add adapter code

extralibs/midas/api.py

@@ -0,0 +1,171 @@
+# based on https://github.com/isl-org/MiDaS
+
+import cv2
+import torch
+import torch.nn as nn
+from torchvision.transforms import Compose
+
+from extralibs.midas.midas.dpt_depth import DPTDepthModel
+from extralibs.midas.midas.midas_net import MidasNet
+from extralibs.midas.midas.midas_net_custom import MidasNet_small
+from extralibs.midas.midas.transforms import Resize, NormalizeImage, PrepareForNet
+
+
+ISL_PATHS = {
+    "dpt_large": "midas_models/dpt_large-midas-2f21e586.pt",
+    "dpt_hybrid": "midas_models/dpt_hybrid-midas-501f0c75.pt",
+    "midas_v21": "",
+    "midas_v21_small": "",
+}
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+def load_midas_transform(model_type):
+    # https://github.com/isl-org/MiDaS/blob/master/run.py
+    # load transform only
+    if model_type == "dpt_large":  # DPT-Large
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "dpt_hybrid":  # DPT-Hybrid
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "midas_v21":
+        net_w, net_h = 384, 384
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+
+    elif model_type == "midas_v21_small":
+        net_w, net_h = 256, 256
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+
+    else:
+        assert False, f"model_type '{model_type}' not implemented, use: --model_type large"
+
+    transform = Compose(
+        [
+            Resize(
+                net_w,
+                net_h,
+                resize_target=None,
+                keep_aspect_ratio=True,
+                ensure_multiple_of=32,
+                resize_method=resize_mode,
+                image_interpolation_method=cv2.INTER_CUBIC,
+            ),
+            normalization,
+            PrepareForNet(),
+        ]
+    )
+
+    return transform
+
+
+def load_model(model_type, model_path=None):
+    # https://github.com/isl-org/MiDaS/blob/master/run.py
+    # load network
+    if model_path is None:
+        model_path = ISL_PATHS[model_type]
+    if model_type == "dpt_large":  # DPT-Large
+        model = DPTDepthModel(
+            path=model_path,
+            backbone="vitl16_384",
+            non_negative=True,
+        )
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "dpt_hybrid":  # DPT-Hybrid
+        model = DPTDepthModel(
+            path=model_path,
+            backbone="vitb_rn50_384",
+            non_negative=True,
+        )
+        net_w, net_h = 384, 384
+        resize_mode = "minimal"
+        normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    elif model_type == "midas_v21":
+        model = MidasNet(model_path, non_negative=True)
+        net_w, net_h = 384, 384
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(
+            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+        )
+
+    elif model_type == "midas_v21_small":
+        model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True,
+                               non_negative=True, blocks={'expand': True})
+        net_w, net_h = 256, 256
+        resize_mode = "upper_bound"
+        normalization = NormalizeImage(
+            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
+        )
+
+    else:
+        print(f"model_type '{model_type}' not implemented, use: --model_type large")
+        assert False
+
+    transform = Compose(
+        [
+            Resize(
+                net_w,
+                net_h,
+                resize_target=None,
+                keep_aspect_ratio=True,
+                ensure_multiple_of=32,
+                resize_method=resize_mode,
+                image_interpolation_method=cv2.INTER_CUBIC,
+            ),
+            normalization,
+            PrepareForNet(),
+        ]
+    )
+
+    return model.eval(), transform
+
+
+class MiDaSInference(nn.Module):
+    MODEL_TYPES_TORCH_HUB = [
+        "DPT_Large",
+        "DPT_Hybrid",
+        "MiDaS_small"
+    ]
+    MODEL_TYPES_ISL = [
+        "dpt_large",
+        "dpt_hybrid",
+        "midas_v21",
+        "midas_v21_small",
+    ]
+
+    def __init__(self, model_type, model_path):
+        super().__init__()
+        assert (model_type in self.MODEL_TYPES_ISL)
+        model, _ = load_model(model_type, model_path)
+        self.model = model
+        self.model.train = disabled_train
+
+    def forward(self, x):
+        # x in 0..1 as produced by calling self.transform on a 0..1 float64 numpy array
+        # NOTE: we expect that the correct transform has been called during dataloading.
+        with torch.no_grad():
+            prediction = self.model(x)
+            prediction = torch.nn.functional.interpolate(
+                prediction.unsqueeze(1),
+                size=x.shape[2:],
+                mode="bicubic",
+                align_corners=False,
+            )
+        assert prediction.shape == (x.shape[0], 1, x.shape[2], x.shape[3])
+        return prediction
+

extralibs/midas/midas/base_model.py

@@ -0,0 +1,16 @@
+import torch
+
+
+class BaseModel(torch.nn.Module):
+    def load(self, path):
+        """Load model from file.
+
+        Args:
+            path (str): file path
+        """
+        parameters = torch.load(path, map_location=torch.device('cpu'))
+
+        if "optimizer" in parameters:
+            parameters = parameters["model"]
+
+        self.load_state_dict(parameters)

