -1], [-1, 1], [-1, -1],  # jk,jm,lk,lm
+                            ], device=targets.device).float() * g  # offsets
+
+        for i in range(self.nl):
+            anchors = self.anchors[i]
+            gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]]  # xyxy gain
+
+            # Match targets to anchors
+            t = targets * gain
+            if nt:
+                # Matches
+                r = t[:, :, 4:6] / anchors[:, None]  # wh ratio
+                j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t']  # compare
+                # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t']  # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
+                t = t[j]  # filter
+
+                # Offsets
+                gxy = t[:, 2:4]  # grid xy
+                gxi = gain[[2, 3]] - gxy  # inverse
+                j, k = ((gxy % 1. < g) & (gxy > 1.)).T
+                l, m = ((gxi % 1. < g) & (gxi > 1.)).T
+                j = torch.stack((torch.ones_like(j), j, k, l, m))
+                t = t.repeat((5, 1, 1))[j]
+                offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
+            else:
+                t = targets[0]
+                offsets = 0
+
+            # Define
+            b, c = t[:, :2].long().T  # image, class
+            gxy = t[:, 2:4]  # grid xy
+            gwh = t[:, 4:6]  # grid wh
+            gij = (gxy - offsets).long()
+            gi, gj = gij.T  # grid xy indices
+
+            # Append
+            a = t[:, 6].long()  # anchor indices
+            indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1)))  # image, anchor, grid indices
+            anch.append(anchors[a])  # anchors
+
+        return indices, anch
+    
+
+class ComputeLossBinOTA:
+    # Compute losses
+    def __init__(self, model, autobalance=False):
+        super(ComputeLossBinOTA, self).__init__()
+        device = next(model.parameters()).device  # get model device
+        h = model.hyp  # hyperparameters
+
+        # Define criteria
+        BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
+        BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
+        #MSEangle = nn.MSELoss().to(device)
+
+        # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
+        self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0))  # positive, negative BCE targets
+
+        # Focal loss
+        g = h['fl_gamma']  # focal loss gamma
+        if g > 0:
+            BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
+
+        det = model.module.model[-1] if is_parallel(model) else model.model[-1]  # Detect() module
+        self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02])  # P3-P7
+        self.ssi = list(det.stride).index(16) if autobalance else 0  # stride 16 index
+        self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
+        for k in 'na', 'nc', 'nl', 'anchors', 'stride', 'bin_count':
+            setattr(self, k, getattr(det, k))
+
+        #xy_bin_sigmoid = SigmoidBin(bin_count=11, min=-0.5, max=1.5, use_loss_regression=False).to(device)
+        wh_bin_sigmoid = SigmoidBin(bin_count=self.bin_count, min=0.0, max=4.0, use_loss_regression=False).to(device)
+        #angle_bin_sigmoid = SigmoidBin(bin_count=31, min=-1.1, max=1.1, use_loss_regression=False).to(device)
+        self.wh_bin_sigmoid = wh_bin_sigmoid
+
+    def __call__(self, p, targets, imgs):  # predictions, targets, model   
+        device = targets.device
+        lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
+        bs, as_, gjs, gis, targets, anchors = self.build_targets(p, targets, imgs)
+        pre_gen_gains = [torch.tensor(pp.shape, device=device)[[3, 2, 3, 2]] for pp in p] 
+    
+
+        # Losses
+        for i, pi in enumerate(p):  # layer index, layer predictions
+            b, a, gj, gi = bs[i], as_[i], gjs[i], gis[i]  # image, anchor, gridy, gridx
+            tobj = torch.zeros_like(pi[..., 0], device=device)  # target obj
+
+            obj_idx = self.wh_bin_sigmoid.get_length()*2 + 2     # x,y, w-bce, h-bce     # xy_bin_sigmoid.get_length()*2
+
+            n = b.shape[0]  # number of targets
+            if n:
+                ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets
+
+                # Regression
+                grid = torch.stack([gi, gj], dim=1)
+                selected_tbox = targets[i][:, 2:6] * pre_gen_gains[i]
+                selected_tbox[:, :2] -= grid
+                
+                #pxy = ps[:, :2].sigmoid() * 2. - 0.5
+                ##pxy = ps[:, :2].sigmoid() * 3. - 1.
+                #pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
+                #pbox = torch.cat((pxy, pwh), 1)  # predicted box
+
+                #x_loss, px = xy_bin_sigmoid.training_loss(ps[..., 0:12], tbox[i][..., 0])
+                #y_loss, py = xy_bin_sigmoid.training_loss(ps[..., 12:24], tbox[i][..., 1])
+                w_loss, pw = self.wh_bin_sigmoid.training_loss(ps[..., 2:(3+self.bin_count)], selected_tbox[..., 2] / anchors[i][..., 0])
+                h_loss, ph = self.wh_bin_sigmoid.training_loss(ps[..., (3+self.bin_count):obj_idx], selected_tbox[..., 3] / anchors[i][..., 1])
+
+                pw *= anchors[i][..., 0]
+                ph *= anchors[i][..., 1]
+
+                px = ps[:, 0].sigmoid() * 2. - 0.5
+                py = ps[:, 1].sigmoid() * 2. - 0.5
+
+                lbox += w_loss + h_loss # + x_loss + y_loss
+
+                #print(f"\n px = {px.shape}, py = {py.shape}, pw = {pw.shape}, ph = {ph.shape} \n")
+
+                pbox = torch.cat((px.unsqueeze(1), py.unsqueeze(1), pw.unsqueeze(1), ph.unsqueeze(1)), 1).to(device)  # predicted box
+
+                
+                
+                
+                iou = bbox_iou(pbox.T, selected_tbox, x1y1x2y2=False, CIoU=True)  # iou(prediction, target)
+                lbox += (1.0 - iou).mean()  # iou loss
+
+                # Objectness
+                tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype)  # iou ratio
+
+                # Classification
+                selected_tcls = targets[i][:, 1].long()
+                if self.nc > 1:  # cls loss (only if multiple classes)
+                    t = torch.full_like(ps[:, (1+obj_idx):], self.cn, device=device)  # targets
+                    t[range(n), selected_tcls] = self.cp
+                    lcls += self.BCEcls(ps[:, (1+obj_idx):], t)  # BCE
+
+                # Append targets to text file
+                # with open('targets.txt', 'a') as file:
+                #     [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
+
+            obji = self.BCEobj(pi[..., obj_idx], tobj)
+            lobj += obji * self.balance[i]  # obj loss
+            if self.autobalance:
+                self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
+
+        if self.autobalance:
+            self.balance = [x / self.balance[self.ssi] for x in self.balance]
+        lbox *= self.hyp['box']
+        lobj *= self.hyp['obj']
+        lcls *= self.hyp['cls']
+        bs = tobj.shape[0]  # batch size
+
+        loss = lbox + lobj + lcls
+        return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
+
+    def build_targets(self, p, targets, imgs):
+        
+        #indices, anch = self.find_positive(p, targets)
+        indices, anch = self.find_3_positive(p, targets)
+        #indices, anch = self.find_4_positive(p, targets)
+        #indices, anch = self.find_5_positive(p, targets)
+        #indices, anch = self.find_9_positive(p, targets)
+
+        matching_bs = [[] for pp in p]
+        matching_as = [[] for pp in p]
+        matching_gjs = [[] for pp in p]
+        matching_gis = [[] for pp in p]
+        matching_targets = [[] for pp in p]
+        matching_anchs = [[] for pp in p]
+        
+        nl = len(p)    
+    
+        for batch_idx in range(p[0].shape[0]):
+        
+            b_idx = targets[:, 0]==batch_idx
+            this_target = targets[b_idx]
+            if this_target.shape[0] == 0:
+                continue
+                
+            txywh = this_target[:, 2:6] * imgs[batch_idx].shape[1]
+            txyxy = xywh2xyxy(txywh)
+
+            pxyxys = []
+            p_cls = []
+            p_obj = []
+            from_which_layer = []
+            all_b = []
+            all_a = []
+            all_gj = []
+            all_gi = []
+            all_anch = []
+            
+            for i, pi in enumerate(p):
+                
+                obj_idx = self.wh_bin_sigmoid.get_length()*2 + 2
+                
+                b, a, gj, gi = indices[i]
+                idx = (b == batch_idx)
+                b, a, gj, gi = b[idx], a[idx], gj[idx], gi[idx]                
+                all_b.append(b)
+                all_a.append(a)
+                all_gj.append(gj)
+                all_gi.append(gi)
+                all_anch.append(anch[i][idx])
+                from_which_layer.append(torch.ones(size=(len(b),)) * i)
+                
+                fg_pred = pi[b, a, gj, gi]                
+                p_obj.append(fg_pred[:, obj_idx:(obj_idx+1)])
+                p_cls.append(fg_pred[:, (obj_idx+1):])
+                
+                grid = torch.stack([gi, gj], dim=1)
+                pxy = (fg_pred[:, :2].sigmoid() * 2. - 0.5 + grid) * self.stride[i] #/ 8.
+                #pwh = (fg_pred[:, 2:4].sigmoid() * 2) ** 2 * anch[i][idx] * self.stride[i] #/ 8.
