first commit

app.py

@@ -0,0 +1,140 @@
+import argparse
+import requests
+import gradio as gr
+import numpy as np
+import cv2
+import torch
+import torch.nn as nn
+from PIL import Image
+from torchvision import transforms
+from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+from timm.data import create_transform
+from config import get_config
+from model import build_model
+
+# Download human-readable labels for ImageNet.
+response = requests.get("https://git.io/JJkYN")
+labels = response.text.split("\n")
+
+def parse_option():
+    parser = argparse.ArgumentParser('UniCL demo script', add_help=False)
+    parser.add_argument('--cfg', type=str, default="configs/unicl_swin_base.yaml", metavar="FILE", help='path to config file', )
+    args, unparsed = parser.parse_known_args()
+
+    config = get_config(args)
+
+    return args, config
+
+def build_transforms(img_size, center_crop=True):
+    t = [transforms.ToPILImage()]
+    if center_crop:
+        size = int((256 / 224) * img_size)
+        t.append(
+            transforms.Resize(size)
+        )
+        t.append(
+            transforms.CenterCrop(img_size)    
+        )
+    else:
+        t.append(
+            transforms.Resize(img_size)
+        )        
+    t.append(transforms.ToTensor())
+    t.append(transforms.Normalize(IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD))
+    return transforms.Compose(t)
+
+def build_transforms4display(img_size, center_crop=True):
+    t = [transforms.ToPILImage()]
+    if center_crop:
+        size = int((256 / 224) * img_size)
+        t.append(
+            transforms.Resize(size)
+        )
+        t.append(
+            transforms.CenterCrop(img_size)    
+        )
+    else:
+        t.append(
+            transforms.Resize(img_size)
+        )  
+    t.append(transforms.ToTensor())
+    return transforms.Compose(t)
+
+args, config = parse_option()
+
+'''
+build model
+'''
+model = build_model(config)
+
+url = 'https://projects4jw.blob.core.windows.net/unicl/release/in21k_yfcc14m_gcc15m_swin_base.pth'
+checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu", check_hash=True)
+model.load_state_dict(checkpoint["model"])
+model.eval()
+
+'''
+build data transform
+'''
+eval_transforms = build_transforms(224, center_crop=True)
+display_transforms = build_transforms4display(224, center_crop=True)
+
+'''
+build upsampler
+'''
+# upsampler = nn.Upsample(scale_factor=16, mode='bilinear')
+
+'''
+borrow code from here: https://github.com/jacobgil/pytorch-grad-cam/blob/master/pytorch_grad_cam/utils/image.py
+'''
+def show_cam_on_image(img: np.ndarray,
+                      mask: np.ndarray,
+                      use_rgb: bool = False,
+                      colormap: int = cv2.COLORMAP_JET) -> np.ndarray:
+    """ This function overlays the cam mask on the image as an heatmap.
+    By default the heatmap is in BGR format.
+    :param img: The base image in RGB or BGR format.
+    :param mask: The cam mask.
+    :param use_rgb: Whether to use an RGB or BGR heatmap, this should be set to True if 'img' is in RGB format.
+    :param colormap: The OpenCV colormap to be used.
+    :returns: The default image with the cam overlay.
+    """
+    heatmap = cv2.applyColorMap(np.uint8(255 * mask), colormap)
+    if use_rgb:
+        heatmap = cv2.cvtColor(heatmap, cv2.COLOR_BGR2RGB)
+    heatmap = np.float32(heatmap) / 255
+
+    if np.max(img) > 1:
+        raise Exception(
+            "The input image should np.float32 in the range [0, 1]")
+
+    cam = 0.7*heatmap + 0.3*img
+    # cam = cam / np.max(cam)
+    return np.uint8(255 * cam)
+
+def recognize_image(image, texts):
+    print(texts)
+    img_t = eval_transforms(image) 
+    img_d = display_transforms(image).permute(1, 2, 0).numpy()
+
+    text_embeddings = model.get_text_embeddings(texts.split(';'))
+
+    # compute output
+    feat_img = model.encode_image(img_t.unsqueeze(0))
+    output = model.logit_scale.exp() * feat_img @ text_embeddings.t()
+    prediction = output.softmax(-1).flatten()
+
+    return {texts.split(';')[i]: float(prediction[i]) for i in range(len(texts.split(';')))}
+
+
+image = gr.inputs.Image()
+label = gr.outputs.Label(num_top_classes=100)
+
+gr.Interface(
+    description="UniCL for Zero-shot Image Recognition Demo (https://github.com/microsoft/unicl)",
+    fn=recognize_image,
+    inputs=["image", "text"],
+    outputs=[                   
+        label,
+    ],
+    examples=[["./donut.png", 'a donut; several donuts; a number of donuts'], ["./horses.png"], ["./pencil.png"], ["./apple_with_ipod.jpg"]],
+).launch()