extralibs/midas/midas/blocks.py

@@ -0,0 +1,342 @@
+import torch
+import torch.nn as nn
+
+from .vit import (
+    _make_pretrained_vitb_rn50_384,
+    _make_pretrained_vitl16_384,
+    _make_pretrained_vitb16_384,
+    forward_vit,
+)
+
+def _make_encoder(backbone, features, use_pretrained, groups=1, expand=False, exportable=True, hooks=None, use_vit_only=False, use_readout="ignore",):
+    if backbone == "vitl16_384":
+        pretrained = _make_pretrained_vitl16_384(
+            use_pretrained, hooks=hooks, use_readout=use_readout
+        )
+        scratch = _make_scratch(
+            [256, 512, 1024, 1024], features, groups=groups, expand=expand
+        )  # ViT-L/16 - 85.0% Top1 (backbone)
+    elif backbone == "vitb_rn50_384":
+        pretrained = _make_pretrained_vitb_rn50_384(
+            use_pretrained,
+            hooks=hooks,
+            use_vit_only=use_vit_only,
+            use_readout=use_readout,
+        )
+        scratch = _make_scratch(
+            [256, 512, 768, 768], features, groups=groups, expand=expand
+        )  # ViT-H/16 - 85.0% Top1 (backbone)
+    elif backbone == "vitb16_384":
+        pretrained = _make_pretrained_vitb16_384(
+            use_pretrained, hooks=hooks, use_readout=use_readout
+        )
+        scratch = _make_scratch(
+            [96, 192, 384, 768], features, groups=groups, expand=expand
+        )  # ViT-B/16 - 84.6% Top1 (backbone)
+    elif backbone == "resnext101_wsl":
+        pretrained = _make_pretrained_resnext101_wsl(use_pretrained)
+        scratch = _make_scratch([256, 512, 1024, 2048], features, groups=groups, expand=expand)     # efficientnet_lite3  
+    elif backbone == "efficientnet_lite3":
+        pretrained = _make_pretrained_efficientnet_lite3(use_pretrained, exportable=exportable)
+        scratch = _make_scratch([32, 48, 136, 384], features, groups=groups, expand=expand)  # efficientnet_lite3     
+    else:
+        print(f"Backbone '{backbone}' not implemented")
+        assert False
+        
+    return pretrained, scratch
+
+
+def _make_scratch(in_shape, out_shape, groups=1, expand=False):
+    scratch = nn.Module()
+
+    out_shape1 = out_shape
+    out_shape2 = out_shape
+    out_shape3 = out_shape
+    out_shape4 = out_shape
+    if expand==True:
+        out_shape1 = out_shape
+        out_shape2 = out_shape*2
+        out_shape3 = out_shape*4
+        out_shape4 = out_shape*8
+
+    scratch.layer1_rn = nn.Conv2d(
+        in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer2_rn = nn.Conv2d(
+        in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer3_rn = nn.Conv2d(
+        in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+    scratch.layer4_rn = nn.Conv2d(
+        in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups
+    )
+
+    return scratch
+
+
+def _make_pretrained_efficientnet_lite3(use_pretrained, exportable=False):
+    efficientnet = torch.hub.load(
+        "rwightman/gen-efficientnet-pytorch",
+        "tf_efficientnet_lite3",
+        pretrained=use_pretrained,
+        exportable=exportable
+    )
+    return _make_efficientnet_backbone(efficientnet)
+
+
+def _make_efficientnet_backbone(effnet):
+    pretrained = nn.Module()
+
+    pretrained.layer1 = nn.Sequential(
+        effnet.conv_stem, effnet.bn1, effnet.act1, *effnet.blocks[0:2]
+    )
+    pretrained.layer2 = nn.Sequential(*effnet.blocks[2:3])
+    pretrained.layer3 = nn.Sequential(*effnet.blocks[3:5])
+    pretrained.layer4 = nn.Sequential(*effnet.blocks[5:9])
+
+    return pretrained
+    
+
+def _make_resnet_backbone(resnet):
+    pretrained = nn.Module()
+    pretrained.layer1 = nn.Sequential(
+        resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool, resnet.layer1
+    )
+
+    pretrained.layer2 = resnet.layer2
+    pretrained.layer3 = resnet.layer3
+    pretrained.layer4 = resnet.layer4
+
+    return pretrained
+
+
+def _make_pretrained_resnext101_wsl(use_pretrained):
+    resnet = torch.hub.load("facebookresearch/WSL-Images", "resnext101_32x8d_wsl")
+    return _make_resnet_backbone(resnet)
+
+
+
+class Interpolate(nn.Module):
+    """Interpolation module.
+    """
+
+    def __init__(self, scale_factor, mode, align_corners=False):
+        """Init.
+
+        Args:
+            scale_factor (float): scaling
+            mode (str): interpolation mode
+        """
+        super(Interpolate, self).__init__()
+
+        self.interp = nn.functional.interpolate
+        self.scale_factor = scale_factor
+        self.mode = mode
+        self.align_corners = align_corners
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: interpolated data
+        """
+
+        x = self.interp(
+            x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners
+        )
+
+        return x
+
+
+class ResidualConvUnit(nn.Module):
+    """Residual convolution module.
+    """
+
+    def __init__(self, features):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.conv1 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True
+        )
+
+        self.conv2 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True
+        )
+
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+        out = self.relu(x)
+        out = self.conv1(out)
+        out = self.relu(out)
+        out = self.conv2(out)
+
+        return out + x
+
+
+class FeatureFusionBlock(nn.Module):
+    """Feature fusion block.
+    """
+
+    def __init__(self, features):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock, self).__init__()
+
+        self.resConfUnit1 = ResidualConvUnit(features)
+        self.resConfUnit2 = ResidualConvUnit(features)
+
+    def forward(self, *xs):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if len(xs) == 2:
+            output += self.resConfUnit1(xs[1])
+
+        output = self.resConfUnit2(output)
+
+        output = nn.functional.interpolate(
+            output, scale_factor=2, mode="bilinear", align_corners=True
+        )
+
+        return output
+
+
+
+
+class ResidualConvUnit_custom(nn.Module):
+    """Residual convolution module.
+    """
+
+    def __init__(self, features, activation, bn):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super().__init__()
+
+        self.bn = bn
+
+        self.groups=1
+
+        self.conv1 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
+        )
+        
+        self.conv2 = nn.Conv2d(
+            features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups
+        )
+
+        if self.bn==True:
+            self.bn1 = nn.BatchNorm2d(features)
+            self.bn2 = nn.BatchNorm2d(features)
+
+        self.activation = activation
+
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input
+
+        Returns:
+            tensor: output
+        """
+        
+        out = self.activation(x)
+        out = self.conv1(out)