+                pw = self.wh_bin_sigmoid.forward(fg_pred[..., 2:(3+self.bin_count)].sigmoid()) * anch[i][idx][:, 0] * self.stride[i]
+                ph = self.wh_bin_sigmoid.forward(fg_pred[..., (3+self.bin_count):obj_idx].sigmoid()) * anch[i][idx][:, 1] * self.stride[i]
+                
+                pxywh = torch.cat([pxy, pw.unsqueeze(1), ph.unsqueeze(1)], dim=-1)
+                pxyxy = xywh2xyxy(pxywh)
+                pxyxys.append(pxyxy)
+            
+            pxyxys = torch.cat(pxyxys, dim=0)
+            if pxyxys.shape[0] == 0:
+                continue
+            p_obj = torch.cat(p_obj, dim=0)
+            p_cls = torch.cat(p_cls, dim=0)
+            from_which_layer = torch.cat(from_which_layer, dim=0)
+            all_b = torch.cat(all_b, dim=0)
+            all_a = torch.cat(all_a, dim=0)
+            all_gj = torch.cat(all_gj, dim=0)
+            all_gi = torch.cat(all_gi, dim=0)
+            all_anch = torch.cat(all_anch, dim=0)
+        
+            pair_wise_iou = box_iou(txyxy, pxyxys)
+
+            pair_wise_iou_loss = -torch.log(pair_wise_iou + 1e-8)
+
+            top_k, _ = torch.topk(pair_wise_iou, min(10, pair_wise_iou.shape[1]), dim=1)
+            dynamic_ks = torch.clamp(top_k.sum(1).int(), min=1)
+
+            gt_cls_per_image = (
+                F.one_hot(this_target[:, 1].to(torch.int64), self.nc)
+                .float()
+                .unsqueeze(1)
+                .repeat(1, pxyxys.shape[0], 1)
+            )
+
+            num_gt = this_target.shape[0]            
+            cls_preds_ = (
+                p_cls.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
+                * p_obj.unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
+            )
+
+            y = cls_preds_.sqrt_()
+            pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
+               torch.log(y/(1-y)) , gt_cls_per_image, reduction="none"
+            ).sum(-1)
+            del cls_preds_
+        
+            cost = (
+                pair_wise_cls_loss
+                + 3.0 * pair_wise_iou_loss
+            )
+
+            matching_matrix = torch.zeros_like(cost)
+
+            for gt_idx in range(num_gt):
+                _, pos_idx = torch.topk(
+                    cost[gt_idx], k=dynamic_ks[gt_idx].item(), largest=False
+                )
+                matching_matrix[gt_idx][pos_idx] = 1.0
+
+            del top_k, dynamic_ks
+            anchor_matching_gt = matching_matrix.sum(0)
+            if (anchor_matching_gt > 1).sum() > 0:
+                _, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0)
+                matching_matrix[:, anchor_matching_gt > 1] *= 0.0
+                matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1.0
+            fg_mask_inboxes = matching_matrix.sum(0) > 0.0
+            matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)
+        
+            from_which_layer = from_which_layer[fg_mask_inboxes]
+            all_b = all_b[fg_mask_inboxes]
+            all_a = all_a[fg_mask_inboxes]
+            all_gj = all_gj[fg_mask_inboxes]
+            all_gi = all_gi[fg_mask_inboxes]
+            all_anch = all_anch[fg_mask_inboxes]
+        
+            this_target = this_target[matched_gt_inds]
+        
+            for i in range(nl):
+                layer_idx = from_which_layer == i
+                matching_bs[i].append(all_b[layer_idx])
+                matching_as[i].append(all_a[layer_idx])
+                matching_gjs[i].append(all_gj[layer_idx])
+                matching_gis[i].append(all_gi[layer_idx])
+                matching_targets[i].append(this_target[layer_idx])
+                matching_anchs[i].append(all_anch[layer_idx])
+
+        for i in range(nl):
+            if matching_targets[i] != []:
+                matching_bs[i] = torch.cat(matching_bs[i], dim=0)
+                matching_as[i] = torch.cat(matching_as[i], dim=0)
+                matching_gjs[i] = torch.cat(matching_gjs[i], dim=0)
+                matching_gis[i] = torch.cat(matching_gis[i], dim=0)
+                matching_targets[i] = torch.cat(matching_targets[i], dim=0)
+                matching_anchs[i] = torch.cat(matching_anchs[i], dim=0)
+            else:
+                matching_bs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_as[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_gjs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_gis[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_targets[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_anchs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+
+        return matching_bs, matching_as, matching_gjs, matching_gis, matching_targets, matching_anchs       
+
+    def find_3_positive(self, p, targets):
+        # Build targets for compute_loss(), input targets(image,class,x,y,w,h)
+        na, nt = self.na, targets.shape[0]  # number of anchors, targets
+        indices, anch = [], []
+        gain = torch.ones(7, device=targets.device).long()  # normalized to gridspace gain
+        ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt)  # same as .repeat_interleave(nt)
+        targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2)  # append anchor indices
+
+        g = 0.5  # bias
+        off = torch.tensor([[0, 0],
+                            [1, 0], [0, 1], [-1, 0], [0, -1],  # j,k,l,m
+                            # [1, 1], [1, -1], [-1, 1], [-1, -1],  # jk,jm,lk,lm
+                            ], device=targets.device).float() * g  # offsets
+
+        for i in range(self.nl):
+            anchors = self.anchors[i]
+            gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]]  # xyxy gain
+
+            # Match targets to anchors
+            t = targets * gain
+            if nt:
+                # Matches
+                r = t[:, :, 4:6] / anchors[:, None]  # wh ratio
+                j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t']  # compare
+                # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t']  # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
+                t = t[j]  # filter
+
+                # Offsets
+                gxy = t[:, 2:4]  # grid xy
+                gxi = gain[[2, 3]] - gxy  # inverse
+                j, k = ((gxy % 1. < g) & (gxy > 1.)).T
+                l, m = ((gxi % 1. < g) & (gxi > 1.)).T
+                j = torch.stack((torch.ones_like(j), j, k, l, m))
+                t = t.repeat((5, 1, 1))[j]
+                offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
+            else:
+                t = targets[0]
+                offsets = 0
+
+            # Define
+            b, c = t[:, :2].long().T  # image, class
+            gxy = t[:, 2:4]  # grid xy
+            gwh = t[:, 4:6]  # grid wh
+            gij = (gxy - offsets).long()
+            gi, gj = gij.T  # grid xy indices
+
+            # Append
+            a = t[:, 6].long()  # anchor indices
+            indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1)))  # image, anchor, grid indices
+            anch.append(anchors[a])  # anchors
+
+        return indices, anch
+
+
+class ComputeLossAuxOTA:
+    # Compute losses
+    def __init__(self, model, autobalance=False):
+        super(ComputeLossAuxOTA, self).__init__()
+        device = next(model.parameters()).device  # get model device
+        h = model.hyp  # hyperparameters
+
+        # Define criteria
+        BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
+        BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
+
+        # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
+        self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0))  # positive, negative BCE targets
+
+        # Focal loss
+        g = h['fl_gamma']  # focal loss gamma
+        if g > 0:
+            BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
+
+        det = model.module.model[-1] if is_parallel(model) else model.model[-1]  # Detect() module
+        self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02])  # P3-P7
+        self.ssi = list(det.stride).index(16) if autobalance else 0  # stride 16 index
+        self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
+        for k in 'na', 'nc', 'nl', 'anchors', 'stride':
+            setattr(self, k, getattr(det, k))
+
+    def __call__(self, p, targets, imgs):  # predictions, targets, model   
+        device = targets.device
+        lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
+        bs_aux, as_aux_, gjs_aux, gis_aux, targets_aux, anchors_aux = self.build_targets2(p[:self.nl], targets, imgs)
+        bs, as_, gjs, gis, targets, anchors = self.build_targets(p[:self.nl], targets, imgs)
+        pre_gen_gains_aux = [torch.tensor(pp.shape, device=device)[[3, 2, 3, 2]] for pp in p[:self.nl]] 
+        pre_gen_gains = [torch.tensor(pp.shape, device=device)[[3, 2, 3, 2]] for pp in p[:self.nl]] 
+    
+
+        # Losses
+        for i in range(self.nl):  # layer index, layer predictions
+            pi = p[i]
+            pi_aux = p[i+self.nl]
+            b, a, gj, gi = bs[i], as_[i], gjs[i], gis[i]  # image, anchor, gridy, gridx
+            b_aux, a_aux, gj_aux, gi_aux = bs_aux[i], as_aux_[i], gjs_aux[i], gis_aux[i]  # image, anchor, gridy, gridx
+            tobj = torch.zeros_like(pi[..., 0], device=device)  # target obj
+            tobj_aux = torch.zeros_like(pi_aux[..., 0], device=device)  # target obj
+
+            n = b.shape[0]  # number of targets
+            if n:
+                ps = pi[b, a, gj, gi]  # prediction subset corresponding to targets
+
+                # Regression
+                grid = torch.stack([gi, gj], dim=1)
+                pxy = ps[:, :2].sigmoid() * 2. - 0.5
+                pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
+                pbox = torch.cat((pxy, pwh), 1)  # predicted box
+                selected_tbox = targets[i][:, 2:6] * pre_gen_gains[i]
+                selected_tbox[:, :2] -= grid
+                iou = bbox_iou(pbox.T, selected_tbox, x1y1x2y2=False, CIoU=True)  # iou(prediction, target)
+                lbox += (1.0 - iou).mean()  # iou loss
+
+                # Objectness
+                tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype)  # iou ratio
+
+                # Classification
+                selected_tcls = targets[i][:, 1].long()
+                if self.nc > 1:  # cls loss (only if multiple classes)
+                    t = torch.full_like(ps[:, 5:], self.cn, device=device)  # targets
+                    t[range(n), selected_tcls] = self.cp
+                    lcls += self.BCEcls(ps[:, 5:], t)  # BCE
+
+                # Append targets to text file
+                # with open('targets.txt', 'a') as file:
+                #     [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
+            
+            n_aux = b_aux.shape[0]  # number of targets
+            if n_aux:
+                ps_aux = pi_aux[b_aux, a_aux, gj_aux, gi_aux]  # prediction subset corresponding to targets
+                grid_aux = torch.stack([gi_aux, gj_aux], dim=1)
+                pxy_aux = ps_aux[:, :2].sigmoid() * 2. - 0.5
+                #pxy_aux = ps_aux[:, :2].sigmoid() * 3. - 1.
+                pwh_aux = (ps_aux[:, 2:4].sigmoid() * 2) ** 2 * anchors_aux[i]
+                pbox_aux = torch.cat((pxy_aux, pwh_aux), 1)  # predicted box
+                selected_tbox_aux = targets_aux[i][:, 2:6] * pre_gen_gains_aux[i]
+                selected_tbox_aux[:, :2] -= grid_aux
+                iou_aux = bbox_iou(pbox_aux.T, selected_tbox_aux, x1y1x2y2=False, CIoU=True)  # iou(prediction, target)
+                lbox += 0.25 * (1.0 - iou_aux).mean()  # iou loss
+
+                # Objectness
+                tobj_aux[b_aux, a_aux, gj_aux, gi_aux] = (1.0 - self.gr) + self.gr * iou_aux.detach().clamp(0).type(tobj_aux.dtype)  # iou ratio
+
+                # Classification
+                selected_tcls_aux = targets_aux[i][:, 1].long()
+                if self.nc > 1:  # cls loss (only if multiple classes)
+                    t_aux = torch.full_like(ps_aux[:, 5:], self.cn, device=device)  # targets
+                    t_aux[range(n_aux), selected_tcls_aux] = self.cp
+                    lcls += 0.25 * self.BCEcls(ps_aux[:, 5:], t_aux)  # BCE
+
+            obji = self.BCEobj(pi[..., 4], tobj)
+            obji_aux = self.BCEobj(pi_aux[..., 4], tobj_aux)
+            lobj += obji * self.balance[i] + 0.25 * obji_aux * self.balance[i] # obj loss
+            if self.autobalance:
+                self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
+
+        if self.autobalance:
+            self.balance = [x / self.balance[self.ssi] for x in self.balance]
+        lbox *= self.hyp['box']
+        lobj *= self.hyp['obj']
+        lcls *= self.hyp['cls']
+        bs = tobj.shape[0]  # batch size
+
+        loss = lbox + lobj + lcls
+        return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
+
+    def build_targets(self, p, targets, imgs):
+        
+        indices, anch = self.find_3_positive(p, targets)
+
+        matching_bs = [[] for pp in p]
+        matching_as = [[] for pp in p]
+        matching_gjs = [[] for pp in p]
+        matching_gis = [[] for pp in p]
+        matching_targets = [[] for pp in p]
+        matching_anchs = [[] for pp in p]
+        
+        nl = len(p)    
+    
+        for batch_idx in range(p[0].shape[0]):
+        
+            b_idx = targets[:, 0]==batch_idx
+            this_target = targets[b_idx]
+            if this_target.shape[0] == 0:
+                continue
+                
+            txywh = this_target[:, 2:6] * imgs[batch_idx].shape[1]
+            txyxy = xywh2xyxy(txywh)
+
+            pxyxys = []
+            p_cls = []
+            p_obj = []
+            from_which_layer = []
+            all_b = []
+            all_a = []
+            all_gj = []
+            all_gi = []
+            all_anch = []
+            
+            for i, pi in enumerate(p):
+                
+                b, a, gj, gi = indices[i]
+                idx = (b == batch_idx)
+                b, a, gj, gi = b[idx], a[idx], gj[idx], gi[idx]                
+                all_b.append(b)
+                all_a.append(a)
+                all_gj.append(gj)
+                all_gi.append(gi)
+                all_anch.append(anch[i][idx])
+                from_which_layer.append(torch.ones(size=(len(b),)) * i)
+                
+                fg_pred = pi[b, a, gj, gi]                
+                p_obj.append(fg_pred[:, 4:5])
+                p_cls.append(fg_pred[:, 5:])
+                
+                grid = torch.stack([gi, gj], dim=1)
+                pxy = (fg_pred[:, :2].sigmoid() * 2. - 0.5 + grid) * self.stride[i] #/ 8.
+                #pxy = (fg_pred[:, :2].sigmoid() * 3. - 1. + grid) * self.stride[i]
+                pwh = (fg_pred[:, 2:4].sigmoid() * 2) ** 2 * anch[i][idx] * self.stride[i] #/ 8.