config.py

@@ -0,0 +1,245 @@
+# --------------------------------------------------------
+# Unified Contrastive Learning (UniCL)
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Jianwei Yang (jianwyan@microsoft.com)
+# Based on Swin Transformer written by Zhe Liu
+# --------------------------------------------------------
+
+import os
+import yaml
+from yacs.config import CfgNode as CN
+
+_C = CN()
+_C.VERBOSE = False
+
+# Base config files
+_C.BASE = ['']
+
+# -----------------------------------------------------------------------------
+# Data settings
+# -----------------------------------------------------------------------------
+_C.DATA = CN()
+# Batch size for a single GPU, could be overwritten by command line argument
+_C.DATA.BATCH_SIZE = 128
+# Path to dataset, could be overwritten by command line argument
+_C.DATA.DATA_PATH = ''
+# Dataset name
+_C.DATA.DATASET = 'imagenet'
+# Input image size
+_C.DATA.IMG_SIZE = 224
+# Interpolation to resize image (random, bilinear, bicubic)
+_C.DATA.INTERPOLATION = 'bicubic'
+# Use zipped dataset instead of folder dataset
+# could be overwritten by command line argument
+_C.DATA.ZIP_MODE = False
+# Cache Data in Memory, could be overwritten by command line argument
+_C.DATA.CACHE_MODE = 'part'
+# Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.
+_C.DATA.PIN_MEMORY = True
+# Number of data loading threads
+_C.DATA.NUM_WORKERS = 8
+
+# -----------------------------------------------------------------------------
+# Model settings
+# -----------------------------------------------------------------------------
+_C.MODEL = CN()
+# Model name
+_C.MODEL.NAME = ''
+# Checkpoint to resume, could be overwritten by command line argument
+_C.MODEL.RESUME = ''
+# Number of classes, overwritten in data preparation
+_C.MODEL.NUM_CLASSES = 0
+# Label Smoothing
+_C.MODEL.LABEL_SMOOTHING = 0.1
+# Whether load pretrained model
+_C.MODEL.PRETRAINED = ''
+# Projection dimension
+_C.MODEL.DIM_PROJECTION = 512
+# Mode specific
+_C.MODEL.SPEC = CN(new_allowed=True)
+# -----------------------------------------------------------------------------
+# Build Image Encoder
+# -----------------------------------------------------------------------------
+_C.MODEL.IMAGE_ENCODER = CN()
+# Image encoder type
+_C.MODEL.IMAGE_ENCODER.TYPE = 'swin'
+# Input image size
+_C.MODEL.IMAGE_ENCODER.IMG_SIZE = 224
+# Dropout rate
+_C.MODEL.IMAGE_ENCODER.DROP_RATE = 0.0
+# Drop path rate
+_C.MODEL.IMAGE_ENCODER.DROP_PATH_RATE = 0.1
+
+# Swin Transformer parameters
+_C.MODEL.IMAGE_ENCODER.SWIN = CN()
+_C.MODEL.IMAGE_ENCODER.SWIN.PATCH_SIZE = 4
+_C.MODEL.IMAGE_ENCODER.SWIN.IN_CHANS = 3
+_C.MODEL.IMAGE_ENCODER.SWIN.EMBED_DIM = 96
+_C.MODEL.IMAGE_ENCODER.SWIN.DEPTHS = [2, 2, 6, 2]
+_C.MODEL.IMAGE_ENCODER.SWIN.NUM_HEADS = [3, 6, 12, 24]
+_C.MODEL.IMAGE_ENCODER.SWIN.WINDOW_SIZE = 7
+_C.MODEL.IMAGE_ENCODER.SWIN.MLP_RATIO = 4.
+_C.MODEL.IMAGE_ENCODER.SWIN.QKV_BIAS = True
+_C.MODEL.IMAGE_ENCODER.SWIN.QK_SCALE = None
+_C.MODEL.IMAGE_ENCODER.SWIN.APE = False
+_C.MODEL.IMAGE_ENCODER.SWIN.PATCH_NORM = True
+
+# FocalNet parameters
+_C.MODEL.IMAGE_ENCODER.FOCAL = CN()
+_C.MODEL.IMAGE_ENCODER.FOCAL.PATCH_SIZE = 4
+_C.MODEL.IMAGE_ENCODER.FOCAL.IN_CHANS = 3
+_C.MODEL.IMAGE_ENCODER.FOCAL.EMBED_DIM = 96
+_C.MODEL.IMAGE_ENCODER.FOCAL.DEPTHS = [2, 2, 6, 2]
+_C.MODEL.IMAGE_ENCODER.FOCAL.MLP_RATIO = 4.
+_C.MODEL.IMAGE_ENCODER.FOCAL.PATCH_NORM = True
+_C.MODEL.IMAGE_ENCODER.FOCAL.FOCAL_LEVELS = [2, 2, 2, 2]
+_C.MODEL.IMAGE_ENCODER.FOCAL.FOCAL_WINDOWS = [3, 3, 3, 3]
+_C.MODEL.IMAGE_ENCODER.FOCAL.FOCAL_FACTORS = [2, 2, 2, 2]
+_C.MODEL.IMAGE_ENCODER.FOCAL.USE_CONV_EMBED = False
+_C.MODEL.IMAGE_ENCODER.FOCAL.USE_LAYERSCALE = False
+_C.MODEL.IMAGE_ENCODER.FOCAL.USE_POSTLN = False
+
+# -----------------------------------------------------------------------------
+# Build Text Encoder
+# -----------------------------------------------------------------------------
+_C.MODEL.TEXT_ENCODER = CN()
+
+_C.MODEL.TEXT_ENCODER.NAME = 'transformer'
+_C.MODEL.TEXT_ENCODER.LOAD_PRETRAINED = False
+_C.MODEL.TEXT_ENCODER.PRETRAINED = ''
+_C.MODEL.TEXT_ENCODER.TOKENIZER = 'clip'
+_C.MODEL.TEXT_ENCODER.CONTEXT_LENGTH = 77
+_C.MODEL.TEXT_ENCODER.WIDTH = 1024
+_C.MODEL.TEXT_ENCODER.HEADS = 16
+_C.MODEL.TEXT_ENCODER.LAYERS = 12
+_C.MODEL.TEXT_ENCODER.AUTOGRESSIVE = True
+
