Merge pull request 'dev' (#2) from dev into main

Reviewed-on: #2

main.py

@@ -70,8 +70,10 @@ def init_anchor(device: "torch.device") -> Anchor:
         raise Exception(f"Unsupported device type: {device.type}")
 
 
-def init_model_runner(preset: presets.Preset, device: "torch.device") -> ModelRunner:
-    return ModelRunner(preset, device)
+def init_model_runner(
+    config: Path, checkpoint_path: Path, device: "torch.device"
+) -> ModelRunner:
+    return ModelRunner(config, checkpoint_path, device)
 
 
 def init_interpolator(
@@ -84,20 +86,21 @@ def init_interpolator(
 class InterpolationPipeline:
     def __init__(
         self,
-        preset: presets.Preset,
+        config: Path,
+        checkpoint_path: Path,
         base_path: Path,
     ):
         self.fs = init_fs(base_path)
         self.video_maker = init_video_maker()
         self.device = init_device()
-        self.model_runner = init_model_runner(preset, self.device)
+        self.model_runner = init_model_runner(config, checkpoint_path, self.device)
         self.interpolator = init_interpolator(self.model_runner, self.device)
 
     def run(self, video_path: Path, output_video: str):
         prev_frames = tuple()
         interpolated_frames: list["np.ndarray"] = []
         part = 0
-        chunk_seconds = 1
+        chunk_seconds = 10
         length = self.video_maker.get_video_duration(video_path)
         last_part_seconds = 1 if length % chunk_seconds else 0
         total_parts = int(length // chunk_seconds) + last_part_seconds
@@ -186,7 +189,8 @@ def runner(
     preset: presets.Preset = presets.LARGE,
 ):
     pipeline = InterpolationPipeline(
-        preset=preset,
+        config=preset.config,
+        checkpoint_path=preset.checkpoint,
         base_path=base_path,
     )
     pipeline.run(video_path, output_video)

onnx_export.py

@@ -1,8 +0,0 @@
-import torch
-from src.export_to_onnx import export_to_onnx
-from src.config.presets import SMALL
-
-
-if __name__ == "__main__":
-    device = torch.device("cuda")
-    export_to_onnx(SMALL, "src/pretrained/amt_s.onnx", device)

pyproject.toml

@@ -7,14 +7,8 @@ requires-python = ">=3.12"
 dependencies = [
     "imageio>=2.37.3",
     "numpy>=2.4.4",
-    "nvidia-modelopt[all]>=0.33.1",
     "omegaconf>=2.3.0",
-    "onnx>=1.21.0",
-    "onnxscript>=0.6.2",
     "opencv-python>=4.13.0.92",
-    "tensorrt>=10.16.1.11",
-    "torch==2.5.1",
-    "torch-tensorrt>=2.5.0",
-    "torchvision>=0.20.1",
+    "torch>=2.11.0",
     "tqdm>=4.67.3",
 ]

src/config/presets.py

@@ -1,4 +1,3 @@
-from typing import Literal
 from pathlib import Path
 from dataclasses import dataclass
 
@@ -7,7 +6,6 @@ from dataclasses import dataclass
 class Preset:
     config: Path
     checkpoint: Path
-    onnx: Path | None = None
 
 
 SMALL = Preset(

src/export_to_onnx.py

@@ -1,29 +0,0 @@
-import torch
-import torchvision
-
-torchvision.disable_beta_transforms_warning()
-torch.backends.cudnn.enabled = False
-from .interpolator import ModelRunner
-from .config.presets import Preset
-
-
-def export_to_onnx(preset: Preset, output_path: str, device: torch.device):
-    model_runner = ModelRunner(preset, device)
-    # model_runner.model.eval()
-    dummy_input = model_runner.get_dummy_input()
-    torch.onnx.export(
-        model_runner.model,
-        dummy_input,
-        output_path,
-        opset_version=17,
-        input_names=['img0', 'img1', 'embt'],
-        output_names=["imgt_pred"],
-        dynamic_axes={
-            "img0": {0: "batch", 2: "height", 3: "width"},
-            "img1": {0: "batch", 2: "height", 3: "width"},
-            "embt": {0: "batch", 2: "height", 3: "width"},
-            "imgt_pred": {0: "batch", 2: "height", 3: "width"},
-        },
-        dynamo=True,
-        use_external_data_format=False,
-    )

src/interpolator.py

@@ -1,16 +1,13 @@
 import logging
 from pathlib import Path
-from typing import Optional
 