+                pxywh = torch.cat([pxy, pwh], dim=-1)
+                pxyxy = xywh2xyxy(pxywh)
+                pxyxys.append(pxyxy)
+            
+            pxyxys = torch.cat(pxyxys, dim=0)
+            if pxyxys.shape[0] == 0:
+                continue
+            p_obj = torch.cat(p_obj, dim=0)
+            p_cls = torch.cat(p_cls, dim=0)
+            from_which_layer = torch.cat(from_which_layer, dim=0)
+            all_b = torch.cat(all_b, dim=0)
+            all_a = torch.cat(all_a, dim=0)
+            all_gj = torch.cat(all_gj, dim=0)
+            all_gi = torch.cat(all_gi, dim=0)
+            all_anch = torch.cat(all_anch, dim=0)
+        
+            pair_wise_iou = box_iou(txyxy, pxyxys)
+
+            pair_wise_iou_loss = -torch.log(pair_wise_iou + 1e-8)
+
+            top_k, _ = torch.topk(pair_wise_iou, min(20, pair_wise_iou.shape[1]), dim=1)
+            dynamic_ks = torch.clamp(top_k.sum(1).int(), min=1)
+
+            gt_cls_per_image = (
+                F.one_hot(this_target[:, 1].to(torch.int64), self.nc)
+                .float()
+                .unsqueeze(1)
+                .repeat(1, pxyxys.shape[0], 1)
+            )
+
+            num_gt = this_target.shape[0]
+            cls_preds_ = (
+                p_cls.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
+                * p_obj.unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
+            )
+
+            y = cls_preds_.sqrt_()
+            pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
+               torch.log(y/(1-y)) , gt_cls_per_image, reduction="none"
+            ).sum(-1)
+            del cls_preds_
+        
+            cost = (
+                pair_wise_cls_loss
+                + 3.0 * pair_wise_iou_loss
+            )
+
+            matching_matrix = torch.zeros_like(cost)
+
+            for gt_idx in range(num_gt):
+                _, pos_idx = torch.topk(
+                    cost[gt_idx], k=dynamic_ks[gt_idx].item(), largest=False
+                )
+                matching_matrix[gt_idx][pos_idx] = 1.0
+
+            del top_k, dynamic_ks
+            anchor_matching_gt = matching_matrix.sum(0)
+            if (anchor_matching_gt > 1).sum() > 0:
+                _, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0)
+                matching_matrix[:, anchor_matching_gt > 1] *= 0.0
+                matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1.0
+            fg_mask_inboxes = matching_matrix.sum(0) > 0.0
+            matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)
+        
+            from_which_layer = from_which_layer[fg_mask_inboxes]
+            all_b = all_b[fg_mask_inboxes]
+            all_a = all_a[fg_mask_inboxes]
+            all_gj = all_gj[fg_mask_inboxes]
+            all_gi = all_gi[fg_mask_inboxes]
+            all_anch = all_anch[fg_mask_inboxes]
+        
+            this_target = this_target[matched_gt_inds]
+        
+            for i in range(nl):
+                layer_idx = from_which_layer == i
+                matching_bs[i].append(all_b[layer_idx])
+                matching_as[i].append(all_a[layer_idx])
+                matching_gjs[i].append(all_gj[layer_idx])
+                matching_gis[i].append(all_gi[layer_idx])
+                matching_targets[i].append(this_target[layer_idx])
+                matching_anchs[i].append(all_anch[layer_idx])
+
+        for i in range(nl):
+            if matching_targets[i] != []:
+                matching_bs[i] = torch.cat(matching_bs[i], dim=0)
+                matching_as[i] = torch.cat(matching_as[i], dim=0)
+                matching_gjs[i] = torch.cat(matching_gjs[i], dim=0)
+                matching_gis[i] = torch.cat(matching_gis[i], dim=0)
+                matching_targets[i] = torch.cat(matching_targets[i], dim=0)
+                matching_anchs[i] = torch.cat(matching_anchs[i], dim=0)
+            else:
+                matching_bs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_as[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_gjs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_gis[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_targets[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_anchs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+
+        return matching_bs, matching_as, matching_gjs, matching_gis, matching_targets, matching_anchs
+
+    def build_targets2(self, p, targets, imgs):
+        
+        indices, anch = self.find_5_positive(p, targets)
+
+        matching_bs = [[] for pp in p]
+        matching_as = [[] for pp in p]
+        matching_gjs = [[] for pp in p]
+        matching_gis = [[] for pp in p]
+        matching_targets = [[] for pp in p]
+        matching_anchs = [[] for pp in p]
+        
+        nl = len(p)    
+    
+        for batch_idx in range(p[0].shape[0]):
+        
+            b_idx = targets[:, 0]==batch_idx
+            this_target = targets[b_idx]
+            if this_target.shape[0] == 0:
+                continue
+                
+            txywh = this_target[:, 2:6] * imgs[batch_idx].shape[1]
+            txyxy = xywh2xyxy(txywh)
+
+            pxyxys = []
+            p_cls = []
+            p_obj = []
+            from_which_layer = []
+            all_b = []
+            all_a = []
+            all_gj = []
+            all_gi = []
+            all_anch = []
+            
+            for i, pi in enumerate(p):
+                
+                b, a, gj, gi = indices[i]
+                idx = (b == batch_idx)
+                b, a, gj, gi = b[idx], a[idx], gj[idx], gi[idx]                
+                all_b.append(b)
+                all_a.append(a)
+                all_gj.append(gj)
+                all_gi.append(gi)
+                all_anch.append(anch[i][idx])
+                from_which_layer.append(torch.ones(size=(len(b),)) * i)
+                
+                fg_pred = pi[b, a, gj, gi]                
+                p_obj.append(fg_pred[:, 4:5])
+                p_cls.append(fg_pred[:, 5:])
+                
+                grid = torch.stack([gi, gj], dim=1)
+                pxy = (fg_pred[:, :2].sigmoid() * 2. - 0.5 + grid) * self.stride[i] #/ 8.
+                #pxy = (fg_pred[:, :2].sigmoid() * 3. - 1. + grid) * self.stride[i]
+                pwh = (fg_pred[:, 2:4].sigmoid() * 2) ** 2 * anch[i][idx] * self.stride[i] #/ 8.
+                pxywh = torch.cat([pxy, pwh], dim=-1)
+                pxyxy = xywh2xyxy(pxywh)
+                pxyxys.append(pxyxy)
+            
+            pxyxys = torch.cat(pxyxys, dim=0)
+            if pxyxys.shape[0] == 0:
+                continue
+            p_obj = torch.cat(p_obj, dim=0)
+            p_cls = torch.cat(p_cls, dim=0)
+            from_which_layer = torch.cat(from_which_layer, dim=0)
+            all_b = torch.cat(all_b, dim=0)
+            all_a = torch.cat(all_a, dim=0)
+            all_gj = torch.cat(all_gj, dim=0)
+            all_gi = torch.cat(all_gi, dim=0)
+            all_anch = torch.cat(all_anch, dim=0)
+        
+            pair_wise_iou = box_iou(txyxy, pxyxys)
+
+            pair_wise_iou_loss = -torch.log(pair_wise_iou + 1e-8)
+
+            top_k, _ = torch.topk(pair_wise_iou, min(20, pair_wise_iou.shape[1]), dim=1)
+            dynamic_ks = torch.clamp(top_k.sum(1).int(), min=1)
+
+            gt_cls_per_image = (
+                F.one_hot(this_target[:, 1].to(torch.int64), self.nc)
+                .float()
+                .unsqueeze(1)
+                .repeat(1, pxyxys.shape[0], 1)
+            )
+
+            num_gt = this_target.shape[0]
+            cls_preds_ = (
+                p_cls.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
+                * p_obj.unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
+            )
+
+            y = cls_preds_.sqrt_()
+            pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
+               torch.log(y/(1-y)) , gt_cls_per_image, reduction="none"
+            ).sum(-1)
+            del cls_preds_
+        
+            cost = (
+                pair_wise_cls_loss
+                + 3.0 * pair_wise_iou_loss
+            )
+
+            matching_matrix = torch.zeros_like(cost)
+
+            for gt_idx in range(num_gt):
+                _, pos_idx = torch.topk(
+                    cost[gt_idx], k=dynamic_ks[gt_idx].item(), largest=False
+                )
+                matching_matrix[gt_idx][pos_idx] = 1.0
+
+            del top_k, dynamic_ks
+            anchor_matching_gt = matching_matrix.sum(0)
+            if (anchor_matching_gt > 1).sum() > 0:
+                _, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0)
+                matching_matrix[:, anchor_matching_gt > 1] *= 0.0
+                matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1.0
+            fg_mask_inboxes = matching_matrix.sum(0) > 0.0
+            matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)
+        
+            from_which_layer = from_which_layer[fg_mask_inboxes]
+            all_b = all_b[fg_mask_inboxes]
+            all_a = all_a[fg_mask_inboxes]
+            all_gj = all_gj[fg_mask_inboxes]
+            all_gi = all_gi[fg_mask_inboxes]
+            all_anch = all_anch[fg_mask_inboxes]
+        
+            this_target = this_target[matched_gt_inds]
+        
+            for i in range(nl):
+                layer_idx = from_which_layer == i
+                matching_bs[i].append(all_b[layer_idx])
+                matching_as[i].append(all_a[layer_idx])
+                matching_gjs[i].append(all_gj[layer_idx])
+                matching_gis[i].append(all_gi[layer_idx])
+                matching_targets[i].append(this_target[layer_idx])
+                matching_anchs[i].append(all_anch[layer_idx])
+
+        for i in range(nl):
+            if matching_targets[i] != []:
+                matching_bs[i] = torch.cat(matching_bs[i], dim=0)
+                matching_as[i] = torch.cat(matching_as[i], dim=0)
+                matching_gjs[i] = torch.cat(matching_gjs[i], dim=0)
+                matching_gis[i] = torch.cat(matching_gis[i], dim=0)
+                matching_targets[i] = torch.cat(matching_targets[i], dim=0)
+                matching_anchs[i] = torch.cat(matching_anchs[i], dim=0)
+            else:
+                matching_bs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_as[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_gjs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_gis[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_targets[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+                matching_anchs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
+
+        return matching_bs, matching_as, matching_gjs, matching_gis, matching_targets, matching_anchs              
+
+    def find_5_positive(self, p, targets):
+        # Build targets for compute_loss(), input targets(image,class,x,y,w,h)
+        na, nt = self.na, targets.shape[0]  # number of anchors, targets
+        indices, anch = [], []
+        gain = torch.ones(7, device=targets.device).long()  # normalized to gridspace gain
+        ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt)  # same as .repeat_interleave(nt)
+        targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2)  # append anchor indices
+
+        g = 1.0  # bias
+        off = torch.tensor([[0, 0],
+                            [1, 0], [0, 1], [-1, 0], [0, -1],  # j,k,l,m
+                            # [1, 1], [1, -1], [-1, 1], [-1, -1],  # jk,jm,lk,lm
+                            ], device=targets.device).float() * g  # offsets
+
+        for i in range(self.nl):
+            anchors = self.anchors[i]
+            gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]]  # xyxy gain
+
+            # Match targets to anchors
+            t = targets * gain
+            if nt:
+                # Matches
+                r = t[:, :, 4:6] / anchors[:, None]  # wh ratio
+                j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t']  # compare
+                # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t']  # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
+                t = t[j]  # filter
+
+                # Offsets
+                gxy = t[:, 2:4]  # grid xy
+                gxi = gain[[2, 3]] - gxy  # inverse
+                j, k = ((gxy % 1. < g) & (gxy > 1.)).T
+                l, m = ((gxi % 1. < g) & (gxi > 1.)).T
+                j = torch.stack((torch.ones_like(j), j, k, l, m))
+                t = t.repeat((5, 1, 1))[j]
+                offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
+            else:
+                t = targets[0]
+                offsets = 0
+
+            # Define
+            b, c = t[:, :2].long().T  # image, class
+            gxy = t[:, 2:4]  # grid xy
+            gwh = t[:, 4:6]  # grid wh
+            gij = (gxy - offsets).long()
+            gi, gj = gij.T  # grid xy indices
+
+            # Append
+            a = t[:, 6].long()  # anchor indices
+            indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1)))  # image, anchor, grid indices
+            anch.append(anchors[a])  # anchors
+
+        return indices, anch                 
+
+    def find_3_positive(self, p, targets):
+        # Build targets for compute_loss(), input targets(image,class,x,y,w,h)
+        na, nt = self.na, targets.shape[0]  # number of anchors, targets
+        indices, anch = [], []
+        gain = torch.ones(7, device=targets.device).long()  # normalized to gridspace gain
+        ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt)  # same as .repeat_interleave(nt)
+        targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2)  # append anchor indices
+
+        g = 0.5  # bias
+        off = torch.tensor([[0, 0],
+                            [1, 0], [0, 1], [-1, 0], [0, -1],  # j,k,l,m
+                            # [1, 1], [1, -1], [-1, 1], [-1, -1],  # jk,jm,lk,lm
+                            ], device=targets.device).float() * g  # offsets
+
+        for i in range(self.nl):
+            anchors = self.anchors[i]
+            gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]]  # xyxy gain
+
+            # Match targets to anchors
+            t = targets * gain
+            if nt:
+                # Matches
+                r = t[:, :, 4:6] / anchors[:, None]  # wh ratio
+                j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t']  # compare
+                # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t']  # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
+                t = t[j]  # filter
+
+                # Offsets
+                gxy = t[:, 2:4]  # grid xy
+                gxi = gain[[2, 3]] - gxy  # inverse
+                j, k = ((gxy % 1. < g) & (gxy > 1.)).T
+                l, m = ((gxi % 1. < g) & (gxi > 1.)).T
+                j = torch.stack((torch.ones_like(j), j, k, l, m))
+                t = t.repeat((5, 1, 1))[j]
+                offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
+            else:
+                t = targets[0]
+                offsets = 0
+
+            # Define
+            b, c = t[:, :2].long().T  # image, class
+            gxy = t[:, 2:4]  # grid xy
+            gwh = t[:, 4:6]  # grid wh
+            gij = (gxy - offsets).long()
+            gi, gj = gij.T  # grid xy indices
+
+            # Append
+            a = t[:, 6].long()  # anchor indices
+            indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1)))  # image, anchor, grid indices
+            anch.append(anchors[a])  # anchors
+
+        return indices, anch

utils/metrics.py

@@ -0,0 +1,227 @@
+# Model validation metrics
+
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+
+from . import general
+
+
+def fitness(x):
+    # Model fitness as a weighted combination of metrics
+    w = [0.0, 0.0, 0.1, 0.9]  # weights for [P, R, mAP@0.5, mAP@0.5:0.95]
+    return (x[:, :4] * w).sum(1)
+
+
+def ap_per_class(tp, conf, pred_cls, target_cls, v5_metric=False, plot=False, save_dir='.', names=()):
+    """ Compute the average precision, given the recall and precision curves.