+# -----------------------------------------------------------------------------
+# Training settings
+# -----------------------------------------------------------------------------
+_C.TRAIN = CN()
+_C.TRAIN.START_EPOCH = 0
+_C.TRAIN.EPOCHS = 32
+_C.TRAIN.WARMUP_EPOCHS = 5
+_C.TRAIN.WEIGHT_DECAY = 0.1
+_C.TRAIN.BASE_LR = 5e-4
+_C.TRAIN.WARMUP_LR = 5e-7
+_C.TRAIN.MIN_LR = 5e-6
+# Clip gradient norm
+_C.TRAIN.CLIP_GRAD = 5.0
+# Auto resume from latest checkpoint
+_C.TRAIN.AUTO_RESUME = True
+# Gradient accumulation steps
+# could be overwritten by command line argument
+_C.TRAIN.ACCUMULATION_STEPS = 0
+# Whether to use gradient checkpointing to save memory
+# could be overwritten by command line argument
+_C.TRAIN.USE_CHECKPOINT = False
+
+# LR scheduler
+_C.TRAIN.LR_SCHEDULER = CN()
+_C.TRAIN.LR_SCHEDULER.NAME = 'cosine'
+# Epoch interval to decay LR, used in StepLRScheduler
+_C.TRAIN.LR_SCHEDULER.DECAY_EPOCHS = 30
+# LR decay rate, used in StepLRScheduler
+_C.TRAIN.LR_SCHEDULER.DECAY_RATE = 0.1
+
+# Optimizer
+_C.TRAIN.OPTIMIZER = CN()
+_C.TRAIN.OPTIMIZER.NAME = 'adamw'
+# Optimizer Epsilon
+_C.TRAIN.OPTIMIZER.EPS = 1e-8
+# Optimizer Betas
+_C.TRAIN.OPTIMIZER.BETAS = (0.9, 0.999)
+# SGD momentum
+_C.TRAIN.OPTIMIZER.MOMENTUM = 0.9
+
+# -----------------------------------------------------------------------------
+# Augmentation settings
+# -----------------------------------------------------------------------------
+_C.AUG = CN()
+# Color jitter factor
+_C.AUG.COLOR_JITTER = 0.4
+# Use AutoAugment policy. "v0" or "original"
+_C.AUG.AUTO_AUGMENT = 'rand-m9-mstd0.5-inc1'
+# Random erase prob
+_C.AUG.REPROB = 0.25
+# Random erase mode
+_C.AUG.REMODE = 'pixel'
+# Random erase count
+_C.AUG.RECOUNT = 1
+# Mixup alpha, mixup enabled if > 0
+_C.AUG.MIXUP = 0.8
+# Cutmix alpha, cutmix enabled if > 0
+_C.AUG.CUTMIX = 1.0
+# Cutmix min/max ratio, overrides alpha and enables cutmix if set
+_C.AUG.CUTMIX_MINMAX = None
+# Probability of performing mixup or cutmix when either/both is enabled
+_C.AUG.MIXUP_PROB = 1.0
+# Probability of switching to cutmix when both mixup and cutmix enabled
+_C.AUG.MIXUP_SWITCH_PROB = 0.5
+# How to apply mixup/cutmix params. Per "batch", "pair", or "elem"
+_C.AUG.MIXUP_MODE = 'batch'
+
+# -----------------------------------------------------------------------------
+# Testing settings
+# -----------------------------------------------------------------------------
+_C.TEST = CN()
+# Whether to use center crop when testing
+_C.TEST.CROP = True
+
+# -----------------------------------------------------------------------------
+# Misc
+# -----------------------------------------------------------------------------
+# Mixed precision opt level, if O0, no amp is used ('O0', 'O1', 'O2')
+# overwritten by command line argument
+_C.AMP_OPT_LEVEL = ''
+# Path to output folder, overwritten by command line argument
+_C.OUTPUT = ''
+# Tag of experiment, overwritten by command line argument
+_C.TAG = 'default'
+# Frequency to save checkpoint
+_C.SAVE_FREQ = 1
+# Frequency to logging info
+_C.PRINT_FREQ = 100
+# Fixed random seed
+_C.SEED = 0
+# Perform evaluation only, overwritten by command line argument
+_C.EVAL_MODE = False
+# Test throughput only, overwritten by command line argument
+_C.THROUGHPUT_MODE = False
+# Debug only so that skip dataloader initialization, overwritten by command line argument
+_C.DEBUG_MODE = False
+# local rank for DistributedDataParallel, given by command line argument
+_C.LOCAL_RANK = 0
+
+
+def _update_config_from_file(config, cfg_file):
+    config.defrost()
+    with open(cfg_file, 'r') as f:
+        yaml_cfg = yaml.load(f, Loader=yaml.FullLoader)
+
+    for cfg in yaml_cfg.setdefault('BASE', ['']):
+        if cfg:
+            _update_config_from_file(
+                config, os.path.join(os.path.dirname(cfg_file), cfg)
+            )
+    print('=> merge config from {}'.format(cfg_file))
+    config.merge_from_file(cfg_file)
+    config.freeze()
+
+
+def update_config(config, args):
+    _update_config_from_file(config, args.cfg)
+    config.freeze()
+
+
+def get_config(args):
+    """Get a yacs CfgNode object with default values."""
+    # Return a clone so that the defaults will not be altered
+    # This is for the "local variable" use pattern
+    config = _C.clone()
+    update_config(config, args)
+
+    return config