-from cv2 import imread
 import torch
-import onnxruntime as ort
 import numpy as np
 from omegaconf import OmegaConf, DictConfig
 
-from src.config.presets import Preset
-from src.utils.torch import img2tensor, tensor2img
+from src.utils.torch import img2tensor, check_dim_and_resize, tensor2img
 from src.utils.build import build_from_cfg
+from src.utils.padder import InputPadder
 
 
 class Anchor:
@@ -23,27 +20,8 @@ class Anchor:
         return f"Anchor(resolution={self.resolution}, memory={self.memory}, memory_bias={self.memory_bias})"
 
 
-class ONNXWrapper:
-    def __init__(self, path):
-        self.session = ort.InferenceSession(path)
-
-        self.input_names = [i.name for i in self.session.get_inputs()]
-        self.output_names = [o.name for o in self.session.get_outputs()]
-
-    def __call__(self, tensor1, tensor2, embt):
-        inputs = {
-            self.input_names[0]: tensor1.cpu().numpy(),
-            self.input_names[1]: tensor2.cpu().numpy(),
-            self.input_names[2]: embt.cpu().numpy(),
-        }
-
-        outputs = self.session.run(self.output_names, inputs)
-
-        return {"imgt_pred": torch.from_numpy(outputs[0])}
-
-
 class ModelRunner:
-    def __init__(self, preset: Preset, device: torch.device) -> None:
+    def __init__(self, config: Path, ckpt_path: Path, device: torch.device) -> None:
         """Initializes the ModelRunner with configuration and checkpoint.
 
         Args:
@@ -51,73 +29,17 @@ class ModelRunner:
             ckpt_path (Path): Path to model checkpoint in .pth format
             device (torch.device): Device to load the model on
         """
-        self.model: Optional[torch.nn.Module] = None
-        self.session: Optional[ONNXWrapper] = None
-        self.embt = torch.tensor(1 / 2).float().view(1, 1, 1, 1).to(device)
-
-        if preset.onnx:
-            self.session = ONNXWrapper(preset.onnx)
-            self.device = device
-            self.embt = self.embt.cpu().numpy()
-            return
-
-        omega_config = OmegaConf.load(preset.config)
+        omega_config = OmegaConf.load(config)
         network_config: DictConfig = omega_config.network
         logging.info(
-            f"Loaded network configuration: {network_config} from [{preset.checkpoint}]"
+            f"Loaded network configuration: {network_config} from [{ckpt_path}]"
         )
         model = build_from_cfg(network_config)
-        checkpoint = torch.load(
-            preset.checkpoint, map_location=device, weights_only=False
-        )
+        checkpoint = torch.load(ckpt_path, map_location=device, weights_only=False)
         model.load_state_dict(checkpoint["state_dict"])
-        model = model.to(device)
+        model = model.to(get_device())
         model.eval()
-        # self.model = torch.compile(model)
         self.model = model
-        self.device = device
-        # self.model = torch.compile(model, backend="tensorrt")
-        if logging.getLogger().isEnabledFor(logging.DEBUG):
-            for name, param in self.model.named_parameters():
-                logging.debug(
-                    f"Parameter: {name}, shape: {param.shape}, dtype: {param.dtype}"
-                )
-
-    def get_dummy_input(self):
-        """Generates a dummy input tensor for ONNX export."""
-        return (
-            img2tensor(imread(filename="example/frame_01.png"), self.device),
-            img2tensor(imread(filename="example/frame_02.png"), self.device),
-            self.embt,
-        )
-
-    def run(self, image1: np.ndarray, image2: np.ndarray) -> np.ndarray:
-        """Runs the model inference to interpolate between two images.
-
-        Args:
-            image1 (np.ndarray): First input image as a NumPy array
-            image2 (np.ndarray): Second input image as a NumPy array
-
-        Returns:
-            np.ndarray: Interpolated image as a NumPy array
-        """
-        if self.session:
-            image1 = img2tensor(image1, self.device).cpu().numpy()
-            image2 = img2tensor(image2, self.device).cpu().numpy()
-            inputs = {
-                "img0": image1,
-                "img1": image2,
-                "embt": self.embt,
-            }
-            outputs = self.session.session.run(None, inputs)
-            return outputs[0]
-
-        tensor1 = img2tensor(image1, self.device)
-        tensor2 = img2tensor(image2, self.device)
-        with torch.no_grad():
-            with torch.amp.autocast(self.device.type):
-                interpolated = self.model(tensor1, tensor2, self.embt)["imgt_pred"]
-        return tensor2img(interpolated.cpu())
 