+    Source: https://github.com/rafaelpadilla/Object-Detection-Metrics.
+    # Arguments
+        tp:  True positives (nparray, nx1 or nx10).
+        conf:  Objectness value from 0-1 (nparray).
+        pred_cls:  Predicted object classes (nparray).
+        target_cls:  True object classes (nparray).
+        plot:  Plot precision-recall curve at mAP@0.5
+        save_dir:  Plot save directory
+    # Returns
+        The average precision as computed in py-faster-rcnn.
+    """
+
+    # Sort by objectness
+    i = np.argsort(-conf)
+    tp, conf, pred_cls = tp[i], conf[i], pred_cls[i]
+
+    # Find unique classes
+    unique_classes = np.unique(target_cls)
+    nc = unique_classes.shape[0]  # number of classes, number of detections
+
+    # Create Precision-Recall curve and compute AP for each class
+    px, py = np.linspace(0, 1, 1000), []  # for plotting
+    ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000))
+    for ci, c in enumerate(unique_classes):
+        i = pred_cls == c
+        n_l = (target_cls == c).sum()  # number of labels
+        n_p = i.sum()  # number of predictions
+
+        if n_p == 0 or n_l == 0:
+            continue
+        else:
+            # Accumulate FPs and TPs
+            fpc = (1 - tp[i]).cumsum(0)
+            tpc = tp[i].cumsum(0)
+
+            # Recall
+            recall = tpc / (n_l + 1e-16)  # recall curve
+            r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0)  # negative x, xp because xp decreases
+
+            # Precision
+            precision = tpc / (tpc + fpc)  # precision curve
+            p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1)  # p at pr_score
+
+            # AP from recall-precision curve
+            for j in range(tp.shape[1]):
+                ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j], v5_metric=v5_metric)
+                if plot and j == 0:
+                    py.append(np.interp(px, mrec, mpre))  # precision at mAP@0.5
+
+    # Compute F1 (harmonic mean of precision and recall)
+    f1 = 2 * p * r / (p + r + 1e-16)
+    if plot:
+        plot_pr_curve(px, py, ap, Path(save_dir) / 'PR_curve.png', names)
+        plot_mc_curve(px, f1, Path(save_dir) / 'F1_curve.png', names, ylabel='F1')
+        plot_mc_curve(px, p, Path(save_dir) / 'P_curve.png', names, ylabel='Precision')
+        plot_mc_curve(px, r, Path(save_dir) / 'R_curve.png', names, ylabel='Recall')
+
+    i = f1.mean(0).argmax()  # max F1 index
+    return p[:, i], r[:, i], ap, f1[:, i], unique_classes.astype('int32')
+
+
+def compute_ap(recall, precision, v5_metric=False):
+    """ Compute the average precision, given the recall and precision curves
+    # Arguments
+        recall:    The recall curve (list)
+        precision: The precision curve (list)
+        v5_metric: Assume maximum recall to be 1.0, as in YOLOv5, MMDetetion etc.
+    # Returns
+        Average precision, precision curve, recall curve
+    """
+
+    # Append sentinel values to beginning and end
+    if v5_metric:  # New YOLOv5 metric, same as MMDetection and Detectron2 repositories
+        mrec = np.concatenate(([0.], recall, [1.0]))
+    else:  # Old YOLOv5 metric, i.e. default YOLOv7 metric
+        mrec = np.concatenate(([0.], recall, [recall[-1] + 0.01]))
+    mpre = np.concatenate(([1.], precision, [0.]))
+
+    # Compute the precision envelope
+    mpre = np.flip(np.maximum.accumulate(np.flip(mpre)))
+
+    # Integrate area under curve
+    method = 'interp'  # methods: 'continuous', 'interp'
+    if method == 'interp':
+        x = np.linspace(0, 1, 101)  # 101-point interp (COCO)
+        ap = np.trapz(np.interp(x, mrec, mpre), x)  # integrate
+    else:  # 'continuous'
+        i = np.where(mrec[1:] != mrec[:-1])[0]  # points where x axis (recall) changes
+        ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])  # area under curve
+
+    return ap, mpre, mrec
+
+
+class ConfusionMatrix:
+    # Updated version of https://github.com/kaanakan/object_detection_confusion_matrix
+    def __init__(self, nc, conf=0.25, iou_thres=0.45):
+        self.matrix = np.zeros((nc + 1, nc + 1))
+        self.nc = nc  # number of classes
+        self.conf = conf
+        self.iou_thres = iou_thres
+
+    def process_batch(self, detections, labels):
+        """
+        Return intersection-over-union (Jaccard index) of boxes.
+        Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
+        Arguments:
+            detections (Array[N, 6]), x1, y1, x2, y2, conf, class
+            labels (Array[M, 5]), class, x1, y1, x2, y2
+        Returns:
+            None, updates confusion matrix accordingly
+        """
+        detections = detections[detections[:, 4] > self.conf]
+        gt_classes = labels[:, 0].int()
+        detection_classes = detections[:, 5].int()
+        iou = general.box_iou(labels[:, 1:], detections[:, :4])
+
+        x = torch.where(iou > self.iou_thres)
+        if x[0].shape[0]:
+            matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()
+            if x[0].shape[0] > 1:
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
+        else:
+            matches = np.zeros((0, 3))
+
+        n = matches.shape[0] > 0
+        m0, m1, _ = matches.transpose().astype(np.int16)
+        for i, gc in enumerate(gt_classes):
+            j = m0 == i
+            if n and sum(j) == 1:
+                self.matrix[gc, detection_classes[m1[j]]] += 1  # correct
+            else:
+                self.matrix[self.nc, gc] += 1  # background FP
+
+        if n:
+            for i, dc in enumerate(detection_classes):
+                if not any(m1 == i):
+                    self.matrix[dc, self.nc] += 1  # background FN
+
+    def matrix(self):
+        return self.matrix
+
+    def plot(self, save_dir='', names=()):
+        try:
+            import seaborn as sn
+
+            array = self.matrix / (self.matrix.sum(0).reshape(1, self.nc + 1) + 1E-6)  # normalize
+            array[array < 0.005] = np.nan  # don't annotate (would appear as 0.00)
+
+            fig = plt.figure(figsize=(12, 9), tight_layout=True)
+            sn.set(font_scale=1.0 if self.nc < 50 else 0.8)  # for label size
+            labels = (0 < len(names) < 99) and len(names) == self.nc  # apply names to ticklabels
+            sn.heatmap(array, annot=self.nc < 30, annot_kws={"size": 8}, cmap='Blues', fmt='.2f', square=True,
+                       xticklabels=names + ['background FP'] if labels else "auto",
+                       yticklabels=names + ['background FN'] if labels else "auto").set_facecolor((1, 1, 1))
+            fig.axes[0].set_xlabel('True')
+            fig.axes[0].set_ylabel('Predicted')
+            fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250)
+        except Exception as e:
+            pass
+
+    def print(self):
+        for i in range(self.nc + 1):
+            print(' '.join(map(str, self.matrix[i])))
+
+
+# Plots ----------------------------------------------------------------------------------------------------------------
+
+def plot_pr_curve(px, py, ap, save_dir='pr_curve.png', names=()):
+    # Precision-recall curve
+    fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
+    py = np.stack(py, axis=1)
+
+    if 0 < len(names) < 21:  # display per-class legend if < 21 classes
+        for i, y in enumerate(py.T):
+            ax.plot(px, y, linewidth=1, label=f'{names[i]} {ap[i, 0]:.3f}')  # plot(recall, precision)
+    else:
+        ax.plot(px, py, linewidth=1, color='grey')  # plot(recall, precision)
+
+    ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean())
+    ax.set_xlabel('Recall')
+    ax.set_ylabel('Precision')
+    ax.set_xlim(0, 1)
+    ax.set_ylim(0, 1)
+    plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
+    fig.savefig(Path(save_dir), dpi=250)
+
+
+def plot_mc_curve(px, py, save_dir='mc_curve.png', names=(), xlabel='Confidence', ylabel='Metric'):
+    # Metric-confidence curve
+    fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
+
+    if 0 < len(names) < 21:  # display per-class legend if < 21 classes
+        for i, y in enumerate(py):
+            ax.plot(px, y, linewidth=1, label=f'{names[i]}')  # plot(confidence, metric)
+    else:
+        ax.plot(px, py.T, linewidth=1, color='grey')  # plot(confidence, metric)
+
+    y = py.mean(0)
+    ax.plot(px, y, linewidth=3, color='blue', label=f'all classes {y.max():.2f} at {px[y.argmax()]:.3f}')
+    ax.set_xlabel(xlabel)
+    ax.set_ylabel(ylabel)
+    ax.set_xlim(0, 1)
+    ax.set_ylim(0, 1)
+    plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
+    fig.savefig(Path(save_dir), dpi=250)

utils/plots.py

@@ -0,0 +1,489 @@
+# Plotting utils
+
+import glob
+import math
+import os
+import random
+from copy import copy
+from pathlib import Path
+
+import cv2
+import matplotlib
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+import seaborn as sns
+import torch
+import yaml
+from PIL import Image, ImageDraw, ImageFont
+from scipy.signal import butter, filtfilt
+
+from utils.general import xywh2xyxy, xyxy2xywh
+from utils.metrics import fitness
+
+# Settings
+matplotlib.rc('font', **{'size': 11})
+matplotlib.use('Agg')  # for writing to files only
+
+
+def color_list():
+    # Return first 10 plt colors as (r,g,b) https://stackoverflow.com/questions/51350872/python-from-color-name-to-rgb
+    def hex2rgb(h):
+        return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4))
+
+    return [hex2rgb(h) for h in matplotlib.colors.TABLEAU_COLORS.values()]  # or BASE_ (8), CSS4_ (148), XKCD_ (949)
+
+
+def hist2d(x, y, n=100):
+    # 2d histogram used in labels.png and evolve.png
+    xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n)
+    hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges))
+    xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1)
+    yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1)
+    return np.log(hist[xidx, yidx])
+
+
+def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5):
+    # https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy
+    def butter_lowpass(cutoff, fs, order):
+        nyq = 0.5 * fs
+        normal_cutoff = cutoff / nyq
+        return butter(order, normal_cutoff, btype='low', analog=False)
+
+    b, a = butter_lowpass(cutoff, fs, order=order)
+    return filtfilt(b, a, data)  # forward-backward filter
+
+
+def plot_one_box(x, img, color=None, label=None, line_thickness=3):
+    # Plots one bounding box on image img
+    tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1  # line/font thickness
+    color = color or [random.randint(0, 255) for _ in range(3)]
+    c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3]))
+    cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA)
+    if label:
+        tf = max(tl - 1, 1)  # font thickness
+        t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0]
+        c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3
+        cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA)  # filled
+        cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA)
+
+
+def plot_one_box_PIL(box, img, color=None, label=None, line_thickness=None):
+    img = Image.fromarray(img)
+    draw = ImageDraw.Draw(img)
+    line_thickness = line_thickness or max(int(min(img.size) / 200), 2)
+    draw.rectangle(box, width=line_thickness, outline=tuple(color))  # plot
+    if label:
+        fontsize = max(round(max(img.size) / 40), 12)
+        font = ImageFont.truetype("Arial.ttf", fontsize)
+        txt_width, txt_height = font.getsize(label)
+        draw.rectangle([box[0], box[1] - txt_height + 4, box[0] + txt_width, box[1]], fill=tuple(color))
+        draw.text((box[0], box[1] - txt_height + 1), label, fill=(255, 255, 255), font=font)
+    return np.asarray(img)
+
+
+def plot_wh_methods():  # from utils.plots import *; plot_wh_methods()
+    # Compares the two methods for width-height anchor multiplication