+        if self.bn==True:
+            out = self.bn1(out)
+       
+        out = self.activation(out)
+        out = self.conv2(out)
+        if self.bn==True:
+            out = self.bn2(out)
+
+        if self.groups > 1:
+            out = self.conv_merge(out)
+
+        return self.skip_add.add(out, x)
+
+        # return out + x
+
+
+class FeatureFusionBlock_custom(nn.Module):
+    """Feature fusion block.
+    """
+
+    def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True):
+        """Init.
+
+        Args:
+            features (int): number of features
+        """
+        super(FeatureFusionBlock_custom, self).__init__()
+
+        self.deconv = deconv
+        self.align_corners = align_corners
+
+        self.groups=1
+
+        self.expand = expand
+        out_features = features
+        if self.expand==True:
+            out_features = features//2
+        
+        self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1)
+
+        self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn)
+        self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn)
+        
+        self.skip_add = nn.quantized.FloatFunctional()
+
+    def forward(self, *xs):
+        """Forward pass.
+
+        Returns:
+            tensor: output
+        """
+        output = xs[0]
+
+        if len(xs) == 2:
+            res = self.resConfUnit1(xs[1])
+            output = self.skip_add.add(output, res)
+            # output += res
+
+        output = self.resConfUnit2(output)
+
+        output = nn.functional.interpolate(
+            output, scale_factor=2, mode="bilinear", align_corners=self.align_corners
+        )
+
+        output = self.out_conv(output)
+
+        return output
+

extralibs/midas/midas/dpt_depth.py

@@ -0,0 +1,110 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .base_model import BaseModel
+from .blocks import (
+    FeatureFusionBlock,
+    FeatureFusionBlock_custom,
+    Interpolate,
+    _make_encoder,
+    forward_vit,
+)
+
+
+def _make_fusion_block(features, use_bn):
+    return FeatureFusionBlock_custom(
+        features,
+        nn.ReLU(False),
+        deconv=False,
+        bn=use_bn,
+        expand=False,
+        align_corners=True,
+    )
+
+
+class DPT(BaseModel):
+    def __init__(
+        self,
+        head,
+        features=256,
+        backbone="vitb_rn50_384",
+        readout="project",
+        channels_last=False,
+        use_bn=False,
+    ):
+
+        super(DPT, self).__init__()
+
+        self.channels_last = channels_last
+
+        hooks = {
+            "vitb_rn50_384": [0, 1, 8, 11],
+            "vitb16_384": [2, 5, 8, 11],
+            "vitl16_384": [5, 11, 17, 23],
+        }
+
+        # Instantiate backbone and reassemble blocks
+        self.pretrained, self.scratch = _make_encoder(
+            backbone,
+            features,
+            False, # Set to true of you want to train from scratch, uses ImageNet weights
+            groups=1,
+            expand=False,
+            exportable=False,
+            hooks=hooks[backbone],
+            use_readout=readout,
+        )
+
+        self.scratch.refinenet1 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet2 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet3 = _make_fusion_block(features, use_bn)
+        self.scratch.refinenet4 = _make_fusion_block(features, use_bn)
+
+        self.scratch.output_conv = head
+
+
+    def forward(self, x):
+        if self.channels_last == True:
+            x.contiguous(memory_format=torch.channels_last)
+
+        layer_1, layer_2, layer_3, layer_4 = forward_vit(self.pretrained, x)
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+
+        out = self.scratch.output_conv(path_1)
+
+        return out
+
+
+class DPTDepthModel(DPT):
+    def __init__(self, path=None, non_negative=True, **kwargs):
+        features = kwargs["features"] if "features" in kwargs else 256
+
+        head = nn.Sequential(
+            nn.Conv2d(features, features // 2, kernel_size=3, stride=1, padding=1),
+            Interpolate(scale_factor=2, mode="bilinear", align_corners=True),
+            nn.Conv2d(features // 2, 32, kernel_size=3, stride=1, padding=1),
+            nn.ReLU(True),
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+            nn.Identity(),
+        )
+
+        super().__init__(head, **kwargs)
+
+        if path is not None:
+           self.load(path)
+           print("Midas depth estimation model loaded.")
+
+    def forward(self, x):
+        return super().forward(x).squeeze(dim=1)
+

extralibs/midas/midas/midas_net.py

@@ -0,0 +1,76 @@
+"""MidashNet: Network for monocular depth estimation trained by mixing several datasets.
+This file contains code that is adapted from
+https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py
+"""
+import torch
+import torch.nn as nn
+
+from .base_model import BaseModel
+from .blocks import FeatureFusionBlock, Interpolate, _make_encoder
+
+
+class MidasNet(BaseModel):
+    """Network for monocular depth estimation.
+    """
+
+    def __init__(self, path=None, features=256, non_negative=True):
+        """Init.
+
+        Args:
+            path (str, optional): Path to saved model. Defaults to None.
+            features (int, optional): Number of features. Defaults to 256.
+            backbone (str, optional): Backbone network for encoder. Defaults to resnet50
+        """
+        print("Loading weights: ", path)
+
+        super(MidasNet, self).__init__()
+
+        use_pretrained = False if path is None else True
+
+        self.pretrained, self.scratch = _make_encoder(backbone="resnext101_wsl", features=features, use_pretrained=use_pretrained)
+
+        self.scratch.refinenet4 = FeatureFusionBlock(features)
+        self.scratch.refinenet3 = FeatureFusionBlock(features)
+        self.scratch.refinenet2 = FeatureFusionBlock(features)
+        self.scratch.refinenet1 = FeatureFusionBlock(features)
+
+        self.scratch.output_conv = nn.Sequential(
+            nn.Conv2d(features, 128, kernel_size=3, stride=1, padding=1),
+            Interpolate(scale_factor=2, mode="bilinear"),
+            nn.Conv2d(128, 32, kernel_size=3, stride=1, padding=1),
+            nn.ReLU(True),
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+        )
+
+        if path:
+            self.load(path)
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input data (image)
+
+        Returns:
+            tensor: depth
+        """
+
+        layer_1 = self.pretrained.layer1(x)
+        layer_2 = self.pretrained.layer2(layer_1)
+        layer_3 = self.pretrained.layer3(layer_2)
+        layer_4 = self.pretrained.layer4(layer_3)
+
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+
+        out = self.scratch.output_conv(path_1)
+
+        return torch.squeeze(out, dim=1)