 
 def get_vram_available(device: torch.device) -> int:
@@ -154,6 +76,7 @@ class ImageInterpolator:
         self.device = device
         self.anchor = anchor
         self.vram_available = get_vram_available(device)
+        self.embt = torch.tensor(1 / 2).float().view(1, 1, 1, 1).to(device)
         self.model_runner = model_runner
         logging.debug(
             f"Initialized ImageInterpolator with device: {device}, anchor: {anchor}, available VRAM: {self.vram_available} bytes"
@@ -167,7 +90,38 @@ class ImageInterpolator:
             image2 (Path): Path to the second input image (only png and jpg formats are supported)
             output_path (Path): Path to save the interpolated image (only png and jpg formats are supported)
         """
-        return self.model_runner.run(image1, image2)
+        logging.debug(f"Reading images: {image1} and {image2}")
+        tensor1 = img2tensor(image1).to(self.device)
+        tensor2 = img2tensor(image2).to(self.device)
+        logging.debug(
+            f"Image shapes after conversion to tensors: {tensor1.shape}, {tensor2.shape}"
+        )
+        tensor1, tensor2 = check_dim_and_resize(tensor1, tensor2)
+        logging.debug(f"Image shapes after resizing: {tensor1.shape}, {tensor2.shape}")
+        h, w = tensor1.shape[2], tensor1.shape[3]
+        logging.debug(f"Interpolating images of size: {h}x{w}")
+
+        scale = self.scale(h, w)
+        logging.debug(f"Calculated scale factor: {scale:.2f}")
+        padding = int(16 / scale)
+        logging.debug(f"Calculated padding: {padding} pixels")
+        padder = InputPadder(tensor1.shape, divisor=padding)
+        tensor1_padded, tensor2_padded = padder.pad(tensor1, tensor2)
+        logging.debug(
+            f"Image shapes after padding: {tensor1_padded.shape}, {tensor2_padded.shape}"
+        )
+
+        tensor1_padded = tensor1_padded.to(self.device)
+        tensor2_padded = tensor2_padded.to(self.device)
+        logging.debug("Running model inference for interpolation")
+        with torch.no_grad():
+            interpolated = self.model_runner.model(
+                tensor1_padded, tensor2_padded, self.embt, scale_factor=scale, eval=True
+            )["imgt_pred"]
+        logging.debug(f"Interpolated image shape before unpadding: {interpolated.shape}")
+        (interpolated,) = padder.unpad(interpolated)
+        logging.debug(f"Interpolated image shape after unpadding: {interpolated.shape}")
+        return tensor2img(interpolated.cpu())
 
     def scale(self, height: int, width: int) -> float:
         scale = (

src/networks/AMT-S.py

@@ -67,14 +67,7 @@ class Model(nn.Module):
         flow = torch.cat([flow0, flow1], dim=1)
         return corr, flow
 
-    def forward(
-        self,
-        img0: torch.Tensor,
-        img1: torch.Tensor,
-        embt: torch.Tensor,
-    ):
-        scale_factor = 1.0
-        eval = False
+    def forward(self, img0, img1, embt, scale_factor=1.0, eval=False, **kwargs):
         mean_ = (
             torch.cat([img0, img1], 2)
             .mean(1, keepdim=True)

src/networks/blocks/feat_enc.py

@@ -1,44 +1,40 @@
-from typing import Any
-
 import torch
 import torch.nn as nn
 
 
 class BottleneckBlock(nn.Module):
-    def __init__(self, in_planes, planes, norm_fn="group", stride=1):
+    def __init__(self, in_planes, planes, norm_fn='group', stride=1):
         super(BottleneckBlock, self).__init__()
   
         self.conv1 = nn.Conv2d(in_planes, planes//4, kernel_size=1, padding=0)
-        self.conv2 = nn.Conv2d(
-            planes // 4, planes // 4, kernel_size=3, padding=1, stride=stride
-        )
+        self.conv2 = nn.Conv2d(planes//4, planes//4, kernel_size=3, padding=1, stride=stride)
         self.conv3 = nn.Conv2d(planes//4, planes, kernel_size=1, padding=0)
         self.relu = nn.ReLU(inplace=True)
 
         num_groups = planes // 8
 
-        if norm_fn == "group":
+        if norm_fn == 'group':
             self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes//4)
             self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes//4)
             self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
             if not stride == 1:
                 self.norm4 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
         