+    # https://github.com/ultralytics/yolov3/issues/168
+    x = np.arange(-4.0, 4.0, .1)
+    ya = np.exp(x)
+    yb = torch.sigmoid(torch.from_numpy(x)).numpy() * 2
+
+    fig = plt.figure(figsize=(6, 3), tight_layout=True)
+    plt.plot(x, ya, '.-', label='YOLOv3')
+    plt.plot(x, yb ** 2, '.-', label='YOLOR ^2')
+    plt.plot(x, yb ** 1.6, '.-', label='YOLOR ^1.6')
+    plt.xlim(left=-4, right=4)
+    plt.ylim(bottom=0, top=6)
+    plt.xlabel('input')
+    plt.ylabel('output')
+    plt.grid()
+    plt.legend()
+    fig.savefig('comparison.png', dpi=200)
+
+
+def output_to_target(output):
+    # Convert model output to target format [batch_id, class_id, x, y, w, h, conf]
+    targets = []
+    for i, o in enumerate(output):
+        for *box, conf, cls in o.cpu().numpy():
+            targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf])
+    return np.array(targets)
+
+
+def plot_images(images, targets, paths=None, fname='images.jpg', names=None, max_size=640, max_subplots=16):
+    # Plot image grid with labels
+
+    if isinstance(images, torch.Tensor):
+        images = images.cpu().float().numpy()
+    if isinstance(targets, torch.Tensor):
+        targets = targets.cpu().numpy()
+
+    # un-normalise
+    if np.max(images[0]) <= 1:
+        images *= 255
+
+    tl = 3  # line thickness
+    tf = max(tl - 1, 1)  # font thickness
+    bs, _, h, w = images.shape  # batch size, _, height, width
+    bs = min(bs, max_subplots)  # limit plot images
+    ns = np.ceil(bs ** 0.5)  # number of subplots (square)
+
+    # Check if we should resize
+    scale_factor = max_size / max(h, w)
+    if scale_factor < 1:
+        h = math.ceil(scale_factor * h)
+        w = math.ceil(scale_factor * w)
+
+    colors = color_list()  # list of colors
+    mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8)  # init
+    for i, img in enumerate(images):
+        if i == max_subplots:  # if last batch has fewer images than we expect
+            break
+
+        block_x = int(w * (i // ns))
+        block_y = int(h * (i % ns))
+
+        img = img.transpose(1, 2, 0)
+        if scale_factor < 1:
+            img = cv2.resize(img, (w, h))
+
+        mosaic[block_y:block_y + h, block_x:block_x + w, :] = img
+        if len(targets) > 0:
+            image_targets = targets[targets[:, 0] == i]
+            boxes = xywh2xyxy(image_targets[:, 2:6]).T
+            classes = image_targets[:, 1].astype('int')
+            labels = image_targets.shape[1] == 6  # labels if no conf column
+            conf = None if labels else image_targets[:, 6]  # check for confidence presence (label vs pred)
+
+            if boxes.shape[1]:
+                if boxes.max() <= 1.01:  # if normalized with tolerance 0.01
+                    boxes[[0, 2]] *= w  # scale to pixels
+                    boxes[[1, 3]] *= h
+                elif scale_factor < 1:  # absolute coords need scale if image scales
+                    boxes *= scale_factor
+            boxes[[0, 2]] += block_x
+            boxes[[1, 3]] += block_y
+            for j, box in enumerate(boxes.T):
+                cls = int(classes[j])
+                color = colors[cls % len(colors)]
+                cls = names[cls] if names else cls
+                if labels or conf[j] > 0.25:  # 0.25 conf thresh
+                    label = '%s' % cls if labels else '%s %.1f' % (cls, conf[j])
+                    plot_one_box(box, mosaic, label=label, color=color, line_thickness=tl)
+
+        # Draw image filename labels
+        if paths:
+            label = Path(paths[i]).name[:40]  # trim to 40 char
+            t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0]
+            cv2.putText(mosaic, label, (block_x + 5, block_y + t_size[1] + 5), 0, tl / 3, [220, 220, 220], thickness=tf,
+                        lineType=cv2.LINE_AA)
+
+        # Image border
+        cv2.rectangle(mosaic, (block_x, block_y), (block_x + w, block_y + h), (255, 255, 255), thickness=3)
+
+    if fname:
+        r = min(1280. / max(h, w) / ns, 1.0)  # ratio to limit image size
+        mosaic = cv2.resize(mosaic, (int(ns * w * r), int(ns * h * r)), interpolation=cv2.INTER_AREA)
+        # cv2.imwrite(fname, cv2.cvtColor(mosaic, cv2.COLOR_BGR2RGB))  # cv2 save
+        Image.fromarray(mosaic).save(fname)  # PIL save
+    return mosaic
+
+
+def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''):
+    # Plot LR simulating training for full epochs
+    optimizer, scheduler = copy(optimizer), copy(scheduler)  # do not modify originals
+    y = []
+    for _ in range(epochs):
+        scheduler.step()
+        y.append(optimizer.param_groups[0]['lr'])
+    plt.plot(y, '.-', label='LR')
+    plt.xlabel('epoch')
+    plt.ylabel('LR')
+    plt.grid()
+    plt.xlim(0, epochs)
+    plt.ylim(0)
+    plt.savefig(Path(save_dir) / 'LR.png', dpi=200)
+    plt.close()
+
+
+def plot_test_txt():  # from utils.plots import *; plot_test()
+    # Plot test.txt histograms
+    x = np.loadtxt('test.txt', dtype=np.float32)
+    box = xyxy2xywh(x[:, :4])
+    cx, cy = box[:, 0], box[:, 1]
+
+    fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True)
+    ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0)
+    ax.set_aspect('equal')
+    plt.savefig('hist2d.png', dpi=300)
+
+    fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True)
+    ax[0].hist(cx, bins=600)
+    ax[1].hist(cy, bins=600)
+    plt.savefig('hist1d.png', dpi=200)
+
+
+def plot_targets_txt():  # from utils.plots import *; plot_targets_txt()
+    # Plot targets.txt histograms
+    x = np.loadtxt('targets.txt', dtype=np.float32).T
+    s = ['x targets', 'y targets', 'width targets', 'height targets']
+    fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)
+    ax = ax.ravel()
+    for i in range(4):
+        ax[i].hist(x[i], bins=100, label='%.3g +/- %.3g' % (x[i].mean(), x[i].std()))
+        ax[i].legend()
+        ax[i].set_title(s[i])
+    plt.savefig('targets.jpg', dpi=200)
+
+
+def plot_study_txt(path='', x=None):  # from utils.plots import *; plot_study_txt()
+    # Plot study.txt generated by test.py
+    fig, ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True)
+    # ax = ax.ravel()
+
+    fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True)
+    # for f in [Path(path) / f'study_coco_{x}.txt' for x in ['yolor-p6', 'yolor-w6', 'yolor-e6', 'yolor-d6']]:
+    for f in sorted(Path(path).glob('study*.txt')):
+        y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T
+        x = np.arange(y.shape[1]) if x is None else np.array(x)
+        s = ['P', 'R', 'mAP@.5', 'mAP@.5:.95', 't_inference (ms/img)', 't_NMS (ms/img)', 't_total (ms/img)']
+        # for i in range(7):
+        #     ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8)
+        #     ax[i].set_title(s[i])
+
+        j = y[3].argmax() + 1
+        ax2.plot(y[6, 1:j], y[3, 1:j] * 1E2, '.-', linewidth=2, markersize=8,
+                 label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO'))
+
+    ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5],
+             'k.-', linewidth=2, markersize=8, alpha=.25, label='EfficientDet')
+
+    ax2.grid(alpha=0.2)
+    ax2.set_yticks(np.arange(20, 60, 5))
+    ax2.set_xlim(0, 57)
+    ax2.set_ylim(30, 55)
+    ax2.set_xlabel('GPU Speed (ms/img)')
+    ax2.set_ylabel('COCO AP val')
+    ax2.legend(loc='lower right')
+    plt.savefig(str(Path(path).name) + '.png', dpi=300)
+
+
+def plot_labels(labels, names=(), save_dir=Path(''), loggers=None):
+    # plot dataset labels
+    print('Plotting labels... ')
+    c, b = labels[:, 0], labels[:, 1:].transpose()  # classes, boxes
+    nc = int(c.max() + 1)  # number of classes
+    colors = color_list()
+    x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height'])
+
+    # seaborn correlogram
+    sns.pairplot(x, corner=True, diag_kind='auto', kind='hist', diag_kws=dict(bins=50), plot_kws=dict(pmax=0.9))
+    plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200)
+    plt.close()
+
+    # matplotlib labels
+    matplotlib.use('svg')  # faster
+    ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel()
+    ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8)
+    ax[0].set_ylabel('instances')
+    if 0 < len(names) < 30:
+        ax[0].set_xticks(range(len(names)))
+        ax[0].set_xticklabels(names, rotation=90, fontsize=10)
+    else:
+        ax[0].set_xlabel('classes')
+    sns.histplot(x, x='x', y='y', ax=ax[2], bins=50, pmax=0.9)
+    sns.histplot(x, x='width', y='height', ax=ax[3], bins=50, pmax=0.9)
+
+    # rectangles
+    labels[:, 1:3] = 0.5  # center
+    labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000
+    img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255)
+    for cls, *box in labels[:1000]:
+        ImageDraw.Draw(img).rectangle(box, width=1, outline=colors[int(cls) % 10])  # plot
+    ax[1].imshow(img)
+    ax[1].axis('off')
+
+    for a in [0, 1, 2, 3]:
+        for s in ['top', 'right', 'left', 'bottom']:
+            ax[a].spines[s].set_visible(False)
+
+    plt.savefig(save_dir / 'labels.jpg', dpi=200)
+    matplotlib.use('Agg')
+    plt.close()
+
+    # loggers
+    for k, v in loggers.items() or {}:
+        if k == 'wandb' and v:
+            v.log({"Labels": [v.Image(str(x), caption=x.name) for x in save_dir.glob('*labels*.jpg')]}, commit=False)
+
+
+def plot_evolution(yaml_file='data/hyp.finetune.yaml'):  # from utils.plots import *; plot_evolution()
+    # Plot hyperparameter evolution results in evolve.txt
+    with open(yaml_file) as f:
+        hyp = yaml.load(f, Loader=yaml.SafeLoader)
+    x = np.loadtxt('evolve.txt', ndmin=2)
+    f = fitness(x)
+    # weights = (f - f.min()) ** 2  # for weighted results
+    plt.figure(figsize=(10, 12), tight_layout=True)
+    matplotlib.rc('font', **{'size': 8})
+    for i, (k, v) in enumerate(hyp.items()):
+        y = x[:, i + 7]
+        # mu = (y * weights).sum() / weights.sum()  # best weighted result
+        mu = y[f.argmax()]  # best single result
+        plt.subplot(6, 5, i + 1)
+        plt.scatter(y, f, c=hist2d(y, f, 20), cmap='viridis', alpha=.8, edgecolors='none')
+        plt.plot(mu, f.max(), 'k+', markersize=15)
+        plt.title('%s = %.3g' % (k, mu), fontdict={'size': 9})  # limit to 40 characters
+        if i % 5 != 0:
+            plt.yticks([])
+        print('%15s: %.3g' % (k, mu))
+    plt.savefig('evolve.png', dpi=200)
+    print('\nPlot saved as evolve.png')
+
+
+def profile_idetection(start=0, stop=0, labels=(), save_dir=''):
+    # Plot iDetection '*.txt' per-image logs. from utils.plots import *; profile_idetection()
+    ax = plt.subplots(2, 4, figsize=(12, 6), tight_layout=True)[1].ravel()
+    s = ['Images', 'Free Storage (GB)', 'RAM Usage (GB)', 'Battery', 'dt_raw (ms)', 'dt_smooth (ms)', 'real-world FPS']
+    files = list(Path(save_dir).glob('frames*.txt'))
+    for fi, f in enumerate(files):
+        try:
+            results = np.loadtxt(f, ndmin=2).T[:, 90:-30]  # clip first and last rows
+            n = results.shape[1]  # number of rows
+            x = np.arange(start, min(stop, n) if stop else n)
+            results = results[:, x]
+            t = (results[0] - results[0].min())  # set t0=0s
+            results[0] = x
+            for i, a in enumerate(ax):
+                if i < len(results):
+                    label = labels[fi] if len(labels) else f.stem.replace('frames_', '')
+                    a.plot(t, results[i], marker='.', label=label, linewidth=1, markersize=5)
+                    a.set_title(s[i])
+                    a.set_xlabel('time (s)')
+                    # if fi == len(files) - 1:
+                    #     a.set_ylim(bottom=0)
+                    for side in ['top', 'right']:
+                        a.spines[side].set_visible(False)
+                else:
+                    a.remove()
+        except Exception as e:
+            print('Warning: Plotting error for %s; %s' % (f, e))