extralibs/midas/midas/midas_net_custom.py

@@ -0,0 +1,128 @@
+"""MidashNet: Network for monocular depth estimation trained by mixing several datasets.
+This file contains code that is adapted from
+https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py
+"""
+import torch
+import torch.nn as nn
+
+from .base_model import BaseModel
+from .blocks import FeatureFusionBlock, FeatureFusionBlock_custom, Interpolate, _make_encoder
+
+
+class MidasNet_small(BaseModel):
+    """Network for monocular depth estimation.
+    """
+
+    def __init__(self, path=None, features=64, backbone="efficientnet_lite3", non_negative=True, exportable=True, channels_last=False, align_corners=True,
+        blocks={'expand': True}):
+        """Init.
+
+        Args:
+            path (str, optional): Path to saved model. Defaults to None.
+            features (int, optional): Number of features. Defaults to 256.
+            backbone (str, optional): Backbone network for encoder. Defaults to resnet50
+        """
+        print("Loading weights: ", path)
+
+        super(MidasNet_small, self).__init__()
+
+        use_pretrained = False if path else True
+                
+        self.channels_last = channels_last
+        self.blocks = blocks
+        self.backbone = backbone
+
+        self.groups = 1
+
+        features1=features
+        features2=features
+        features3=features
+        features4=features
+        self.expand = False
+        if "expand" in self.blocks and self.blocks['expand'] == True:
+            self.expand = True
+            features1=features
+            features2=features*2
+            features3=features*4
+            features4=features*8
+
+        self.pretrained, self.scratch = _make_encoder(self.backbone, features, use_pretrained, groups=self.groups, expand=self.expand, exportable=exportable)
+  
+        self.scratch.activation = nn.ReLU(False)    
+
+        self.scratch.refinenet4 = FeatureFusionBlock_custom(features4, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet3 = FeatureFusionBlock_custom(features3, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet2 = FeatureFusionBlock_custom(features2, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners)
+        self.scratch.refinenet1 = FeatureFusionBlock_custom(features1, self.scratch.activation, deconv=False, bn=False, align_corners=align_corners)
+
+        
+        self.scratch.output_conv = nn.Sequential(
+            nn.Conv2d(features, features//2, kernel_size=3, stride=1, padding=1, groups=self.groups),
+            Interpolate(scale_factor=2, mode="bilinear"),
+            nn.Conv2d(features//2, 32, kernel_size=3, stride=1, padding=1),
+            self.scratch.activation,
+            nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0),
+            nn.ReLU(True) if non_negative else nn.Identity(),
+            nn.Identity(),
+        )
+        
+        if path:
+            self.load(path)
+
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x (tensor): input data (image)
+
+        Returns:
+            tensor: depth
+        """
+        if self.channels_last==True:
+            print("self.channels_last = ", self.channels_last)
+            x.contiguous(memory_format=torch.channels_last)
+
+
+        layer_1 = self.pretrained.layer1(x)
+        layer_2 = self.pretrained.layer2(layer_1)
+        layer_3 = self.pretrained.layer3(layer_2)
+        layer_4 = self.pretrained.layer4(layer_3)
+        
+        layer_1_rn = self.scratch.layer1_rn(layer_1)
+        layer_2_rn = self.scratch.layer2_rn(layer_2)
+        layer_3_rn = self.scratch.layer3_rn(layer_3)
+        layer_4_rn = self.scratch.layer4_rn(layer_4)
+
+
+        path_4 = self.scratch.refinenet4(layer_4_rn)
+        path_3 = self.scratch.refinenet3(path_4, layer_3_rn)
+        path_2 = self.scratch.refinenet2(path_3, layer_2_rn)
+        path_1 = self.scratch.refinenet1(path_2, layer_1_rn)
+        
+        out = self.scratch.output_conv(path_1)
+
+        return torch.squeeze(out, dim=1)
+
+
+
+def fuse_model(m):
+    prev_previous_type = nn.Identity()
+    prev_previous_name = ''
+    previous_type = nn.Identity()
+    previous_name = ''
+    for name, module in m.named_modules():
+        if prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d and type(module) == nn.ReLU:
+            # print("FUSED ", prev_previous_name, previous_name, name)
+            torch.quantization.fuse_modules(m, [prev_previous_name, previous_name, name], inplace=True)
+        elif prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d:
+            # print("FUSED ", prev_previous_name, previous_name)
+            torch.quantization.fuse_modules(m, [prev_previous_name, previous_name], inplace=True)
+        # elif previous_type == nn.Conv2d and type(module) == nn.ReLU:
+        #    print("FUSED ", previous_name, name)
+        #    torch.quantization.fuse_modules(m, [previous_name, name], inplace=True)
+
+        prev_previous_type = previous_type
+        prev_previous_name = previous_name
+        previous_type = type(module)
+        previous_name = name

extralibs/midas/midas/transforms.py

@@ -0,0 +1,234 @@
+import numpy as np
+import cv2
+import math
+
+
+def apply_min_size(sample, size, image_interpolation_method=cv2.INTER_AREA):
+    """Rezise the sample to ensure the given size. Keeps aspect ratio.
+
+    Args:
+        sample (dict): sample
+        size (tuple): image size
+
+    Returns:
+        tuple: new size
+    """
+    shape = list(sample["disparity"].shape)
+
+    if shape[0] >= size[0] and shape[1] >= size[1]:
+        return sample
+
+    scale = [0, 0]
+    scale[0] = size[0] / shape[0]
+    scale[1] = size[1] / shape[1]
+
+    scale = max(scale)
+
+    shape[0] = math.ceil(scale * shape[0])
+    shape[1] = math.ceil(scale * shape[1])
+
+    # resize
+    sample["image"] = cv2.resize(
+        sample["image"], tuple(shape[::-1]), interpolation=image_interpolation_method
+    )
+
+    sample["disparity"] = cv2.resize(
+        sample["disparity"], tuple(shape[::-1]), interpolation=cv2.INTER_NEAREST
+    )
+    sample["mask"] = cv2.resize(
+        sample["mask"].astype(np.float32),
+        tuple(shape[::-1]),
+        interpolation=cv2.INTER_NEAREST,
+    )
+    sample["mask"] = sample["mask"].astype(bool)
+
+    return tuple(shape)
+
+
+class Resize(object):
+    """Resize sample to given size (width, height).
+    """
+
+    def __init__(
+        self,
+        width,
+        height,
+        resize_target=True,
+        keep_aspect_ratio=False,
+        ensure_multiple_of=1,
+        resize_method="lower_bound",
+        image_interpolation_method=cv2.INTER_AREA,
+    ):
+        """Init.
+
+        Args:
+            width (int): desired output width
+            height (int): desired output height
+            resize_target (bool, optional):
+                True: Resize the full sample (image, mask, target).
+                False: Resize image only.
+                Defaults to True.
+            keep_aspect_ratio (bool, optional):
+                True: Keep the aspect ratio of the input sample.
+                Output sample might not have the given width and height, and
+                resize behaviour depends on the parameter 'resize_method'.
+                Defaults to False.
+            ensure_multiple_of (int, optional):
+                Output width and height is constrained to be multiple of this parameter.
+                Defaults to 1.
+            resize_method (str, optional):
+                "lower_bound": Output will be at least as large as the given size.
+                "upper_bound": Output will be at max as large as the given size. (Output size might be smaller than given size.)
+                "minimal": Scale as least as possible.  (Output size might be smaller than given size.)
+                Defaults to "lower_bound".
+        """
+        self.__width = width
+        self.__height = height
+
+        self.__resize_target = resize_target
+        self.__keep_aspect_ratio = keep_aspect_ratio
+        self.__multiple_of = ensure_multiple_of
+        self.__resize_method = resize_method
+        self.__image_interpolation_method = image_interpolation_method
+
+    def constrain_to_multiple_of(self, x, min_val=0, max_val=None):
+        y = (np.round(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        if max_val is not None and y > max_val:
+            y = (np.floor(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        if y < min_val:
+            y = (np.ceil(x / self.__multiple_of) * self.__multiple_of).astype(int)
+
+        return y
+
+    def get_size(self, width, height):
+        # determine new height and width
+        scale_height = self.__height / height
+        scale_width = self.__width / width
+
+        if self.__keep_aspect_ratio:
+            if self.__resize_method == "lower_bound":
+                # scale such that output size is lower bound
+                if scale_width > scale_height:
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            elif self.__resize_method == "upper_bound":
+                # scale such that output size is upper bound
+                if scale_width < scale_height:
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            elif self.__resize_method == "minimal":
+                # scale as least as possbile
+                if abs(1 - scale_width) < abs(1 - scale_height):
+                    # fit width
+                    scale_height = scale_width
+                else:
+                    # fit height
+                    scale_width = scale_height
+            else:
+                raise ValueError(
+                    f"resize_method {self.__resize_method} not implemented"
+                )
+
+        if self.__resize_method == "lower_bound":
+            new_height = self.constrain_to_multiple_of(
+                scale_height * height, min_val=self.__height
+            )
+            new_width = self.constrain_to_multiple_of(
+                scale_width * width, min_val=self.__width
+            )
+        elif self.__resize_method == "upper_bound":
+            new_height = self.constrain_to_multiple_of(
+                scale_height * height, max_val=self.__height
+            )
+            new_width = self.constrain_to_multiple_of(
+                scale_width * width, max_val=self.__width
+            )
+        elif self.__resize_method == "minimal":
+            new_height = self.constrain_to_multiple_of(scale_height * height)
+            new_width = self.constrain_to_multiple_of(scale_width * width)
+        else:
+            raise ValueError(f"resize_method {self.__resize_method} not implemented")
+
+        return (new_width, new_height)
+
+    def __call__(self, sample):
+        width, height = self.get_size(
+            sample["image"].shape[1], sample["image"].shape[0]
+        )
+
+        # resize sample
+        sample["image"] = cv2.resize(
+            sample["image"],
+            (width, height),
+            interpolation=self.__image_interpolation_method,
+        )
+
+        if self.__resize_target:
+            if "disparity" in sample:
+                sample["disparity"] = cv2.resize(
+                    sample["disparity"],
+                    (width, height),
+                    interpolation=cv2.INTER_NEAREST,
+                )
+
+            if "depth" in sample:
+                sample["depth"] = cv2.resize(
+                    sample["depth"], (width, height), interpolation=cv2.INTER_NEAREST
+                )
+
+            sample["mask"] = cv2.resize(
+                sample["mask"].astype(np.float32),
+                (width, height),
+                interpolation=cv2.INTER_NEAREST,
+            )
+            sample["mask"] = sample["mask"].astype(bool)
+
+        return sample
+
+
+class NormalizeImage(object):
+    """Normlize image by given mean and std.
+    """
+
+    def __init__(self, mean, std):
+        self.__mean = mean
+        self.__std = std
+
+    def __call__(self, sample):
+        sample["image"] = (sample["image"] - self.__mean) / self.__std
+
+        return sample
+
+
+class PrepareForNet(object):
+    """Prepare sample for usage as network input.
+    """
+
+    def __init__(self):
+        pass
+
+    def __call__(self, sample):
+        image = np.transpose(sample["image"], (2, 0, 1))
+        sample["image"] = np.ascontiguousarray(image).astype(np.float32)
+
+        if "mask" in sample:
+            sample["mask"] = sample["mask"].astype(np.float32)
+            sample["mask"] = np.ascontiguousarray(sample["mask"])
+
+        if "disparity" in sample:
+            disparity = sample["disparity"].astype(np.float32)
+            sample["disparity"] = np.ascontiguousarray(disparity)
+
+        if "depth" in sample:
+            depth = sample["depth"].astype(np.float32)
+            sample["depth"] = np.ascontiguousarray(depth)
+
+        return sample