-        elif norm_fn == "batch":
+        elif norm_fn == 'batch':
             self.norm1 = nn.BatchNorm2d(planes//4)
             self.norm2 = nn.BatchNorm2d(planes//4)
             self.norm3 = nn.BatchNorm2d(planes)
             if not stride == 1:
                 self.norm4 = nn.BatchNorm2d(planes)
         
-        elif norm_fn == "instance":
+        elif norm_fn == 'instance':
             self.norm1 = nn.InstanceNorm2d(planes//4)
             self.norm2 = nn.InstanceNorm2d(planes//4)
             self.norm3 = nn.InstanceNorm2d(planes)
             if not stride == 1:
                 self.norm4 = nn.InstanceNorm2d(planes)
 
-        elif norm_fn == "none":
+        elif norm_fn == 'none':
             self.norm1 = nn.Sequential()
             self.norm2 = nn.Sequential()
             self.norm3 = nn.Sequential()
@@ -50,8 +46,8 @@ class BottleneckBlock(nn.Module):
         
         else:    
             self.downsample = nn.Sequential(
-                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm4
-            )
+                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm4)
+
 
     def forward(self, x):
         y = x
@@ -66,36 +62,34 @@ class BottleneckBlock(nn.Module):
 
 
 class ResidualBlock(nn.Module):
-    def __init__(self, in_planes, planes, norm_fn="group", stride=1):
+    def __init__(self, in_planes, planes, norm_fn='group', stride=1):
         super(ResidualBlock, self).__init__()
   
-        self.conv1 = nn.Conv2d(
-            in_planes, planes, kernel_size=3, padding=1, stride=stride
-        )
+        self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, padding=1, stride=stride)
         self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1)
         self.relu = nn.ReLU(inplace=True)
 
         num_groups = planes // 8
 
-        if norm_fn == "group":
+        if norm_fn == 'group':
             self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
             self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
             if not stride == 1:
                 self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
         
-        elif norm_fn == "batch":
+        elif norm_fn == 'batch':
             self.norm1 = nn.BatchNorm2d(planes)
             self.norm2 = nn.BatchNorm2d(planes)
             if not stride == 1:
                 self.norm3 = nn.BatchNorm2d(planes)
         
-        elif norm_fn == "instance":
+        elif norm_fn == 'instance':
             self.norm1 = nn.InstanceNorm2d(planes)
             self.norm2 = nn.InstanceNorm2d(planes)
             if not stride == 1:
                 self.norm3 = nn.InstanceNorm2d(planes)
 
-        elif norm_fn == "none":
+        elif norm_fn == 'none':
             self.norm1 = nn.Sequential()
             self.norm2 = nn.Sequential()
             if not stride == 1:
@@ -106,8 +100,8 @@ class ResidualBlock(nn.Module):
         
         else:    
             self.downsample = nn.Sequential(
-                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm3
-            )
+                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm3)
+
 
     def forward(self, x):
         y = x
@@ -121,20 +115,20 @@ class ResidualBlock(nn.Module):
 
 
 class SmallEncoder(nn.Module):
-    def __init__(self, output_dim=128, norm_fn="batch", dropout=0.0):
+    def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0):
         super(SmallEncoder, self).__init__()
         self.norm_fn = norm_fn
 
-        if self.norm_fn == "group":
+        if self.norm_fn == 'group':
             self.norm1 = nn.GroupNorm(num_groups=8, num_channels=32)
             
-        elif self.norm_fn == "batch":
+        elif self.norm_fn == 'batch':
             self.norm1 = nn.BatchNorm2d(32)
 
-        elif self.norm_fn == "instance":
+        elif self.norm_fn == 'instance':
             self.norm1 = nn.InstanceNorm2d(32)
 
-        elif self.norm_fn == "none":
+        elif self.norm_fn == 'none':
             self.norm1 = nn.Sequential()
 
         self.conv1 = nn.Conv2d(3, 32, kernel_size=7, stride=2, padding=3)
@@ -153,7 +147,7 @@ class SmallEncoder(nn.Module):
 
         for m in self.modules():
             if isinstance(m, nn.Conv2d):
-                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
             elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)):
                 if m.weight is not None:
                     nn.init.constant_(m.weight, 1)
@@ -168,15 +162,14 @@ class SmallEncoder(nn.Module):
         self.in_planes = dim
         return nn.Sequential(*layers)
 