configs/unicl_focalnet_giant.yaml

@@ -0,0 +1,16 @@
+MODEL:
+  NAME: unicl_focalnet_giant
+  DIM_PROJECTION: 1024 
+  IMAGE_ENCODER:
+    TYPE: focalnet_giant_lrf
+    DROP_PATH_RATE: 0.5
+    FOCAL:
+      USE_POSTLN: False
+      USE_CONV_EMBED: False    
+      EMBED_DIM: 512
+      USE_LAYERSCALE: True
+  TEXT_ENCODER:
+    NAME: 'transformer'
+    WIDTH: 1024
+    HEADS: 16
+    LAYERS: 16   

configs/unicl_swin_base.yaml

@@ -0,0 +1,16 @@
+MODEL:
+  NAME: unicl_swin_base
+  DIM_PROJECTION: 512  
+  IMAGE_ENCODER:
+    TYPE: swin
+    DROP_PATH_RATE: 0.5
+    SWIN:
+      EMBED_DIM: 128
+      DEPTHS: [ 2, 2, 18, 2 ]
+      NUM_HEADS: [ 4, 8, 16, 32 ]
+      WINDOW_SIZE: 7
+  TEXT_ENCODER:
+    NAME: 'transformer'
+    WIDTH: 512
+    HEADS: 8
+    LAYERS: 12  

configs/unicl_swin_tiny.yaml

@@ -0,0 +1,16 @@
+MODEL:
+  NAME: unicl_swin_tiny
+  DIM_PROJECTION: 512  
+  IMAGE_ENCODER:
+    TYPE: swin
+    DROP_PATH_RATE: 0.2
+    SWIN:
+      EMBED_DIM: 96
+      DEPTHS: [ 2, 2, 6, 2 ]
+      NUM_HEADS: [ 3, 6, 12, 24 ]
+      WINDOW_SIZE: 7
+  TEXT_ENCODER:
+    NAME: 'transformer'
+    WIDTH: 512
+    HEADS: 8
+    LAYERS: 12  

model/__init__.py

@@ -0,0 +1 @@
+from .model import build_unicl_model as build_model

model/image_encoder/__init__.py

@@ -0,0 +1 @@
+from .build import build_model as build_image_encoder

model/image_encoder/build.py

@@ -0,0 +1,59 @@
+from timm.models import create_model
+from .swin_transformer import SwinTransformer
+from . import focalnet
+
+def build_model(config):
+    model_type = config.TYPE
+    print(f"Creating model: {model_type}")
+    
+    if "swin" in model_type:
+        model = SwinTransformer(
+            num_classes=0,
+            img_size=config.IMG_SIZE,
+            patch_size=config.SWIN.PATCH_SIZE,
+            in_chans=config.SWIN.IN_CHANS,
+            embed_dim=config.SWIN.EMBED_DIM,
+            depths=config.SWIN.DEPTHS,
+            num_heads=config.SWIN.NUM_HEADS,
+            window_size=config.SWIN.WINDOW_SIZE,
+            mlp_ratio=config.SWIN.MLP_RATIO,
+            qkv_bias=config.SWIN.QKV_BIAS,
+            qk_scale=config.SWIN.QK_SCALE,
+            drop_rate=config.DROP_RATE,
+            drop_path_rate=config.DROP_PATH_RATE,
+            ape=config.SWIN.APE,
+            patch_norm=config.SWIN.PATCH_NORM,
+            use_checkpoint=False
+            ) 
+    elif "focal" in model_type:
+        model = create_model(
+            model_type, 
+            pretrained=False, 
+            img_size=config.IMG_SIZE,
+            num_classes=0,
+            drop_path_rate=config.DROP_PATH_RATE,
+            use_conv_embed=config.FOCAL.USE_CONV_EMBED, 
+            use_layerscale=config.FOCAL.USE_LAYERSCALE,
+            use_postln=config.FOCAL.USE_POSTLN
+        )  
+
+    elif "vit" in model_type:
+        model = create_model(
+            model_type,
+            pretrained=is_pretrained,
+            img_size=config.DATA.IMG_SIZE,
+            num_classes=config.MODEL.NUM_CLASSES,
+        )
+    elif "resnet" in model_type:
+        model = create_model(
+            model_type,
+            pretrained=is_pretrained,
+            num_classes=config.MODEL.NUM_CLASSES
+        )
+    else:
+        model = create_model(
+            model_type,
+            pretrained=is_pretrained,
+            num_classes=config.MODEL.NUM_CLASSES
+        )        
+    return model