extralibs/midas/midas/vit.py

@@ -0,0 +1,489 @@
+import torch
+import torch.nn as nn
+import timm
+import types
+import math
+import torch.nn.functional as F
+
+
+class Slice(nn.Module):
+    def __init__(self, start_index=1):
+        super(Slice, self).__init__()
+        self.start_index = start_index
+
+    def forward(self, x):
+        return x[:, self.start_index :]
+
+
+class AddReadout(nn.Module):
+    def __init__(self, start_index=1):
+        super(AddReadout, self).__init__()
+        self.start_index = start_index
+
+    def forward(self, x):
+        if self.start_index == 2:
+            readout = (x[:, 0] + x[:, 1]) / 2
+        else:
+            readout = x[:, 0]
+        return x[:, self.start_index :] + readout.unsqueeze(1)
+
+
+class ProjectReadout(nn.Module):
+    def __init__(self, in_features, start_index=1):
+        super(ProjectReadout, self).__init__()
+        self.start_index = start_index
+
+        self.project = nn.Sequential(nn.Linear(2 * in_features, in_features), nn.GELU())
+
+    def forward(self, x):
+        readout = x[:, 0].unsqueeze(1).expand_as(x[:, self.start_index :])
+        features = torch.cat((x[:, self.start_index :], readout), -1)
+
+        return self.project(features)
+
+
+class Transpose(nn.Module):
+    def __init__(self, dim0, dim1):
+        super(Transpose, self).__init__()
+        self.dim0 = dim0
+        self.dim1 = dim1
+
+    def forward(self, x):
+        x = x.transpose(self.dim0, self.dim1)
+        return x
+
+
+activations = {}
+def forward_vit(pretrained, x):
+    b, c, h, w = x.shape
+
+    glob = pretrained.model.forward_flex(x)
+    pretrained.activations = activations
+
+    layer_1 = pretrained.activations["1"]
+    layer_2 = pretrained.activations["2"]
+    layer_3 = pretrained.activations["3"]
+    layer_4 = pretrained.activations["4"]
+
+    layer_1 = pretrained.act_postprocess1[0:2](layer_1)
+    layer_2 = pretrained.act_postprocess2[0:2](layer_2)
+    layer_3 = pretrained.act_postprocess3[0:2](layer_3)
+    layer_4 = pretrained.act_postprocess4[0:2](layer_4)
+
+    unflatten = nn.Sequential(
+        nn.Unflatten(
+            2,
+            torch.Size(
+                [
+                    h // pretrained.model.patch_size[1],
+                    w // pretrained.model.patch_size[0],
+                ]
+            ),
+        )
+    )
+
+    if layer_1.ndim == 3:
+        layer_1 = unflatten(layer_1)
+    if layer_2.ndim == 3:
+        layer_2 = unflatten(layer_2)
+    if layer_3.ndim == 3:
+        layer_3 = unflatten(layer_3)
+    if layer_4.ndim == 3:
+        layer_4 = unflatten(layer_4)
+
+    layer_1 = pretrained.act_postprocess1[3 : len(pretrained.act_postprocess1)](layer_1)
+    layer_2 = pretrained.act_postprocess2[3 : len(pretrained.act_postprocess2)](layer_2)
+    layer_3 = pretrained.act_postprocess3[3 : len(pretrained.act_postprocess3)](layer_3)
+    layer_4 = pretrained.act_postprocess4[3 : len(pretrained.act_postprocess4)](layer_4)
+
+    return layer_1, layer_2, layer_3, layer_4
+
+
+def _resize_pos_embed(self, posemb, gs_h, gs_w):
+    posemb_tok, posemb_grid = (
+        posemb[:, : self.start_index],
+        posemb[0, self.start_index :],
+    )
+
+    gs_old = int(math.sqrt(len(posemb_grid)))
+
+    posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)
+    posemb_grid = F.interpolate(posemb_grid, size=(gs_h, gs_w), mode="bilinear")
+    posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_h * gs_w, -1)
+
+    posemb = torch.cat([posemb_tok, posemb_grid], dim=1)
+
+    return posemb
+
+
+def forward_flex(self, x):
+    b, c, h, w = x.shape
+
+    pos_embed = self._resize_pos_embed(
+        self.pos_embed, h // self.patch_size[1], w // self.patch_size[0]
+    )
+
+    B = x.shape[0]
+
+    if hasattr(self.patch_embed, "backbone"):
+        x = self.patch_embed.backbone(x)
+        if isinstance(x, (list, tuple)):
+            x = x[-1]  # last feature if backbone outputs list/tuple of features
+
+    x = self.patch_embed.proj(x).flatten(2).transpose(1, 2)
+
+    if getattr(self, "dist_token", None) is not None:
+        cls_tokens = self.cls_token.expand(
+            B, -1, -1
+        )  # stole cls_tokens impl from Phil Wang, thanks
+        dist_token = self.dist_token.expand(B, -1, -1)
+        x = torch.cat((cls_tokens, dist_token, x), dim=1)
+    else:
+        cls_tokens = self.cls_token.expand(
+            B, -1, -1
+        )  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+
+    x = x + pos_embed
+    x = self.pos_drop(x)
+
+    for blk in self.blocks:
+        x = blk(x)
+
+    x = self.norm(x)
+
+    return x
+
+
+def get_activation(name):
+    def hook(model, input, output):
+        activations[name] = output
+    return hook
+
+
+def get_readout_oper(vit_features, features, use_readout, start_index=1):
+    if use_readout == "ignore":
+        readout_oper = [Slice(start_index)] * len(features)
+    elif use_readout == "add":
+        readout_oper = [AddReadout(start_index)] * len(features)
+    elif use_readout == "project":
+        readout_oper = [
+            ProjectReadout(vit_features, start_index) for out_feat in features
+        ]
+    else:
+        assert (
+            False
+        ), "wrong operation for readout token, use_readout can be 'ignore', 'add', or 'project'"
+
+    return readout_oper
+
+
+def _make_vit_b16_backbone(
+    model,
+    features=[96, 192, 384, 768],
+    size=[384, 384],
+    hooks=[2, 5, 8, 11],
+    vit_features=768,
+    use_readout="ignore",
+    start_index=1,
+):
+    pretrained = nn.Module()
+
+    pretrained.model = model
+    pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
+    pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
+    pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
+    pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
+
+    pretrained.activations = activations
+
+    readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
+
+    # 32, 48, 136, 384
+    pretrained.act_postprocess1 = nn.Sequential(
+        readout_oper[0],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[0],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.ConvTranspose2d(
+            in_channels=features[0],
+            out_channels=features[0],
+            kernel_size=4,
+            stride=4,
+            padding=0,
+            bias=True,
+            dilation=1,
+            groups=1,
+        ),
+    )
+
+    pretrained.act_postprocess2 = nn.Sequential(
+        readout_oper[1],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[1],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.ConvTranspose2d(
+            in_channels=features[1],
+            out_channels=features[1],
+            kernel_size=2,
+            stride=2,
+            padding=0,
+            bias=True,
+            dilation=1,
+            groups=1,
+        ),
+    )
+
+    pretrained.act_postprocess3 = nn.Sequential(
+        readout_oper[2],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[2],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+    )
+
+    pretrained.act_postprocess4 = nn.Sequential(
+        readout_oper[3],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[3],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.Conv2d(
+            in_channels=features[3],
+            out_channels=features[3],
+            kernel_size=3,
+            stride=2,
+            padding=1,
+        ),
+    )
+
+    pretrained.model.start_index = start_index
+    pretrained.model.patch_size = [16, 16]
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
+    pretrained.model._resize_pos_embed = types.MethodType(
+        _resize_pos_embed, pretrained.model
+    )
+
+    return pretrained
+
+
+def _make_pretrained_vitl16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_large_patch16_384", pretrained=pretrained)
+
+    hooks = [5, 11, 17, 23] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model,
+        features=[256, 512, 1024, 1024],
+        hooks=hooks,
+        vit_features=1024,
+        use_readout=use_readout,
+    )
+
+
+def _make_pretrained_vitb16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_base_patch16_384", pretrained=pretrained)
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout
+    )
+
+
+def _make_pretrained_deitb16_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model("vit_deit_base_patch16_384", pretrained=pretrained)
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout
+    )
+
+
+def _make_pretrained_deitb16_distil_384(pretrained, use_readout="ignore", hooks=None):
+    model = timm.create_model(
+        "vit_deit_base_distilled_patch16_384", pretrained=pretrained
+    )
+
+    hooks = [2, 5, 8, 11] if hooks == None else hooks
+    return _make_vit_b16_backbone(
+        model,
+        features=[96, 192, 384, 768],
+        hooks=hooks,
+        use_readout=use_readout,
+        start_index=2,
+    )
+
+
+def _make_vit_b_rn50_backbone(
+    model,
+    features=[256, 512, 768, 768],
+    size=[384, 384],
+    hooks=[0, 1, 8, 11],
+    vit_features=768,
+    use_vit_only=False,
+    use_readout="ignore",
+    start_index=1,
+):
+    pretrained = nn.Module()
+
+    pretrained.model = model
+    
+    if use_vit_only == True:
+        pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1"))
+        pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2"))
+    else:
+        pretrained.model.patch_embed.backbone.stages[0].register_forward_hook(
+            get_activation("1")
+        )
+        pretrained.model.patch_embed.backbone.stages[1].register_forward_hook(
+            get_activation("2")
+        )
+
+    pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3"))
+    pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4"))
+
+    pretrained.activations = activations
+
+    readout_oper = get_readout_oper(vit_features, features, use_readout, start_index)
+
+    if use_vit_only == True:
+        pretrained.act_postprocess1 = nn.Sequential(
+            readout_oper[0],
+            Transpose(1, 2),
+            nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+            nn.Conv2d(
+                in_channels=vit_features,
+                out_channels=features[0],
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            ),
+            nn.ConvTranspose2d(
+                in_channels=features[0],
+                out_channels=features[0],
+                kernel_size=4,
+                stride=4,
+                padding=0,
+                bias=True,
+                dilation=1,
+                groups=1,
+            ),
+        )
+
+        pretrained.act_postprocess2 = nn.Sequential(
+            readout_oper[1],
+            Transpose(1, 2),
+            nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+            nn.Conv2d(
+                in_channels=vit_features,
+                out_channels=features[1],
+                kernel_size=1,
+                stride=1,
+                padding=0,
+            ),
+            nn.ConvTranspose2d(
+                in_channels=features[1],
+                out_channels=features[1],
+                kernel_size=2,
+                stride=2,
+                padding=0,
+                bias=True,
+                dilation=1,
+                groups=1,
+            ),
+        )
+    else:
+        pretrained.act_postprocess1 = nn.Sequential(
+            nn.Identity(), nn.Identity(), nn.Identity()
+        )
+        pretrained.act_postprocess2 = nn.Sequential(
+            nn.Identity(), nn.Identity(), nn.Identity()
+        )
+
+    pretrained.act_postprocess3 = nn.Sequential(
+        readout_oper[2],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[2],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+    )
+
+    pretrained.act_postprocess4 = nn.Sequential(
+        readout_oper[3],
+        Transpose(1, 2),
+        nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])),
+        nn.Conv2d(
+            in_channels=vit_features,
+            out_channels=features[3],
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        ),
+        nn.Conv2d(
+            in_channels=features[3],
+            out_channels=features[3],
+            kernel_size=3,
+            stride=2,
+            padding=1,
+        ),
+    )
+
+    pretrained.model.start_index = start_index
+    pretrained.model.patch_size = [16, 16]
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model)
+
+    # We inject this function into the VisionTransformer instances so that
+    # we can use it with interpolated position embeddings without modifying the library source.
+    pretrained.model._resize_pos_embed = types.MethodType(
+        _resize_pos_embed, pretrained.model
+    )
+
+    return pretrained
+
+
+def _make_pretrained_vitb_rn50_384(
+    pretrained, use_readout="ignore", hooks=None, use_vit_only=False
+):
+    model = timm.create_model("vit_base_resnet50_384", pretrained=pretrained)
+
+    hooks = [0, 1, 8, 11] if hooks == None else hooks
+    return _make_vit_b_rn50_backbone(
+        model,
+        features=[256, 512, 768, 768],
+        size=[384, 384],
+        hooks=hooks,
+        use_vit_only=use_vit_only,
+        use_readout=use_readout,
+    )