+
+    ax[1].legend()
+    plt.savefig(Path(save_dir) / 'idetection_profile.png', dpi=200)
+
+
+def plot_results_overlay(start=0, stop=0):  # from utils.plots import *; plot_results_overlay()
+    # Plot training 'results*.txt', overlaying train and val losses
+    s = ['train', 'train', 'train', 'Precision', 'mAP@0.5', 'val', 'val', 'val', 'Recall', 'mAP@0.5:0.95']  # legends
+    t = ['Box', 'Objectness', 'Classification', 'P-R', 'mAP-F1']  # titles
+    for f in sorted(glob.glob('results*.txt') + glob.glob('../../Downloads/results*.txt')):
+        results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T
+        n = results.shape[1]  # number of rows
+        x = range(start, min(stop, n) if stop else n)
+        fig, ax = plt.subplots(1, 5, figsize=(14, 3.5), tight_layout=True)
+        ax = ax.ravel()
+        for i in range(5):
+            for j in [i, i + 5]:
+                y = results[j, x]
+                ax[i].plot(x, y, marker='.', label=s[j])
+                # y_smooth = butter_lowpass_filtfilt(y)
+                # ax[i].plot(x, np.gradient(y_smooth), marker='.', label=s[j])
+
+            ax[i].set_title(t[i])
+            ax[i].legend()
+            ax[i].set_ylabel(f) if i == 0 else None  # add filename
+        fig.savefig(f.replace('.txt', '.png'), dpi=200)
+
+
+def plot_results(start=0, stop=0, bucket='', id=(), labels=(), save_dir=''):
+    # Plot training 'results*.txt'. from utils.plots import *; plot_results(save_dir='runs/train/exp')
+    fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True)
+    ax = ax.ravel()
+    s = ['Box', 'Objectness', 'Classification', 'Precision', 'Recall',
+         'val Box', 'val Objectness', 'val Classification', 'mAP@0.5', 'mAP@0.5:0.95']
+    if bucket:
+        # files = ['https://storage.googleapis.com/%s/results%g.txt' % (bucket, x) for x in id]
+        files = ['results%g.txt' % x for x in id]
+        c = ('gsutil cp ' + '%s ' * len(files) + '.') % tuple('gs://%s/results%g.txt' % (bucket, x) for x in id)
+        os.system(c)
+    else:
+        files = list(Path(save_dir).glob('results*.txt'))
+    assert len(files), 'No results.txt files found in %s, nothing to plot.' % os.path.abspath(save_dir)
+    for fi, f in enumerate(files):
+        try:
+            results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T
+            n = results.shape[1]  # number of rows
+            x = range(start, min(stop, n) if stop else n)
+            for i in range(10):
+                y = results[i, x]
+                if i in [0, 1, 2, 5, 6, 7]:
+                    y[y == 0] = np.nan  # don't show zero loss values
+                    # y /= y[0]  # normalize
+                label = labels[fi] if len(labels) else f.stem
+                ax[i].plot(x, y, marker='.', label=label, linewidth=2, markersize=8)
+                ax[i].set_title(s[i])
+                # if i in [5, 6, 7]:  # share train and val loss y axes
+                #     ax[i].get_shared_y_axes().join(ax[i], ax[i - 5])
+        except Exception as e:
+            print('Warning: Plotting error for %s; %s' % (f, e))
+
+    ax[1].legend()
+    fig.savefig(Path(save_dir) / 'results.png', dpi=200)
+    
+    
+def output_to_keypoint(output):
+    # Convert model output to target format [batch_id, class_id, x, y, w, h, conf]
+    targets = []
+    for i, o in enumerate(output):
+        kpts = o[:,6:]
+        o = o[:,:6]
+        for index, (*box, conf, cls) in enumerate(o.detach().cpu().numpy()):
+            targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf, *list(kpts.detach().cpu().numpy()[index])])
+    return np.array(targets)
+
+
+def plot_skeleton_kpts(im, kpts, steps, orig_shape=None):
+    #Plot the skeleton and keypointsfor coco datatset
+    palette = np.array([[255, 128, 0], [255, 153, 51], [255, 178, 102],
+                        [230, 230, 0], [255, 153, 255], [153, 204, 255],
+                        [255, 102, 255], [255, 51, 255], [102, 178, 255],
+                        [51, 153, 255], [255, 153, 153], [255, 102, 102],
+                        [255, 51, 51], [153, 255, 153], [102, 255, 102],
+                        [51, 255, 51], [0, 255, 0], [0, 0, 255], [255, 0, 0],
+                        [255, 255, 255]])
+
+    skeleton = [[16, 14], [14, 12], [17, 15], [15, 13], [12, 13], [6, 12],
+                [7, 13], [6, 7], [6, 8], [7, 9], [8, 10], [9, 11], [2, 3],
+                [1, 2], [1, 3], [2, 4], [3, 5], [4, 6], [5, 7]]
+
+    pose_limb_color = palette[[9, 9, 9, 9, 7, 7, 7, 0, 0, 0, 0, 0, 16, 16, 16, 16, 16, 16, 16]]
+    pose_kpt_color = palette[[16, 16, 16, 16, 16, 0, 0, 0, 0, 0, 0, 9, 9, 9, 9, 9, 9]]
+    radius = 5
+    num_kpts = len(kpts) // steps
+
+    for kid in range(num_kpts):
+        r, g, b = pose_kpt_color[kid]
+        x_coord, y_coord = kpts[steps * kid], kpts[steps * kid + 1]
+        if not (x_coord % 640 == 0 or y_coord % 640 == 0):
+            if steps == 3:
+                conf = kpts[steps * kid + 2]
+                if conf < 0.5:
+                    continue
+            cv2.circle(im, (int(x_coord), int(y_coord)), radius, (int(r), int(g), int(b)), -1)
+
+    for sk_id, sk in enumerate(skeleton):
+        r, g, b = pose_limb_color[sk_id]
+        pos1 = (int(kpts[(sk[0]-1)*steps]), int(kpts[(sk[0]-1)*steps+1]))
+        pos2 = (int(kpts[(sk[1]-1)*steps]), int(kpts[(sk[1]-1)*steps+1]))
+        if steps == 3:
+            conf1 = kpts[(sk[0]-1)*steps+2]
+            conf2 = kpts[(sk[1]-1)*steps+2]
+            if conf1<0.5 or conf2<0.5:
+                continue
+        if pos1[0]%640 == 0 or pos1[1]%640==0 or pos1[0]<0 or pos1[1]<0:
+            continue
+        if pos2[0] % 640 == 0 or pos2[1] % 640 == 0 or pos2[0]<0 or pos2[1]<0:
+            continue
+        cv2.line(im, pos1, pos2, (int(r), int(g), int(b)), thickness=2)

utils/torch_utils.py

@@ -0,0 +1,374 @@
+# YOLOR PyTorch utils
+
+import datetime
+import logging
+import math
+import os
+import platform
+import subprocess
+import time
+from contextlib import contextmanager
+from copy import deepcopy
+from pathlib import Path
+
+import torch
+import torch.backends.cudnn as cudnn
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision
+
+try:
+    import thop  # for FLOPS computation
+except ImportError:
+    thop = None
+logger = logging.getLogger(__name__)
+
+
+@contextmanager
+def torch_distributed_zero_first(local_rank: int):
+    """
+    Decorator to make all processes in distributed training wait for each local_master to do something.
+    """
+    if local_rank not in [-1, 0]:
+        torch.distributed.barrier()
+    yield
+    if local_rank == 0:
+        torch.distributed.barrier()
+
+
+def init_torch_seeds(seed=0):
+    # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html
+    torch.manual_seed(seed)
+    if seed == 0:  # slower, more reproducible
+        cudnn.benchmark, cudnn.deterministic = False, True
+    else:  # faster, less reproducible
+        cudnn.benchmark, cudnn.deterministic = True, False
+
+
+def date_modified(path=__file__):
+    # return human-readable file modification date, i.e. '2021-3-26'
+    t = datetime.datetime.fromtimestamp(Path(path).stat().st_mtime)
+    return f'{t.year}-{t.month}-{t.day}'
+
+
+def git_describe(path=Path(__file__).parent):  # path must be a directory
+    # return human-readable git description, i.e. v5.0-5-g3e25f1e https://git-scm.com/docs/git-describe
+    s = f'git -C {path} describe --tags --long --always'
+    try:
+        return subprocess.check_output(s, shell=True, stderr=subprocess.STDOUT).decode()[:-1]
+    except subprocess.CalledProcessError as e:
+        return ''  # not a git repository
+
+
+def select_device(device='', batch_size=None):
+    # device = 'cpu' or '0' or '0,1,2,3'
+    s = f'YOLOR 🚀 {git_describe() or date_modified()} torch {torch.__version__} '  # string
+    cpu = device.lower() == 'cpu'
+    if cpu:
+        os.environ['CUDA_VISIBLE_DEVICES'] = '-1'  # force torch.cuda.is_available() = False
+    elif device:  # non-cpu device requested
+        os.environ['CUDA_VISIBLE_DEVICES'] = device  # set environment variable
+        assert torch.cuda.is_available(), f'CUDA unavailable, invalid device {device} requested'  # check availability
+
+    cuda = not cpu and torch.cuda.is_available()
+    if cuda:
+        n = torch.cuda.device_count()
+        if n > 1 and batch_size:  # check that batch_size is compatible with device_count
+            assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}'
+        space = ' ' * len(s)
+        for i, d in enumerate(device.split(',') if device else range(n)):
+            p = torch.cuda.get_device_properties(i)
+            s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / 1024 ** 2}MB)\n"  # bytes to MB
+    else:
+        s += 'CPU\n'
+
+    logger.info(s.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else s)  # emoji-safe
+    return torch.device('cuda:0' if cuda else 'cpu')
+
+
+def time_synchronized():
+    # pytorch-accurate time
+    if torch.cuda.is_available():
+        torch.cuda.synchronize()
+    return time.time()
+
+
+def profile(x, ops, n=100, device=None):
+    # profile a pytorch module or list of modules. Example usage:
+    #     x = torch.randn(16, 3, 640, 640)  # input
+    #     m1 = lambda x: x * torch.sigmoid(x)
+    #     m2 = nn.SiLU()
+    #     profile(x, [m1, m2], n=100)  # profile speed over 100 iterations
+
+    device = device or torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
+    x = x.to(device)
+    x.requires_grad = True
+    print(torch.__version__, device.type, torch.cuda.get_device_properties(0) if device.type == 'cuda' else '')
+    print(f"\n{'Params':>12s}{'GFLOPS':>12s}{'forward (ms)':>16s}{'backward (ms)':>16s}{'input':>24s}{'output':>24s}")
+    for m in ops if isinstance(ops, list) else [ops]:
+        m = m.to(device) if hasattr(m, 'to') else m  # device
+        m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m  # type
+        dtf, dtb, t = 0., 0., [0., 0., 0.]  # dt forward, backward
+        try:
+            flops = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2  # GFLOPS
+        except:
+            flops = 0
+
+        for _ in range(n):
+            t[0] = time_synchronized()
+            y = m(x)
+            t[1] = time_synchronized()
+            try:
+                _ = y.sum().backward()
+                t[2] = time_synchronized()
+            except:  # no backward method
+                t[2] = float('nan')
+            dtf += (t[1] - t[0]) * 1000 / n  # ms per op forward
+            dtb += (t[2] - t[1]) * 1000 / n  # ms per op backward
+
+        s_in = tuple(x.shape) if isinstance(x, torch.Tensor) else 'list'
+        s_out = tuple(y.shape) if isinstance(y, torch.Tensor) else 'list'
+        p = sum(list(x.numel() for x in m.parameters())) if isinstance(m, nn.Module) else 0  # parameters
+        print(f'{p:12}{flops:12.4g}{dtf:16.4g}{dtb:16.4g}{str(s_in):>24s}{str(s_out):>24s}')
+
+
+def is_parallel(model):
+    return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel)
+
+
+def intersect_dicts(da, db, exclude=()):
+    # Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values
+    return {k: v for k, v in da.items() if k in db and not any(x in k for x in exclude) and v.shape == db[k].shape}
+
+
+def initialize_weights(model):
+    for m in model.modules():
+        t = type(m)
+        if t is nn.Conv2d:
+            pass  # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+        elif t is nn.BatchNorm2d:
+            m.eps = 1e-3
+            m.momentum = 0.03
+        elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6]:
+            m.inplace = True
+
+
+def find_modules(model, mclass=nn.Conv2d):
+    # Finds layer indices matching module class 'mclass'
+    return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)]
+
+
+def sparsity(model):
+    # Return global model sparsity
+    a, b = 0., 0.