model/image_encoder/focalnet.py

@@ -0,0 +1,649 @@
+# --------------------------------------------------------
+# FocalNets -- Focal Modulation Networks
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Jianwei Yang (jianwyan@microsoft.com)
+# --------------------------------------------------------
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+from timm.models.registry import register_model
+
+from torchvision import transforms
+from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+from timm.data import create_transform
+from timm.data.transforms import _pil_interp
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, 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
+
+class FocalModulation(nn.Module):
+    def __init__(self, dim, focal_window, focal_level, focal_factor=2, bias=True, proj_drop=0.):
+        super().__init__()
+
+        self.dim = dim
+        self.focal_window = focal_window
+        self.focal_level = focal_level
+        self.focal_factor = focal_factor
+
+        self.f = nn.Linear(dim, 2*dim + (self.focal_level+1), bias=bias)
+        self.h = nn.Conv2d(dim, dim, kernel_size=1, stride=1, bias=bias)
+
+        self.act = nn.GELU()
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.focal_layers = nn.ModuleList()
+                
+        self.kernel_sizes = []
+        for k in range(self.focal_level):
+            kernel_size = self.focal_factor*k + self.focal_window
+            self.focal_layers.append(
+                nn.Sequential(
+                    nn.Conv2d(dim, dim, kernel_size=kernel_size, stride=1, 
+                    groups=dim, padding=kernel_size//2, bias=False),
+                    nn.GELU(),
+                    )
+                )              
+            self.kernel_sizes.append(kernel_size)          
+    def forward(self, x):
+        """
+        Args:
+            x: input features with shape of (B, H, W, C)
+        """
+        C = x.shape[-1]
+
+        # pre linear projection
+        x = self.f(x).permute(0, 3, 1, 2).contiguous()
+        q, ctx, self.gates = torch.split(x, (C, C, self.focal_level+1), 1)
+        
+        # context aggreation
+        ctx_all = 0 
+        for l in range(self.focal_level):         
+            ctx = self.focal_layers[l](ctx)
+            ctx_all = ctx_all + ctx*self.gates[:, l:l+1]
+        ctx_global = self.act(ctx.mean(2, keepdim=True).mean(3, keepdim=True))
+        ctx_all = ctx_all + ctx_global*self.gates[:,self.focal_level:]
+
+        # focal modulation
+        self.modulator = self.h(ctx_all)
+        x_out = q*self.modulator
+        x_out = x_out.permute(0, 2, 3, 1).contiguous()
+        
+        # post linear porjection
+        x_out = self.proj(x_out)
+        x_out = self.proj_drop(x_out)
+        return x_out
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}'
+
+    def flops(self, N):
+        # calculate flops for 1 window with token length of N
+        flops = 0
+
+        flops += N * self.dim * (self.dim * 2 + (self.focal_level+1))
+
+        # focal convolution
+        for k in range(self.focal_level):
+            flops += N * (self.kernel_sizes[k]**2+1) * self.dim
+
+        # global gating
+        flops += N * 1 * self.dim 
+
+        #  self.linear
+        flops += N * self.dim * (self.dim + 1)
+
+        # x = self.proj(x)
+        flops += N * self.dim * self.dim
+        return flops
+
+class FocalNetBlock(nn.Module):
+    r""" Focal Modulation Network Block.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        drop (float, optional): Dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+        focal_level (int): Number of focal levels. 
+        focal_window (int): Focal window size at first focal level
+        use_layerscale (bool): Whether use layerscale
+        layerscale_value (float): Initial layerscale value
+        use_postln (bool): Whether use layernorm after modulation
+    """
+
+    def __init__(self, dim, input_resolution, mlp_ratio=4., drop=0., drop_path=0., 
+                    act_layer=nn.GELU, norm_layer=nn.LayerNorm,
+                    focal_level=1, focal_window=3,
+                    use_layerscale=False, layerscale_value=1e-4, 
+                    use_postln=False):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.mlp_ratio = mlp_ratio
+
+        self.focal_window = focal_window
+        self.focal_level = focal_level
+        self.use_postln = use_postln
+
+        self.norm1 = norm_layer(dim)
+        self.modulation = FocalModulation(dim, proj_drop=drop, focal_window=focal_window, focal_level=self.focal_level)
+
+        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)
+
+        self.alpha = 3.0 if self.use_postln else 1.0
+
+        self.gamma_1 = 1.0
+        self.gamma_2 = 1.0    
+        if use_layerscale:
+            self.gamma_1 = nn.Parameter(layerscale_value * torch.ones((dim)), requires_grad=True)
+            self.gamma_2 = nn.Parameter(layerscale_value * torch.ones((dim)), requires_grad=True)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x):
+        H, W = self.H, self.W
+        B, L, C = x.shape
+        shortcut = x
+