extralibs/midas/utils.py

@@ -0,0 +1,189 @@
+"""Utils for monoDepth."""
+import sys
+import re
+import numpy as np
+import cv2
+import torch
+
+
+def read_pfm(path):
+    """Read pfm file.
+
+    Args:
+        path (str): path to file
+
+    Returns:
+        tuple: (data, scale)
+    """
+    with open(path, "rb") as file:
+
+        color = None
+        width = None
+        height = None
+        scale = None
+        endian = None
+
+        header = file.readline().rstrip()
+        if header.decode("ascii") == "PF":
+            color = True
+        elif header.decode("ascii") == "Pf":
+            color = False
+        else:
+            raise Exception("Not a PFM file: " + path)
+
+        dim_match = re.match(r"^(\d+)\s(\d+)\s$", file.readline().decode("ascii"))
+        if dim_match:
+            width, height = list(map(int, dim_match.groups()))
+        else:
+            raise Exception("Malformed PFM header.")
+
+        scale = float(file.readline().decode("ascii").rstrip())
+        if scale < 0:
+            # little-endian
+            endian = "<"
+            scale = -scale
+        else:
+            # big-endian
+            endian = ">"
+
+        data = np.fromfile(file, endian + "f")
+        shape = (height, width, 3) if color else (height, width)
+
+        data = np.reshape(data, shape)
+        data = np.flipud(data)
+
+        return data, scale
+
+
+def write_pfm(path, image, scale=1):
+    """Write pfm file.
+
+    Args:
+        path (str): pathto file
+        image (array): data
+        scale (int, optional): Scale. Defaults to 1.
+    """
+
+    with open(path, "wb") as file:
+        color = None
+
+        if image.dtype.name != "float32":
+            raise Exception("Image dtype must be float32.")
+
+        image = np.flipud(image)
+
+        if len(image.shape) == 3 and image.shape[2] == 3:  # color image
+            color = True
+        elif (
+            len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1
+        ):  # greyscale
+            color = False
+        else:
+            raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.")
+
+        file.write("PF\n" if color else "Pf\n".encode())
+        file.write("%d %d\n".encode() % (image.shape[1], image.shape[0]))
+
+        endian = image.dtype.byteorder
+
+        if endian == "<" or endian == "=" and sys.byteorder == "little":
+            scale = -scale
+
+        file.write("%f\n".encode() % scale)
+
+        image.tofile(file)
+
+
+def read_image(path):
+    """Read image and output RGB image (0-1).
+
+    Args:
+        path (str): path to file
+
+    Returns:
+        array: RGB image (0-1)
+    """
+    img = cv2.imread(path)
+
+    if img.ndim == 2:
+        img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255.0
+
+    return img
+
+
+def resize_image(img):
+    """Resize image and make it fit for network.
+
+    Args:
+        img (array): image
+
+    Returns:
+        tensor: data ready for network
+    """
+    height_orig = img.shape[0]
+    width_orig = img.shape[1]
+
+    if width_orig > height_orig:
+        scale = width_orig / 384
+    else:
+        scale = height_orig / 384
+
+    height = (np.ceil(height_orig / scale / 32) * 32).astype(int)
+    width = (np.ceil(width_orig / scale / 32) * 32).astype(int)
+
+    img_resized = cv2.resize(img, (width, height), interpolation=cv2.INTER_AREA)
+
+    img_resized = (
+        torch.from_numpy(np.transpose(img_resized, (2, 0, 1))).contiguous().float()
+    )
+    img_resized = img_resized.unsqueeze(0)
+
+    return img_resized
+
+
+def resize_depth(depth, width, height):
+    """Resize depth map and bring to CPU (numpy).
+
+    Args:
+        depth (tensor): depth
+        width (int): image width
+        height (int): image height
+
+    Returns:
+        array: processed depth
+    """
+    depth = torch.squeeze(depth[0, :, :, :]).to("cpu")
+
+    depth_resized = cv2.resize(
+        depth.numpy(), (width, height), interpolation=cv2.INTER_CUBIC
+    )
+
+    return depth_resized
+
+def write_depth(path, depth, bits=1):
+    """Write depth map to pfm and png file.
+
+    Args:
+        path (str): filepath without extension
+        depth (array): depth
+    """
+    write_pfm(path + ".pfm", depth.astype(np.float32))
+
+    depth_min = depth.min()
+    depth_max = depth.max()
+
+    max_val = (2**(8*bits))-1
+
+    if depth_max - depth_min > np.finfo("float").eps:
+        out = max_val * (depth - depth_min) / (depth_max - depth_min)
+    else:
+        out = np.zeros(depth.shape, dtype=depth.type)
+
+    if bits == 1:
+        cv2.imwrite(path + ".png", out.astype("uint8"))
+    elif bits == 2:
+        cv2.imwrite(path + ".png", out.astype("uint16"))
+
+    return

lvdm/models/ddpm3d.py

@@ -1432,3 +1432,53 @@ class DiffusionWrapper(pl.LightningModule):
 
         return out
 
+
+class T2VAdapterDepth(LatentDiffusion):
+    def __init__(self, depth_stage_config, adapter_config, *args, **kwargs):
+        super(T2VAdapterDepth, self).__init__(*args, **kwargs)
+        self.adapter = instantiate_from_config(adapter_config)
+        self.condtype = adapter_config.cond_name
+        self.depth_stage_model = instantiate_from_config(depth_stage_config)
+
+    def prepare_midas_input(self, batch_x):
+        # input: b,c,h,w
+        x_midas = torch.nn.functional.interpolate(batch_x, size=(384, 384), mode='bicubic')
+        return x_midas
+    
+    @torch.no_grad()
+    def get_batch_depth(self, batch_x, target_size, encode_bs=1):
+        b, c, t, h, w = batch_x.shape
+        merge_x = rearrange(batch_x, 'b c t h w -> (b t) c h w')
+        split_x = torch.split(merge_x, encode_bs, dim=0)
+        cond_depth_list = []
+        for x in split_x:
+            x_midas = self.prepare_midas_input(x)
+            cond_depth = self.depth_stage_model(x_midas)
+            cond_depth = torch.nn.functional.interpolate(
+                    cond_depth,
+                    size=target_size,
+                    mode="bicubic",
+                    align_corners=False,
+                )
+            depth_min, depth_max = torch.amin(cond_depth, dim=[1, 2, 3], keepdim=True), torch.amax(cond_depth, dim=[1, 2, 3], keepdim=True)
+            cond_depth = 2. * (cond_depth - depth_min) / (depth_max - depth_min + 1e-7) - 1.
+            cond_depth_list.append(cond_depth)
+        batch_cond_depth=torch.cat(cond_depth_list, dim=0)
+        batch_cond_depth = rearrange(batch_cond_depth, '(b t) c h w -> b c t h w', b=b, t=t)
+        return batch_cond_depth
+    
+    def get_adapter_features(self, extra_cond, encode_bs=1):
+        b, c, t, h, w = extra_cond.shape
+        ## process in 2D manner
+        merge_extra_cond = rearrange(extra_cond, 'b c t h w -> (b t) c h w')
+        split_extra_cond = torch.split(merge_extra_cond, encode_bs, dim=0)
+        features_adapter_list = []
+        for extra_cond in split_extra_cond:
+            features_adapter = self.adapter(extra_cond)
+            features_adapter_list.append(features_adapter)
+        merge_features_adapter_list = []
+        for i in range(len(features_adapter_list[0])):
+            merge_features_adapter = torch.cat([features_adapter_list[num][i] for num in range(len(features_adapter_list))], dim=0)
+            merge_features_adapter_list.append(merge_features_adapter)
+        merge_features_adapter_list = [rearrange(feature, '(b t) c h w -> b c t h w', b=b, t=t) for feature in merge_features_adapter_list]
+        return merge_features_adapter_list

lvdm/models/modules/adapter.py

@@ -0,0 +1,105 @@
+import torch
+import torch.nn as nn
+from collections import OrderedDict
+from lvdm.models.modules.util import (
+    zero_module,
+    conv_nd,
+    avg_pool_nd
+)
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, in_c, out_c, down, ksize=3, sk=False, use_conv=True):
+        super().__init__()
+        ps = ksize // 2
+        if in_c != out_c or sk == False:
+            self.in_conv = nn.Conv2d(in_c, out_c, ksize, 1, ps)
+        else:
+            # print('n_in')
+            self.in_conv = None
+        self.block1 = nn.Conv2d(out_c, out_c, 3, 1, 1)
+        self.act = nn.ReLU()
+        self.block2 = nn.Conv2d(out_c, out_c, ksize, 1, ps)
+        if sk == False:
+            self.skep = nn.Conv2d(in_c, out_c, ksize, 1, ps)
+        else:
+            self.skep = None
+
+        self.down = down
+        if self.down == True:
+            self.down_opt = Downsample(in_c, use_conv=use_conv)
+
+    def forward(self, x):
+        if self.down == True:
+            x = self.down_opt(x)
+        if self.in_conv is not None:  # edit
+            x = self.in_conv(x)
+
+        h = self.block1(x)
+        h = self.act(h)
+        h = self.block2(h)
+        if self.skep is not None:
+            return h + self.skep(x)
+        else:
+            return h + x
+
+
+class Adapter(nn.Module):
+    def __init__(self, channels=[320, 640, 1280, 1280], nums_rb=3, cin=64, ksize=3, sk=False, use_conv=True):
+        super(Adapter, self).__init__()
+        self.unshuffle = nn.PixelUnshuffle(8)
+        self.channels = channels
+        self.nums_rb = nums_rb
+        self.body = []
+        for i in range(len(channels)):
+            for j in range(nums_rb):
+                if (i != 0) and (j == 0):
+                    self.body.append(
+                        ResnetBlock(channels[i - 1], channels[i], down=True, ksize=ksize, sk=sk, use_conv=use_conv))
+                else:
+                    self.body.append(
+                        ResnetBlock(channels[i], channels[i], down=False, ksize=ksize, sk=sk, use_conv=use_conv))
+        self.body = nn.ModuleList(self.body)
+        self.conv_in = nn.Conv2d(cin, channels[0], 3, 1, 1)
+
+    def forward(self, x):
+        # unshuffle
+        x = self.unshuffle(x)
+        # extract features
+        features = []
+        x = self.conv_in(x)
+        for i in range(len(self.channels)):
+            for j in range(self.nums_rb):
+                idx = i * self.nums_rb + j
+                x = self.body[idx](x)
+            features.append(x)
+
+        return features

lvdm/models/modules/lora.py

@@ -622,7 +622,7 @@ def net_load_lora(net, checkpoint_path, alpha=1.0, remove=False):
     state_dict = torch.load(checkpoint_path)
     for k, v in state_dict.items():
         state_dict[k] = v.to(net.device)
-    # import pdb;pdb.set_trace()
+
     for key in state_dict:
         if ".alpha" in key or key in visited:
             continue
@@ -685,7 +685,7 @@ def net_load_lora_v2(net, checkpoint_path, alpha=1.0, remove=False, origin_weigh
     state_dict = torch.load(checkpoint_path)
     for k, v in state_dict.items():
         state_dict[k] = v.to(net.device)
-    # import pdb;pdb.set_trace()
+
     for key in state_dict:
         if ".alpha" in key or key in visited:
             continue

lvdm/models/modules/openaimodel3d.py

@@ -629,7 +629,7 @@ class UNetModel(nn.Module):
         self.middle_block.apply(convert_module_to_f32)
         self.output_blocks.apply(convert_module_to_f32)
 