+    for p in model.parameters():
+        a += p.numel()
+        b += (p == 0).sum()
+    return b / a
+
+
+def prune(model, amount=0.3):
+    # Prune model to requested global sparsity
+    import torch.nn.utils.prune as prune
+    print('Pruning model... ', end='')
+    for name, m in model.named_modules():
+        if isinstance(m, nn.Conv2d):
+            prune.l1_unstructured(m, name='weight', amount=amount)  # prune
+            prune.remove(m, 'weight')  # make permanent
+    print(' %.3g global sparsity' % sparsity(model))
+
+
+def fuse_conv_and_bn(conv, bn):
+    # Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/
+    fusedconv = nn.Conv2d(conv.in_channels,
+                          conv.out_channels,
+                          kernel_size=conv.kernel_size,
+                          stride=conv.stride,
+                          padding=conv.padding,
+                          groups=conv.groups,
+                          bias=True).requires_grad_(False).to(conv.weight.device)
+
+    # prepare filters
+    w_conv = conv.weight.clone().view(conv.out_channels, -1)
+    w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
+    fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape))
+
+    # prepare spatial bias
+    b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias
+    b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
+    fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)
+
+    return fusedconv
+
+
+def model_info(model, verbose=False, img_size=640):
+    # Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320]
+    n_p = sum(x.numel() for x in model.parameters())  # number parameters
+    n_g = sum(x.numel() for x in model.parameters() if x.requires_grad)  # number gradients
+    if verbose:
+        print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma'))
+        for i, (name, p) in enumerate(model.named_parameters()):
+            name = name.replace('module_list.', '')
+            print('%5g %40s %9s %12g %20s %10.3g %10.3g' %
+                  (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))
+
+    try:  # FLOPS
+        from thop import profile
+        stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32
+        img = torch.zeros((1, model.yaml.get('ch', 3), stride, stride), device=next(model.parameters()).device)  # input
+        flops = profile(deepcopy(model), inputs=(img,), verbose=False)[0] / 1E9 * 2  # stride GFLOPS
+        img_size = img_size if isinstance(img_size, list) else [img_size, img_size]  # expand if int/float
+        fs = ', %.1f GFLOPS' % (flops * img_size[0] / stride * img_size[1] / stride)  # 640x640 GFLOPS
+    except (ImportError, Exception):
+        fs = ''
+
+    logger.info(f"Model Summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}")
+
+
+def load_classifier(name='resnet101', n=2):
+    # Loads a pretrained model reshaped to n-class output
+    model = torchvision.models.__dict__[name](pretrained=True)
+
+    # ResNet model properties
+    # input_size = [3, 224, 224]
+    # input_space = 'RGB'
+    # input_range = [0, 1]
+    # mean = [0.485, 0.456, 0.406]
+    # std = [0.229, 0.224, 0.225]
+
+    # Reshape output to n classes
+    filters = model.fc.weight.shape[1]
+    model.fc.bias = nn.Parameter(torch.zeros(n), requires_grad=True)
+    model.fc.weight = nn.Parameter(torch.zeros(n, filters), requires_grad=True)
+    model.fc.out_features = n
+    return model
+
+
+def scale_img(img, ratio=1.0, same_shape=False, gs=32):  # img(16,3,256,416)
+    # scales img(bs,3,y,x) by ratio constrained to gs-multiple
+    if ratio == 1.0:
+        return img
+    else:
+        h, w = img.shape[2:]
+        s = (int(h * ratio), int(w * ratio))  # new size
+        img = F.interpolate(img, size=s, mode='bilinear', align_corners=False)  # resize
+        if not same_shape:  # pad/crop img
+            h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)]
+        return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447)  # value = imagenet mean
+
+
+def copy_attr(a, b, include=(), exclude=()):
+    # Copy attributes from b to a, options to only include [...] and to exclude [...]
+    for k, v in b.__dict__.items():
+        if (len(include) and k not in include) or k.startswith('_') or k in exclude:
+            continue
+        else:
+            setattr(a, k, v)
+
+
+class ModelEMA:
+    """ Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models
+    Keep a moving average of everything in the model state_dict (parameters and buffers).
+    This is intended to allow functionality like
+    https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage
+    A smoothed version of the weights is necessary for some training schemes to perform well.
+    This class is sensitive where it is initialized in the sequence of model init,
+    GPU assignment and distributed training wrappers.
+    """
+
+    def __init__(self, model, decay=0.9999, updates=0):
+        # Create EMA
+        self.ema = deepcopy(model.module if is_parallel(model) else model).eval()  # FP32 EMA
+        # if next(model.parameters()).device.type != 'cpu':
+        #     self.ema.half()  # FP16 EMA
+        self.updates = updates  # number of EMA updates
+        self.decay = lambda x: decay * (1 - math.exp(-x / 2000))  # decay exponential ramp (to help early epochs)
+        for p in self.ema.parameters():
+            p.requires_grad_(False)
+
+    def update(self, model):
+        # Update EMA parameters
+        with torch.no_grad():
+            self.updates += 1
+            d = self.decay(self.updates)
+
+            msd = model.module.state_dict() if is_parallel(model) else model.state_dict()  # model state_dict
+            for k, v in self.ema.state_dict().items():
+                if v.dtype.is_floating_point:
+                    v *= d
+                    v += (1. - d) * msd[k].detach()
+
+    def update_attr(self, model, include=(), exclude=('process_group', 'reducer')):
+        # Update EMA attributes
+        copy_attr(self.ema, model, include, exclude)
+
+
+class BatchNormXd(torch.nn.modules.batchnorm._BatchNorm):
+    def _check_input_dim(self, input):
+        # The only difference between BatchNorm1d, BatchNorm2d, BatchNorm3d, etc
+        # is this method that is overwritten by the sub-class
+        # This original goal of this method was for tensor sanity checks
+        # If you're ok bypassing those sanity checks (eg. if you trust your inference
+        # to provide the right dimensional inputs), then you can just use this method
+        # for easy conversion from SyncBatchNorm
+        # (unfortunately, SyncBatchNorm does not store the original class - if it did
+        #  we could return the one that was originally created)
+        return
+
+def revert_sync_batchnorm(module):
+    # this is very similar to the function that it is trying to revert:
+    # https://github.com/pytorch/pytorch/blob/c8b3686a3e4ba63dc59e5dcfe5db3430df256833/torch/nn/modules/batchnorm.py#L679
+    module_output = module
+    if isinstance(module, torch.nn.modules.batchnorm.SyncBatchNorm):
+        new_cls = BatchNormXd
+        module_output = BatchNormXd(module.num_features,
+                                               module.eps, module.momentum,
+                                               module.affine,
+                                               module.track_running_stats)
+        if module.affine:
+            with torch.no_grad():
+                module_output.weight = module.weight
+                module_output.bias = module.bias
+        module_output.running_mean = module.running_mean
+        module_output.running_var = module.running_var
+        module_output.num_batches_tracked = module.num_batches_tracked
+        if hasattr(module, "qconfig"):
+            module_output.qconfig = module.qconfig
+    for name, child in module.named_children():
+        module_output.add_module(name, revert_sync_batchnorm(child))
+    del module
+    return module_output
+
+
+class TracedModel(nn.Module):
+
+    def __init__(self, model=None, device=None, img_size=(640,640)): 
+        super(TracedModel, self).__init__()
+        
+        print(" Convert model to Traced-model... ") 
+        self.stride = model.stride
+        self.names = model.names
+        self.model = model
+
+        self.model = revert_sync_batchnorm(self.model)
+        self.model.to('cpu')
+        self.model.eval()
+
+        self.detect_layer = self.model.model[-1]
+        self.model.traced = True
+        
+        rand_example = torch.rand(1, 3, img_size, img_size)
+        
+        traced_script_module = torch.jit.trace(self.model, rand_example, strict=False)
+        #traced_script_module = torch.jit.script(self.model)
+        traced_script_module.save("traced_model.pt")
+        print(" traced_script_module saved! ")
+        self.model = traced_script_module
+        self.model.to(device)
+        self.detect_layer.to(device)
+        print(" model is traced! \n") 
+
+    def forward(self, x, augment=False, profile=False):
+        out = self.model(x)
+        out = self.detect_layer(out)
+        return out

utils/wandb_logging/__init__.py

@@ -0,0 +1 @@
+# init

utils/wandb_logging/log_dataset.py

@@ -0,0 +1,24 @@
+import argparse
+
+import yaml
+
+from wandb_utils import WandbLogger
+
+WANDB_ARTIFACT_PREFIX = 'wandb-artifact://'
+
+
+def create_dataset_artifact(opt):
+    with open(opt.data) as f:
+        data = yaml.load(f, Loader=yaml.SafeLoader)  # data dict
+    logger = WandbLogger(opt, '', None, data, job_type='Dataset Creation')
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default='data/coco.yaml', help='data.yaml path')
+    parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset')
+    parser.add_argument('--project', type=str, default='YOLOR', help='name of W&B Project')
+    opt = parser.parse_args()
+    opt.resume = False  # Explicitly disallow resume check for dataset upload job
+
+    create_dataset_artifact(opt)

utils/wandb_logging/wandb_utils.py

@@ -0,0 +1,306 @@
+import json
+import sys
+from pathlib import Path
+
+import torch
+import yaml
+from tqdm import tqdm
+
+sys.path.append(str(Path(__file__).parent.parent.parent))  # add utils/ to path
+from utils.datasets import LoadImagesAndLabels
+from utils.datasets import img2label_paths
+from utils.general import colorstr, xywh2xyxy, check_dataset
+
+try:
+    import wandb
+    from wandb import init, finish
+except ImportError:
+    wandb = None
+
+WANDB_ARTIFACT_PREFIX = 'wandb-artifact://'
+
+
+def remove_prefix(from_string, prefix=WANDB_ARTIFACT_PREFIX):
+    return from_string[len(prefix):]
+
+
+def check_wandb_config_file(data_config_file):
+    wandb_config = '_wandb.'.join(data_config_file.rsplit('.', 1))  # updated data.yaml path
+    if Path(wandb_config).is_file():
+        return wandb_config
+    return data_config_file
+
+
+def get_run_info(run_path):
+    run_path = Path(remove_prefix(run_path, WANDB_ARTIFACT_PREFIX))
+    run_id = run_path.stem
+    project = run_path.parent.stem
+    model_artifact_name = 'run_' + run_id + '_model'
+    return run_id, project, model_artifact_name
+
+
+def check_wandb_resume(opt):
+    process_wandb_config_ddp_mode(opt) if opt.global_rank not in [-1, 0] else None
+    if isinstance(opt.resume, str):
+        if opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
+            if opt.global_rank not in [-1, 0]:  # For resuming DDP runs
+                run_id, project, model_artifact_name = get_run_info(opt.resume)
+                api = wandb.Api()
+                artifact = api.artifact(project + '/' + model_artifact_name + ':latest')
+                modeldir = artifact.download()
+                opt.weights = str(Path(modeldir) / "last.pt")
+            return True
+    return None
+
+
+def process_wandb_config_ddp_mode(opt):
+    with open(opt.data) as f:
+        data_dict = yaml.load(f, Loader=yaml.SafeLoader)  # data dict
+    train_dir, val_dir = None, None
+    if isinstance(data_dict['train'], str) and data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX):
+        api = wandb.Api()
+        train_artifact = api.artifact(remove_prefix(data_dict['train']) + ':' + opt.artifact_alias)
+        train_dir = train_artifact.download()
+        train_path = Path(train_dir) / 'data/images/'
+        data_dict['train'] = str(train_path)
+
+    if isinstance(data_dict['val'], str) and data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX):
+        api = wandb.Api()
+        val_artifact = api.artifact(remove_prefix(data_dict['val']) + ':' + opt.artifact_alias)
+        val_dir = val_artifact.download()
+        val_path = Path(val_dir) / 'data/images/'
+        data_dict['val'] = str(val_path)
+    if train_dir or val_dir:
+        ddp_data_path = str(Path(val_dir) / 'wandb_local_data.yaml')
+        with open(ddp_data_path, 'w') as f:
+            yaml.dump(data_dict, f)
+        opt.data = ddp_data_path
+
+
+class WandbLogger():
+    def __init__(self, opt, name, run_id, data_dict, job_type='Training'):
+        # Pre-training routine --
+        self.job_type = job_type
+        self.wandb, self.wandb_run, self.data_dict = wandb, None if not wandb else wandb.run, data_dict