+        # Focal Modulation
+        if not self.use_postln:
+            x = self.norm1(x)
+        x = x.view(B, H, W, C)
+        x = self.modulation(x).view(B, H * W, C)
+
+        # FFN
+        x = shortcut*self.alpha + self.drop_path(self.gamma_1 * x)
+        if self.use_postln:
+            x = self.norm1(x)
+
+        if not self.use_postln:
+            x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
+        else:
+            x = x*self.alpha + self.drop_path(self.gamma_2 * self.mlp(x))
+            x = self.norm2(x)
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, " \
+               f"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
+        flops += self.modulation.flops(H*W)
+
+        # mlp
+        flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
+        # norm2
+        flops += self.dim * H * W
+        return flops
+
+class BasicLayer(nn.Module):
+    """ A basic Focal Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        focal_level (int): Number of focal levels
+        focal_window (int): Focal window size at first focal level
+        use_layerscale (bool): Whether use layerscale
+        layerscale_value (float): Initial layerscale value
+        use_postln (bool): Whether use layernorm after modulation
+    """
+
+    def __init__(self, dim, out_dim, input_resolution, depth,
+                 mlp_ratio=4., drop=0., drop_path=0., norm_layer=nn.LayerNorm, 
+                 downsample=None, use_checkpoint=False, 
+                 focal_level=1, focal_window=1, 
+                 use_conv_embed=False, 
+                 use_layerscale=False, layerscale_value=1e-4, use_postln=False):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+        
+        # build blocks
+        self.blocks = nn.ModuleList([
+            FocalNetBlock(
+                dim=dim, 
+                input_resolution=input_resolution,
+                mlp_ratio=mlp_ratio, 
+                drop=drop, 
+                drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                norm_layer=norm_layer,
+                focal_level=focal_level,
+                focal_window=focal_window, 
+                use_layerscale=use_layerscale, 
+                layerscale_value=layerscale_value,
+                use_postln=use_postln, 
+            )
+            for i in range(depth)])
+
+        if downsample is not None:
+            self.downsample = downsample(
+                img_size=input_resolution, 
+                patch_size=2, 
+                in_chans=dim, 
+                embed_dim=out_dim, 
+                use_conv_embed=use_conv_embed, 
+                norm_layer=norm_layer, 
+                is_stem=False
+            )
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x = blk(x)
+
+        if self.downsample is not None:
+            x = x.transpose(1, 2).reshape(x.shape[0], -1, H, W)
+            x, Ho, Wo = self.downsample(x)
+        else:
+            Ho, Wo = H, W        
+        return x, Ho, Wo
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+    def flops(self):
+        flops = 0
+        for blk in self.blocks:
+            flops += blk.flops()
+        if self.downsample is not None:
+            flops += self.downsample.flops()
+        return flops
+
+class PatchEmbed(nn.Module):
+    r""" Image to Patch Embedding
+
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=(224, 224), patch_size=4, in_chans=3, embed_dim=96, use_conv_embed=False, norm_layer=None, is_stem=False):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        if use_conv_embed:
+            # if we choose to use conv embedding, then we treat the stem and non-stem differently
+            if is_stem:
+                kernel_size = 7; padding = 2; stride = 4
+            else:
+                kernel_size = 3; padding = 1; stride = 2
+            self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding)
+        else:
+            self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+
+        x = self.proj(x)        
+        H, W = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)  # B Ph*Pw C
+        if self.norm is not None:
+            x = self.norm(x)
+        return x, H, W
+
+    def flops(self):
+        Ho, Wo = self.patches_resolution
+        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
+        if self.norm is not None:
+            flops += Ho * Wo * self.embed_dim
+        return flops
+
+class FocalNet(nn.Module):
+    r""" Focal Modulation Networks (FocalNets)
+
+    Args:
+        img_size (int | tuple(int)): Input image size. Default 224
+        patch_size (int | tuple(int)): Patch size. Default: 4
+        in_chans (int): Number of input image channels. Default: 3
+        num_classes (int): Number of classes for classification head. Default: 1000
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each Focal Transformer layer.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        drop_rate (float): Dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False 
+        focal_levels (list): How many focal levels at all stages. Note that this excludes the finest-grain level. Default: [1, 1, 1, 1] 