-    def forward(self, x, timesteps=None, time_emb_replace=None, context=None, y=None, **kwargs):
+    def forward(self, x, timesteps=None, time_emb_replace=None, context=None, features_adapter=None, y=None, **kwargs):
         """
         Apply the model to an input batch.
         :param x: an [N x C x ...] Tensor of inputs.
@@ -651,9 +651,17 @@ class UNetModel(nn.Module):
             emb = emb + self.label_emb(y)
 
         h = x.type(self.dtype)
-        for module in self.input_blocks:
+        adapter_idx = 0
+        for id, module in enumerate(self.input_blocks):
             h = module(h, emb, context, **kwargs)
+            ## plug-in adapter features
+            if ((id+1)%3 == 0) and features_adapter is not None:
+                h = h + features_adapter[adapter_idx]
+                adapter_idx += 1
             hs.append(h)
+        if features_adapter is not None:
+            assert len(features_adapter)==adapter_idx, 'Mismatch features adapter'
+
         h = self.middle_block(h, emb, context, **kwargs)
         for module in self.output_blocks:
             h = th.cat([h, hs.pop()], dim=1)

lvdm/samplers/ddim.py

@@ -197,7 +197,7 @@ class DDIMSampler(object):
     def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
                       temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
                       unconditional_guidance_scale=1., unconditional_conditioning=None, sample_noise=None,
-                      cond_fn=None,uc_type=None, model_kwargs={}, 
+                      cond_fn=None, uc_type=None, 
                       **kwargs,
                       ):
         b, *_, device = *x.shape, x.device
@@ -206,15 +206,15 @@ class DDIMSampler(object):
         else:
             is_video = False
         if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
-            e_t = self.model.apply_model(x, t, c, **model_kwargs) # unet denoiser
+            e_t = self.model.apply_model(x, t, c, **kwargs) # unet denoiser
         else:
             # with unconditional condition
             if isinstance(c, torch.Tensor):
-                e_t = self.model.apply_model(x, t, c, **model_kwargs)
-                e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **model_kwargs)
+                e_t = self.model.apply_model(x, t, c, **kwargs)
+                e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
             elif isinstance(c, dict):
-                e_t = self.model.apply_model(x, t, c, **model_kwargs)
-                e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **model_kwargs)
+                e_t = self.model.apply_model(x, t, c, **kwargs)
+                e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
             else:
                 raise NotImplementedError
             # text cfg

lvdm/utils/saving_utils.py

@@ -14,6 +14,24 @@ from torchvision.utils import make_grid
 from torch import Tensor
 from torchvision.transforms.functional import to_tensor
 
+
+def tensor_to_mp4(video, savepath, fps, rescale=True, nrow=None):
+    """
+    video: torch.Tensor, b,c,t,h,w, 0-1
+    if -1~1, enable rescale=True
+    """
+    n = video.shape[0]
+    video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w
+    nrow = int(np.sqrt(n)) if nrow is None else nrow
+    frame_grids = [torchvision.utils.make_grid(framesheet, nrow=nrow) for framesheet in video] # [3, grid_h, grid_w]
+    grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [T, 3, grid_h, grid_w]
+    grid = torch.clamp(grid.float(), -1., 1.)
+    if rescale:
+        grid = (grid + 1.0) / 2.0
+    grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1) # [T, 3, grid_h, grid_w] -> [T, grid_h, grid_w, 3]
+    #print(f'Save video to {savepath}')
+    torchvision.io.write_video(savepath, grid, fps=fps, video_codec='h264', options={'crf': '10'})
+
 # ----------------------------------------------------------------------------------------------
 def savenp2sheet(imgs, savepath, nrow=None):
     """ save multiple imgs (in numpy array type) to a img sheet.

models/adapter_t2v_depth/model_config.yaml

@@ -0,0 +1,89 @@
+model:
+  target: lvdm.models.ddpm3d.T2VAdapterDepth
+  params:
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: video
+    cond_stage_key: caption
+    image_size:
+    - 32
+    - 32
+    video_length: 16
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    scale_by_std: false
+    scale_factor: 0.18215
+
+    unet_config:
+      target: lvdm.models.modules.openaimodel3d.UNetModel
+      params:
+        image_size: 32
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        num_heads: 8
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: true
+        legacy: false
+        kernel_size_t: 1
+        padding_t: 0
+        temporal_length: 16
+        use_relative_position: true
+        
+    first_stage_config:
+      target: lvdm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: lvdm.models.modules.condition_modules.FrozenCLIPEmbedder
+      
+    depth_stage_config:
+      target: extralibs.midas.api.MiDaSInference
+      params:
+        model_type: "dpt_hybrid"
+        model_path: models/adapter_t2v_depth/dpt_hybrid-midas.pt
+
+    adapter_config:
+        target: lvdm.models.modules.adapter.Adapter
+        cond_name: depth
+        params:
+          cin: 64
+          channels: [320, 640, 1280, 1280]
+          nums_rb: 2
+          ksize: 1
+          sk: True
+          use_conv: False

sample_adapter.sh

@@ -0,0 +1,22 @@
+PROMPT="An ostrich walking in the desert, photorealistic, 4k"
+VIDEO="input/flamingo.mp4"
+OUTDIR="results/"
+
+NAME="video_adapter"
+CONFIG_PATH="models/adapter_t2v_depth/model_config.yaml"
+BASE_PATH="models/base_t2v/model.ckpt"
+ADAPTER_PATH="models/adapter_t2v_depth/adapter.pth"
+
+python scripts/sample_text2video_adapter.py \
+    --seed 123 \
+    --ckpt_path $BASE_PATH \
+    --adapter_ckpt $ADAPTER_PATH \
+    --base $CONFIG_PATH \
+    --savedir $OUTDIR/$NAME \
+    --bs 1 --height 256 --width 256 \
+    --frame_stride -1 \
+    --unconditional_guidance_scale 15.0 \
+    --ddim_steps 50 \
+    --ddim_eta 1.0 \
+    --prompt "$PROMPT" \
+    --video $VIDEO

scripts/ddp_wrapper.py

@@ -0,0 +1,47 @@
+import datetime
+import argparse, importlib
+from pytorch_lightning import seed_everything
+
+import torch
+import torch.distributed as dist
+
+
+def setup_dist(local_rank):
+    if dist.is_initialized():
+        return
+    torch.cuda.set_device(local_rank)
+    torch.distributed.init_process_group('nccl', init_method='env://')
+
+
+def get_dist_info():
+    if dist.is_available():
+        initialized = dist.is_initialized()
+    else:
+        initialized = False
+    if initialized:
+        rank = dist.get_rank()
+        world_size = dist.get_world_size()
+    else:
+        rank = 0
+        world_size = 1
+    return rank, world_size
+
+
+if __name__ == '__main__':
+    now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--module", type=str, help="module name", default="inference")
+    parser.add_argument("--local_rank", type=int, nargs="?", help="for ddp", default=0)
+    args, unknown = parser.parse_known_args()
+    inference_api = importlib.import_module(args.module, package=None)
+
+    inference_parser = inference_api.get_parser()
+    inference_args, unknown = inference_parser.parse_known_args()
+
+    seed_everything(inference_args.seed)
+    setup_dist(args.local_rank)
+    torch.backends.cudnn.benchmark = True
+    rank, gpu_num = get_dist_info()
+
+    print("@CoVideoGen Inference [rank%d]: %s"%(rank, now))
+    inference_api.run_inference(inference_args, rank)