-    def forward(
-        self, x: torch.Tensor | list[torch.Tensor] | tuple[torch.Tensor, ...]
-    ):
+
+    def forward(self, x):
 
         # if input is list, combine batch dimension
-        batch_dim = None
-        if is_list := isinstance(x, tuple) or isinstance(x, list):
+        is_list = isinstance(x, tuple) or isinstance(x, list)
+        if is_list:
             batch_dim = x[0].shape[0]
-            x: torch.Tensor = torch.cat(x, dim=0)
+            x = torch.cat(x, dim=0)
 
         x = self.conv1(x)
         x = self.norm1(x)
@@ -190,30 +183,26 @@ class SmallEncoder(nn.Module):
         if self.training and self.dropout is not None:
             x = self.dropout(x)
 
-        if is_list and batch_dim is not None:
-            return torch.split(x, [batch_dim, batch_dim], dim=0)
+        if is_list:
+            x = torch.split(x, [batch_dim, batch_dim], dim=0)
 
         return x
 
-    def __call__(self, *args: Any, **kwds: Any) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
-        return super().__call__(*args, **kwds)
-
-
 class BasicEncoder(nn.Module):
-    def __init__(self, output_dim=128, norm_fn="batch", dropout=0.0):
+    def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0):
         super(BasicEncoder, self).__init__()
         self.norm_fn = norm_fn
 
-        if self.norm_fn == "group":
+        if self.norm_fn == 'group':
             self.norm1 = nn.GroupNorm(num_groups=8, num_channels=64)
             
-        elif self.norm_fn == "batch":
+        elif self.norm_fn == 'batch':
             self.norm1 = nn.BatchNorm2d(64)
 
-        elif self.norm_fn == "instance":
+        elif self.norm_fn == 'instance':
             self.norm1 = nn.InstanceNorm2d(64)
 
-        elif self.norm_fn == "none":
+        elif self.norm_fn == 'none':
             self.norm1 = nn.Sequential()
 
         self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3)
@@ -233,7 +222,7 @@ class BasicEncoder(nn.Module):
 
         for m in self.modules():
             if isinstance(m, nn.Conv2d):
-                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
             elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)):
                 if m.weight is not None:
                     nn.init.constant_(m.weight, 1)
@@ -248,6 +237,7 @@ class BasicEncoder(nn.Module):
         self.in_planes = dim
         return nn.Sequential(*layers)
 
+
     def forward(self, x):
 
         # if input is list, combine batch dimension
@@ -274,22 +264,21 @@ class BasicEncoder(nn.Module):
 
         return x
 
-
 class LargeEncoder(nn.Module):
-    def __init__(self, output_dim=128, norm_fn="batch", dropout=0.0):
+    def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0):
         super(LargeEncoder, self).__init__()
         self.norm_fn = norm_fn
 
-        if self.norm_fn == "group":
+        if self.norm_fn == 'group':
             self.norm1 = nn.GroupNorm(num_groups=8, num_channels=64)
             
-        elif self.norm_fn == "batch":
+        elif self.norm_fn == 'batch':
             self.norm1 = nn.BatchNorm2d(64)
 
-        elif self.norm_fn == "instance":
+        elif self.norm_fn == 'instance':
             self.norm1 = nn.InstanceNorm2d(64)
 
-        elif self.norm_fn == "none":
+        elif self.norm_fn == 'none':
             self.norm1 = nn.Sequential()
 
         self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3)
@@ -310,7 +299,7 @@ class LargeEncoder(nn.Module):
 
         for m in self.modules():
             if isinstance(m, nn.Conv2d):
-                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
             elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)):
                 if m.weight is not None:
                     nn.init.constant_(m.weight, 1)
@@ -325,6 +314,7 @@ class LargeEncoder(nn.Module):
         self.in_planes = dim
         return nn.Sequential(*layers)
 
+
     def forward(self, x):
 
         # if input is list, combine batch dimension

src/networks/blocks/ifrnet.py

@@ -4,78 +4,36 @@ import torch.nn.functional as F
 from src.utils.flow_utils import warp
 
 
-def resize(x: torch.Tensor, scale_factor: float) -> torch.Tensor:
-    return F.interpolate(
-        x, scale_factor=scale_factor, mode="bilinear", align_corners=False
-    )
+def resize(x, scale_factor):
+    return F.interpolate(x, scale_factor=scale_factor, mode="bilinear", align_corners=False)
 