+        focal_windows (list): The focal window size at all stages. Default: [7, 5, 3, 1] 
+        use_conv_embed (bool): Whether use convolutional embedding. We noted that using convolutional embedding usually improve the performance, but we do not use it by default. Default: False 
+        use_layerscale (bool): Whether use layerscale proposed in CaiT. Default: False 
+        layerscale_value (float): Value for layer scale. Default: 1e-4 
+        use_postln (bool): Whether use layernorm after modulation (it helps stablize training of large models)
+    """
+    def __init__(self, 
+                img_size=224, 
+                patch_size=4, 
+                in_chans=3, 
+                num_classes=1000,
+                embed_dim=96, 
+                depths=[2, 2, 6, 2], 
+                mlp_ratio=4., 
+                drop_rate=0., 
+                drop_path_rate=0.1,
+                norm_layer=nn.LayerNorm, 
+                patch_norm=True,
+                use_checkpoint=False,                 
+                focal_levels=[2, 2, 2, 2], 
+                focal_windows=[3, 3, 3, 3], 
+                use_conv_embed=False, 
+                use_layerscale=False, 
+                layerscale_value=1e-4, 
+                use_postln=False, 
+                **kwargs):
+        super().__init__()
+
+        self.num_layers = len(depths)
+        embed_dim = [embed_dim * (2 ** i) for i in range(self.num_layers)]
+
+        self.num_classes = num_classes
+        self.embed_dim = embed_dim
+        self.patch_norm = patch_norm
+        self.num_features = embed_dim[-1]
+        self.mlp_ratio = mlp_ratio
+        
+        # split image into patches using either non-overlapped embedding or overlapped embedding
+        self.patch_embed = PatchEmbed(
+            img_size=to_2tuple(img_size), 
+            patch_size=patch_size, 
+            in_chans=in_chans, 
+            embed_dim=embed_dim[0], 
+            use_conv_embed=use_conv_embed, 
+            norm_layer=norm_layer if self.patch_norm else None, 
+            is_stem=True)
+
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.patches_resolution
+        self.patches_resolution = patches_resolution
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(dim=embed_dim[i_layer], 
+                               out_dim=embed_dim[i_layer+1] if (i_layer < self.num_layers - 1) else None,  
+                               input_resolution=(patches_resolution[0] // (2 ** i_layer),
+                                                 patches_resolution[1] // (2 ** i_layer)),
+                               depth=depths[i_layer],
+                               mlp_ratio=self.mlp_ratio,
+                               drop=drop_rate, 
+                               drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                               norm_layer=norm_layer, 
+                               downsample=PatchEmbed if (i_layer < self.num_layers - 1) else None,
+                               focal_level=focal_levels[i_layer], 
+                               focal_window=focal_windows[i_layer], 
+                               use_conv_embed=use_conv_embed,
+                               use_checkpoint=use_checkpoint, 
+                               use_layerscale=use_layerscale, 
+                               layerscale_value=layerscale_value, 
+                               use_postln=use_postln,
+                    )
+            self.layers.append(layer)
+
+        self.norm = norm_layer(self.num_features)
+        self.avgpool = nn.AdaptiveAvgPool1d(1)
+        self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
+        self.dim_out = self.num_features
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {''}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {''}
+
+    def forward_features(self, x):
+        x, H, W = self.patch_embed(x)
+        x = self.pos_drop(x)
+
+        for layer in self.layers:
+            x, H, W = layer(x, H, W)
+        x = self.norm(x)  # B L C
+        x = self.avgpool(x.transpose(1, 2))  # B C 1
+        x = torch.flatten(x, 1)
+        return x
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        x = self.head(x)
+        return x
+
+    def flops(self):
+        flops = 0
+        flops += self.patch_embed.flops()
+        for i, layer in enumerate(self.layers):
+            flops += layer.flops()
+        flops += self.num_features * self.patches_resolution[0] * self.patches_resolution[1] // (2 ** self.num_layers)
+        flops += self.num_features * self.num_classes
+        return flops
+
+def build_transforms(img_size, center_crop=False):
+    t = []
+    if center_crop:
+        size = int((256 / 224) * img_size)
+        t.append(
+            transforms.Resize(size, interpolation=_pil_interp('bicubic'))
+        )
+        t.append(
+            transforms.CenterCrop(img_size)    
+        )
+    else:
+        t.append(
+            transforms.Resize(img_size, interpolation=_pil_interp('bicubic'))
+        )        
+    t.append(transforms.ToTensor())