scripts/sample_text2video_adapter.py

@@ -0,0 +1,206 @@
+import argparse, os, sys, glob
+import datetime, time
+from omegaconf import OmegaConf
+
+import torch
+from decord import VideoReader, cpu
+import torchvision
+from pytorch_lightning import seed_everything
+
+from lvdm.samplers.ddim import DDIMSampler
+from lvdm.utils.common_utils import instantiate_from_config
+from lvdm.utils.saving_utils import tensor_to_mp4
+
+
+def get_filelist(data_dir, ext='*'):
+    file_list = glob.glob(os.path.join(data_dir, '*.%s'%ext))
+    file_list.sort()
+    return file_list
+
+def load_model_checkpoint(model, ckpt, adapter_ckpt=None):
+    print('>>> Loading checkpoints ...')
+    if adapter_ckpt:
+        ## main model
+        state_dict = torch.load(ckpt, map_location="cpu")
+        if "state_dict" in list(state_dict.keys()):
+            state_dict = state_dict["state_dict"]
+        model.load_state_dict(state_dict, strict=False)
+        print('@model checkpoint loaded.')
+        ## adapter
+        state_dict = torch.load(adapter_ckpt, map_location="cpu")
+        if "state_dict" in list(state_dict.keys()):
+            state_dict = state_dict["state_dict"]
+        model.adapter.load_state_dict(state_dict, strict=True)
+        print('@adapter checkpoint loaded.')
+    else:
+        state_dict = torch.load(ckpt, map_location="cpu")
+        if "state_dict" in list(state_dict.keys()):
+            state_dict = state_dict["state_dict"]
+        model.load_state_dict(state_dict, strict=True)
+        print('@model checkpoint loaded.')
+    return model
+
+def load_prompts(prompt_file):
+    f = open(prompt_file, 'r')
+    prompt_list = []
+    for idx, line in enumerate(f.readlines()):
+        l = line.strip()
+        if len(l) != 0:
+            prompt_list.append(l)
+        f.close()
+    return prompt_list
+
+def load_video(filepath, frame_stride, video_size=(256,256), video_frames=16):
+    vidreader = VideoReader(filepath, ctx=cpu(0), width=video_size[1], height=video_size[0])
+    max_frames = len(vidreader)
+    temp_stride = max_frames // video_frames if frame_stride == -1 else frame_stride
+    if temp_stride * (video_frames-1) >= max_frames:
+        print(f'Warning: default frame stride is used because the input video clip {max_frames} is not long enough.')
+        temp_stride = max_frames // video_frames
+    frame_indices = [temp_stride*i for i in range(video_frames)]
+    frames = vidreader.get_batch(frame_indices)
+        
+    ## [t,h,w,c] -> [c,t,h,w]
+    frame_tensor = torch.tensor(frames.asnumpy()).permute(3, 0, 1, 2).float()
+    frame_tensor = (frame_tensor / 255. - 0.5) * 2    
+    return frame_tensor
+
+
+def save_results(prompt, samples, inputs, filename, realdir, fakedir, fps=10):
+    ## save prompt
+    prompt = prompt[0] if isinstance(prompt, list) else prompt
+    path = os.path.join(realdir, "%s.txt"%filename)
+    with open(path, 'w') as f:
+        f.write(f'{prompt}')
+        f.close()
+
+    ## save video
+    videos = [inputs, samples]
+    savedirs = [realdir, fakedir]
+    for idx, video in enumerate(videos):
+        if video is None:
+            continue
+        # b,c,t,h,w
+        video = video.detach().cpu()
+        video = torch.clamp(video.float(), -1., 1.)
+        n = video.shape[0]
+        video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w
+        frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n)) for framesheet in video] #[3, 1*h, n*w]
+        grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w]
+        grid = (grid + 1.0) / 2.0
+        grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
+        path = os.path.join(savedirs[idx], "%s.mp4"%filename)
+        torchvision.io.write_video(path, grid, fps=fps, video_codec='h264', options={'crf': '10'})
+
+
+def adapter_guided_synthesis(model, prompts, videos, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1., \
+                        unconditional_guidance_scale=1.0, unconditional_guidance_scale_temporal=None, **kwargs):
+    ddim_sampler = DDIMSampler(model)
+
+    batch_size = noise_shape[0]
+    ## get condition embeddings (support single prompt only)
+    if isinstance(prompts, str):
+        prompts = [prompts]
+    cond = model.get_learned_conditioning(prompts)
+    if unconditional_guidance_scale != 1.0:
+        prompts = batch_size * [""]
+        uc = model.get_learned_conditioning(prompts)
+    else:
+        uc = None
+    
+    ## adapter features: process in 2D manner
+    b, c, t, h, w = videos.shape
+    extra_cond = model.get_batch_depth(videos, (h,w))
+    features_adapter = model.get_adapter_features(extra_cond)
+
+    batch_variants = []
+    for _ in range(n_samples):
+        if ddim_sampler is not None:
+            samples, _ = ddim_sampler.sample(S=ddim_steps,
+                                            conditioning=cond,
+                                            batch_size=noise_shape[0],
+                                            shape=noise_shape[1:],
+                                            verbose=False,
+                                            unconditional_guidance_scale=unconditional_guidance_scale,
+                                            unconditional_conditioning=uc,
+                                            eta=ddim_eta,
+                                            temporal_length=noise_shape[2],
+                                            conditional_guidance_scale_temporal=unconditional_guidance_scale_temporal,
+                                            features_adapter=features_adapter,
+                                            **kwargs
+                                            )        
+        ## reconstruct from latent to pixel space
+        batch_images = model.decode_first_stage(samples, decode_bs=1, return_cpu=False)
+        batch_variants.append(batch_images)
+    ## variants, batch, c, t, h, w
+    batch_variants = torch.stack(batch_variants)
+    return batch_variants.permute(1, 0, 2, 3, 4, 5), extra_cond
+
+
+def run_inference(args, gpu_idx):
+    ## model config
+    config = OmegaConf.load(args.base)
+    model_config = config.pop("model", OmegaConf.create())
+    model = instantiate_from_config(model_config)
+    model = model.cuda(gpu_idx)
+    assert os.path.exists(args.ckpt_path), "Error: checkpoint Not Found!"
+    model = load_model_checkpoint(model, args.ckpt_path, args.adapter_ckpt)
+    model.eval()
+
+    ## run over data
+    assert (args.height % 16 == 0) and (args.width % 16 == 0), "Error: image size [h,w] should be multiples of 16!"
+    ## latent noise shape
+    h, w = args.height // 8, args.width // 8
+    channels = model.channels
+    frames = model.temporal_length
+    noise_shape = [args.bs, channels, frames, h, w]
+    
+    ## inference
+    start = time.time()
+    prompt = args.prompt
+    video = load_video(args.video, args.frame_stride, video_size=(args.height, args.width), video_frames=16)
+    video = video.unsqueeze(0).to("cuda")
+    with torch.no_grad():
+        batch_samples, batch_conds = adapter_guided_synthesis(model, prompt, video, noise_shape, args.n_samples, args.ddim_steps, args.ddim_eta, \
+                                                args.unconditional_guidance_scale, args.unconditional_guidance_scale_temporal)
+    batch_samples = batch_samples[0]
+    os.makedirs(args.savedir, exist_ok=True)
+    filename = f"{args.prompt}_seed{args.seed}"
+    filename = filename.replace("/", "_slash_") if "/" in filename else filename
+    filename = filename.replace(" ", "_") if " " in filename else filename
+    tensor_to_mp4(video=batch_conds.detach().cpu(), savepath=os.path.join(args.savedir, f'{filename}_depth.mp4'), fps=10)
+    tensor_to_mp4(video=batch_samples.detach().cpu(), savepath=os.path.join(args.savedir, f'{filename}_sample.mp4'), fps=10)
+
+    print(f"Saved in {args.savedir}. Time used: {(time.time() - start):.2f} seconds")
+
+
+def get_parser():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--savedir", type=str, default=None, help="results saving path")
+    parser.add_argument("--ckpt_path", type=str, default=None, help="checkpoint path")
+    parser.add_argument("--adapter_ckpt", type=str, default=None, help="adapter checkpoint path")
+    parser.add_argument("--base", type=str, help="config (yaml) path")
+    parser.add_argument("--prompt", type=str, default=None, help="prompt string")
+    parser.add_argument("--video", type=str, default=None, help="video path")
+    parser.add_argument("--n_samples", type=int, default=1, help="num of samples per prompt",)
+    parser.add_argument("--ddim_steps", type=int, default=50, help="steps of ddim if positive, otherwise use DDPM",)
+    parser.add_argument("--ddim_eta", type=float, default=1.0, help="eta for ddim sampling (0.0 yields deterministic sampling)",)
+    parser.add_argument("--bs", type=int, default=1, help="batch size for inference")
+    parser.add_argument("--height", type=int, default=512, help="image height, in pixel space")
+    parser.add_argument("--width", type=int, default=512, help="image width, in pixel space")
+    parser.add_argument("--frame_stride", type=int, default=-1, help="frame extracting from input video")
+    parser.add_argument("--unconditional_guidance_scale", type=float, default=1.0, help="prompt classifier-free guidance")
+    parser.add_argument("--unconditional_guidance_scale_temporal", type=float, default=None, help="temporal consistency guidance")
+    parser.add_argument("--seed", type=int, default=2023, help="seed for seed_everything")
+    return parser
+
+
+if __name__ == '__main__':
+    now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
+    print("@CoVideoGen cond-Inference: %s"%now)
+    parser = get_parser()
+    args = parser.parse_args()
+
+    seed_everything(args.seed)
+    rank = 0
+    run_inference(args, rank)