+        # It's more elegant to stick to 1 wandb.init call, but useful config data is overwritten in the WandbLogger's wandb.init call
+        if isinstance(opt.resume, str):  # checks resume from artifact
+            if opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
+                run_id, project, model_artifact_name = get_run_info(opt.resume)
+                model_artifact_name = WANDB_ARTIFACT_PREFIX + model_artifact_name
+                assert wandb, 'install wandb to resume wandb runs'
+                # Resume wandb-artifact:// runs here| workaround for not overwriting wandb.config
+                self.wandb_run = wandb.init(id=run_id, project=project, resume='allow')
+                opt.resume = model_artifact_name
+        elif self.wandb:
+            self.wandb_run = wandb.init(config=opt,
+                                        resume="allow",
+                                        project='YOLOR' if opt.project == 'runs/train' else Path(opt.project).stem,
+                                        name=name,
+                                        job_type=job_type,
+                                        id=run_id) if not wandb.run else wandb.run
+        if self.wandb_run:
+            if self.job_type == 'Training':
+                if not opt.resume:
+                    wandb_data_dict = self.check_and_upload_dataset(opt) if opt.upload_dataset else data_dict
+                    # Info useful for resuming from artifacts
+                    self.wandb_run.config.opt = vars(opt)
+                    self.wandb_run.config.data_dict = wandb_data_dict
+                self.data_dict = self.setup_training(opt, data_dict)
+            if self.job_type == 'Dataset Creation':
+                self.data_dict = self.check_and_upload_dataset(opt)
+        else:
+            prefix = colorstr('wandb: ')
+            print(f"{prefix}Install Weights & Biases for YOLOR logging with 'pip install wandb' (recommended)")
+
+    def check_and_upload_dataset(self, opt):
+        assert wandb, 'Install wandb to upload dataset'
+        check_dataset(self.data_dict)
+        config_path = self.log_dataset_artifact(opt.data,
+                                                opt.single_cls,
+                                                'YOLOR' if opt.project == 'runs/train' else Path(opt.project).stem)
+        print("Created dataset config file ", config_path)
+        with open(config_path) as f:
+            wandb_data_dict = yaml.load(f, Loader=yaml.SafeLoader)
+        return wandb_data_dict
+
+    def setup_training(self, opt, data_dict):
+        self.log_dict, self.current_epoch, self.log_imgs = {}, 0, 16  # Logging Constants
+        self.bbox_interval = opt.bbox_interval
+        if isinstance(opt.resume, str):
+            modeldir, _ = self.download_model_artifact(opt)
+            if modeldir:
+                self.weights = Path(modeldir) / "last.pt"
+                config = self.wandb_run.config
+                opt.weights, opt.save_period, opt.batch_size, opt.bbox_interval, opt.epochs, opt.hyp = str(
+                    self.weights), config.save_period, config.total_batch_size, config.bbox_interval, config.epochs, \
+                                                                                                       config.opt['hyp']
+            data_dict = dict(self.wandb_run.config.data_dict)  # eliminates the need for config file to resume
+        if 'val_artifact' not in self.__dict__:  # If --upload_dataset is set, use the existing artifact, don't download
+            self.train_artifact_path, self.train_artifact = self.download_dataset_artifact(data_dict.get('train'),
+                                                                                           opt.artifact_alias)
+            self.val_artifact_path, self.val_artifact = self.download_dataset_artifact(data_dict.get('val'),
+                                                                                       opt.artifact_alias)
+            self.result_artifact, self.result_table, self.val_table, self.weights = None, None, None, None
+            if self.train_artifact_path is not None:
+                train_path = Path(self.train_artifact_path) / 'data/images/'
+                data_dict['train'] = str(train_path)
+            if self.val_artifact_path is not None:
+                val_path = Path(self.val_artifact_path) / 'data/images/'
+                data_dict['val'] = str(val_path)
+                self.val_table = self.val_artifact.get("val")
+                self.map_val_table_path()
+        if self.val_artifact is not None:
+            self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation")
+            self.result_table = wandb.Table(["epoch", "id", "prediction", "avg_confidence"])
+        if opt.bbox_interval == -1:
+            self.bbox_interval = opt.bbox_interval = (opt.epochs // 10) if opt.epochs > 10 else 1
+        return data_dict
+
+    def download_dataset_artifact(self, path, alias):
+        if isinstance(path, str) and path.startswith(WANDB_ARTIFACT_PREFIX):
+            dataset_artifact = wandb.use_artifact(remove_prefix(path, WANDB_ARTIFACT_PREFIX) + ":" + alias)
+            assert dataset_artifact is not None, "'Error: W&B dataset artifact doesn\'t exist'"
+            datadir = dataset_artifact.download()
+            return datadir, dataset_artifact
+        return None, None
+
+    def download_model_artifact(self, opt):
+        if opt.resume.startswith(WANDB_ARTIFACT_PREFIX):
+            model_artifact = wandb.use_artifact(remove_prefix(opt.resume, WANDB_ARTIFACT_PREFIX) + ":latest")
+            assert model_artifact is not None, 'Error: W&B model artifact doesn\'t exist'
+            modeldir = model_artifact.download()
+            epochs_trained = model_artifact.metadata.get('epochs_trained')
+            total_epochs = model_artifact.metadata.get('total_epochs')
+            assert epochs_trained < total_epochs, 'training to %g epochs is finished, nothing to resume.' % (
+                total_epochs)
+            return modeldir, model_artifact
+        return None, None
+
+    def log_model(self, path, opt, epoch, fitness_score, best_model=False):
+        model_artifact = wandb.Artifact('run_' + wandb.run.id + '_model', type='model', metadata={
+            'original_url': str(path),
+            'epochs_trained': epoch + 1,
+            'save period': opt.save_period,
+            'project': opt.project,
+            'total_epochs': opt.epochs,
+            'fitness_score': fitness_score
+        })
+        model_artifact.add_file(str(path / 'last.pt'), name='last.pt')
+        wandb.log_artifact(model_artifact,
+                           aliases=['latest', 'epoch ' + str(self.current_epoch), 'best' if best_model else ''])
+        print("Saving model artifact on epoch ", epoch + 1)
+
+    def log_dataset_artifact(self, data_file, single_cls, project, overwrite_config=False):
+        with open(data_file) as f:
+            data = yaml.load(f, Loader=yaml.SafeLoader)  # data dict
+        nc, names = (1, ['item']) if single_cls else (int(data['nc']), data['names'])
+        names = {k: v for k, v in enumerate(names)}  # to index dictionary
+        self.train_artifact = self.create_dataset_table(LoadImagesAndLabels(
+            data['train']), names, name='train') if data.get('train') else None
+        self.val_artifact = self.create_dataset_table(LoadImagesAndLabels(
+            data['val']), names, name='val') if data.get('val') else None
+        if data.get('train'):
+            data['train'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'train')
+        if data.get('val'):
+            data['val'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'val')
+        path = data_file if overwrite_config else '_wandb.'.join(data_file.rsplit('.', 1))  # updated data.yaml path
+        data.pop('download', None)
+        with open(path, 'w') as f:
+            yaml.dump(data, f)
+
+        if self.job_type == 'Training':  # builds correct artifact pipeline graph
+            self.wandb_run.use_artifact(self.val_artifact)
+            self.wandb_run.use_artifact(self.train_artifact)
+            self.val_artifact.wait()
+            self.val_table = self.val_artifact.get('val')
+            self.map_val_table_path()
+        else:
+            self.wandb_run.log_artifact(self.train_artifact)
+            self.wandb_run.log_artifact(self.val_artifact)
+        return path
+
+    def map_val_table_path(self):
+        self.val_table_map = {}
+        print("Mapping dataset")
+        for i, data in enumerate(tqdm(self.val_table.data)):
+            self.val_table_map[data[3]] = data[0]
+
+    def create_dataset_table(self, dataset, class_to_id, name='dataset'):
+        # TODO: Explore multiprocessing to slpit this loop parallely| This is essential for speeding up the the logging
+        artifact = wandb.Artifact(name=name, type="dataset")
+        img_files = tqdm([dataset.path]) if isinstance(dataset.path, str) and Path(dataset.path).is_dir() else None
+        img_files = tqdm(dataset.img_files) if not img_files else img_files
+        for img_file in img_files:
+            if Path(img_file).is_dir():
+                artifact.add_dir(img_file, name='data/images')
+                labels_path = 'labels'.join(dataset.path.rsplit('images', 1))
+                artifact.add_dir(labels_path, name='data/labels')
+            else:
+                artifact.add_file(img_file, name='data/images/' + Path(img_file).name)
+                label_file = Path(img2label_paths([img_file])[0])
+                artifact.add_file(str(label_file),
+                                  name='data/labels/' + label_file.name) if label_file.exists() else None
+        table = wandb.Table(columns=["id", "train_image", "Classes", "name"])
+        class_set = wandb.Classes([{'id': id, 'name': name} for id, name in class_to_id.items()])
+        for si, (img, labels, paths, shapes) in enumerate(tqdm(dataset)):
+            height, width = shapes[0]
+            labels[:, 2:] = (xywh2xyxy(labels[:, 2:].view(-1, 4))) * torch.Tensor([width, height, width, height])
+            box_data, img_classes = [], {}
+            for cls, *xyxy in labels[:, 1:].tolist():
+                cls = int(cls)
+                box_data.append({"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]},
+                                 "class_id": cls,
+                                 "box_caption": "%s" % (class_to_id[cls]),
+                                 "scores": {"acc": 1},
+                                 "domain": "pixel"})
+                img_classes[cls] = class_to_id[cls]
+            boxes = {"ground_truth": {"box_data": box_data, "class_labels": class_to_id}}  # inference-space
+            table.add_data(si, wandb.Image(paths, classes=class_set, boxes=boxes), json.dumps(img_classes),
+                           Path(paths).name)
+        artifact.add(table, name)
+        return artifact
+
+    def log_training_progress(self, predn, path, names):
+        if self.val_table and self.result_table:
+            class_set = wandb.Classes([{'id': id, 'name': name} for id, name in names.items()])
+            box_data = []
+            total_conf = 0
+            for *xyxy, conf, cls in predn.tolist():
+                if conf >= 0.25:
+                    box_data.append(
+                        {"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]},
+                         "class_id": int(cls),
+                         "box_caption": "%s %.3f" % (names[cls], conf),
+                         "scores": {"class_score": conf},
+                         "domain": "pixel"})
+                    total_conf = total_conf + conf
+            boxes = {"predictions": {"box_data": box_data, "class_labels": names}}  # inference-space
+            id = self.val_table_map[Path(path).name]
+            self.result_table.add_data(self.current_epoch,
+                                       id,
+                                       wandb.Image(self.val_table.data[id][1], boxes=boxes, classes=class_set),
+                                       total_conf / max(1, len(box_data))
+                                       )
+
+    def log(self, log_dict):
+        if self.wandb_run:
+            for key, value in log_dict.items():
+                self.log_dict[key] = value
+
+    def end_epoch(self, best_result=False):
+        if self.wandb_run:
+            wandb.log(self.log_dict)
+            self.log_dict = {}
+            if self.result_artifact:
+                train_results = wandb.JoinedTable(self.val_table, self.result_table, "id")
+                self.result_artifact.add(train_results, 'result')
+                wandb.log_artifact(self.result_artifact, aliases=['latest', 'epoch ' + str(self.current_epoch),
+                                                                  ('best' if best_result else '')])
+                self.result_table = wandb.Table(["epoch", "id", "prediction", "avg_confidence"])
+                self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation")
+
+    def finish_run(self):
+        if self.wandb_run:
+            if self.log_dict:
+                wandb.log(self.log_dict)
+            wandb.run.finish()