+    t.append(transforms.Normalize(IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD))
+    return transforms.Compose(t)
+
+def build_transforms4display(img_size, center_crop=False):
+    t = []
+    if center_crop:
+        size = int((256 / 224) * img_size)
+        t.append(
+            transforms.Resize(size, interpolation=_pil_interp('bicubic'))
+        )
+        t.append(
+            transforms.CenterCrop(img_size)    
+        )
+    else:
+        t.append(
+            transforms.Resize(img_size, interpolation=_pil_interp('bicubic'))
+        )  
+    t.append(transforms.ToTensor())
+    return transforms.Compose(t)
+
+model_urls = {
+    "focalnet_tiny_srf": "",
+    "focalnet_small_srf": "",
+    "focalnet_base_srf": "",
+    "focalnet_tiny_lrf": "",
+    "focalnet_small_lrf": "",
+    "focalnet_base_lrf": "",    
+}
+
+@register_model
+def focalnet_tiny_srf(pretrained=False, **kwargs):
+    model = FocalNet(depths=[2, 2, 6, 2], embed_dim=96, **kwargs)
+    if pretrained:
+        url = model_urls['focalnet_tiny_srf']
+        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu", check_hash=True)
+        model.load_state_dict(checkpoint["model"])
+    return model
+
+@register_model
+def focalnet_small_srf(pretrained=False, **kwargs):
+    model = FocalNet(depths=[2, 2, 18, 2], embed_dim=96, **kwargs)
+    if pretrained:
+        url = model_urls['focalnet_small_srf']
+        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
+        model.load_state_dict(checkpoint["model"])
+    return model
+
+@register_model
+def focalnet_base_srf(pretrained=False, **kwargs):
+    model = FocalNet(depths=[2, 2, 18, 2], embed_dim=128, **kwargs)
+    if pretrained:
+        url = model_urls['focalnet_base_srf']
+        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
+        model.load_state_dict(checkpoint["model"])
+    return model
+
+@register_model
+def focalnet_tiny_lrf(pretrained=False, **kwargs):
+    model = FocalNet(depths=[2, 2, 6, 2], embed_dim=96, focal_levels=[3, 3, 3, 3], **kwargs)
+    if pretrained:
+        url = model_urls['focalnet_tiny_lrf']
+        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu", check_hash=True)
+        model.load_state_dict(checkpoint["model"])
+    return model
+
+@register_model
+def focalnet_small_lrf(pretrained=False, **kwargs):
+    model = FocalNet(depths=[2, 2, 18, 2], embed_dim=96, focal_levels=[3, 3, 3, 3], **kwargs)
+    if pretrained:
+        url = model_urls['focalnet_small_lrf']
+        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
+        model.load_state_dict(checkpoint["model"])
+    return model
+
+@register_model
+def focalnet_base_lrf(pretrained=False, **kwargs):
+    model = FocalNet(depths=[2, 2, 18, 2], embed_dim=128, focal_levels=[3, 3, 3, 3], **kwargs)
+    if pretrained:
+        url = model_urls['focalnet_base_lrf']
+        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
+        model.load_state_dict(checkpoint["model"])
+    return model
+
+@register_model
+def focalnet_giant_lrf(pretrained=False, **kwargs):
+    model = FocalNet(depths=[2, 2, 42, 2], embed_dim=512, focal_levels=[3, 3, 3, 3], **kwargs)
+    if pretrained:
+        url = model_urls['focalnet_giant_lrf']
+        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
+        model.load_state_dict(checkpoint["model"])
+    return model
+
+@register_model
+def focalnet_tiny_iso_16(pretrained=False, **kwargs):
+    model = FocalNet(depths=[12], patch_size=16, embed_dim=192, focal_levels=[3], focal_windows=[3], **kwargs)
+    if pretrained:
+        url = model_urls['focalnet_tiny_iso_16']
+        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu", check_hash=True)
+        model.load_state_dict(checkpoint["model"])
+    return model
+
+@register_model
+def focalnet_small_iso_16(pretrained=False, **kwargs):
+    model = FocalNet(depths=[12], patch_size=16, embed_dim=384, focal_levels=[3], focal_windows=[3], **kwargs)
+    if pretrained:
+        url = model_urls['focalnet_small_iso_16']
+        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
+        model.load_state_dict(checkpoint["model"])
+    return model
+
+@register_model
+def focalnet_base_iso_16(pretrained=False, **kwargs):
+    model = FocalNet(depths=[12], patch_size=16, embed_dim=768, focal_levels=[3], focal_windows=[3], use_layerscale=True, use_postln=True, **kwargs)
+    if pretrained:
+        url = model_urls['focalnet_base_iso_16']
+        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
+        model.load_state_dict(checkpoint["model"])
+    return model
+
+if __name__ == '__main__':
+    img_size = 224
+    x = torch.rand(16, 3, img_size, img_size).cuda()
+    # model = FocalNet(depths=[2, 2, 6, 2], embed_dim=96)
+    # model = FocalNet(depths=[12], patch_size=16, embed_dim=768, focal_levels=[3], focal_windows=[3], focal_factors=[2])
+    model = FocalNet(depths=[2, 2, 6, 2], embed_dim=96, focal_levels=[3, 3, 3, 3]).cuda()
+    print(model); model(x)
+
+    flops = model.flops()
+    print(f"number of GFLOPs: {flops / 1e9}")
+
+    n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    print(f"number of params: {n_parameters}")