-
-def convrelu(
-    in_channels,
-    out_channels,
-    kernel_size=3,
-    stride=1,
-    padding=1,
-    dilation=1,
-    groups=1,
-    bias=True,
-):
+def convrelu(in_channels, out_channels, kernel_size=3, stride=1, padding=1, dilation=1, groups=1, bias=True):
     return nn.Sequential(
-        nn.Conv2d(
-            in_channels,
-            out_channels,
-            kernel_size,
-            stride,
-            padding,
-            dilation,
-            groups,
-            bias=bias,
-        ),
-        nn.PReLU(out_channels),
+        nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias=bias), 
+        nn.PReLU(out_channels)
     )
 
-
 class ResBlock(nn.Module):
     def __init__(self, in_channels, side_channels, bias=True):
         super(ResBlock, self).__init__()
         self.side_channels = side_channels
         self.conv1 = nn.Sequential(
-            nn.Conv2d(
-                in_channels, in_channels, kernel_size=3, stride=1, padding=1, bias=bias
-            ),
-            nn.PReLU(in_channels),
+            nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1, bias=bias), 
+            nn.PReLU(in_channels)
         )
         self.conv2 = nn.Sequential(
-            nn.Conv2d(
-                side_channels,
-                side_channels,
-                kernel_size=3,
-                stride=1,
-                padding=1,
-                bias=bias,
-            ),
-            nn.PReLU(side_channels),
+            nn.Conv2d(side_channels, side_channels, kernel_size=3, stride=1, padding=1, bias=bias), 
+            nn.PReLU(side_channels)
         )
         self.conv3 = nn.Sequential(
-            nn.Conv2d(
-                in_channels, in_channels, kernel_size=3, stride=1, padding=1, bias=bias
-            ),
-            nn.PReLU(in_channels),
+            nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1, bias=bias), 
+            nn.PReLU(in_channels)
         )
         self.conv4 = nn.Sequential(
-            nn.Conv2d(
-                side_channels,
-                side_channels,
-                kernel_size=3,
-                stride=1,
-                padding=1,
-                bias=bias,
-            ),
-            nn.PReLU(side_channels),
-        )
-        self.conv5 = nn.Conv2d(
-            in_channels, in_channels, kernel_size=3, stride=1, padding=1, bias=bias
+            nn.Conv2d(side_channels, side_channels, kernel_size=3, stride=1, padding=1, bias=bias), 
+            nn.PReLU(side_channels)
         )
+        self.conv5 = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1, bias=bias)
         self.prelu = nn.PReLU(in_channels)
 
     def forward(self, x):
@@ -94,7 +52,6 @@ class ResBlock(nn.Module):
         out = self.prelu(x + out)
         return out
     
-
 class Encoder(nn.Module):
     def __init__(self, channels, large=False):
         super(Encoder, self).__init__()
@@ -103,32 +60,29 @@ class Encoder(nn.Module):
         for idx, ch in enumerate(channels, 1):
             k = 7 if large and idx == 1 else 3
             p = 3 if k ==7 else 1
-            self.register_module(
-                f"pyramid{idx}",
+            self.register_module(f'pyramid{idx}', 
             nn.Sequential(
-                    convrelu(prev_ch, ch, k, 2, p), convrelu(ch, ch, 3, 1, 1)
-                ),
-            )
+                convrelu(prev_ch, ch, k, 2, p),
+                convrelu(ch, ch, 3, 1, 1)
+            ))
             prev_ch = ch
                 
     def forward(self, in_x):
         fs = []
         for idx in range(len(self.channels)):
-            out_x = getattr(self, f"pyramid{idx + 1}")(in_x)
+            out_x = getattr(self, f'pyramid{idx+1}')(in_x)
             fs.append(out_x)
             in_x = out_x
         return fs
     
-
 class InitDecoder(nn.Module):
     def __init__(self, in_ch, out_ch, skip_ch) -> None:
         super().__init__()
         self.convblock = nn.Sequential(
             convrelu(in_ch*2+1, in_ch*2), 
             ResBlock(in_ch*2, skip_ch), 
-            nn.ConvTranspose2d(in_ch * 2, out_ch + 4, 4, 2, 1, bias=True),
+            nn.ConvTranspose2d(in_ch*2, out_ch+4, 4, 2, 1, bias=True)
         )
-
     def forward(self, f0, f1, embt):
         h, w = f0.shape[2:]
         embt = embt.repeat(1, 1, h, w)
@@ -137,16 +91,14 @@ class InitDecoder(nn.Module):
         ft_ = out[:, 4:, ...]
         return flow0, flow1, ft_
     
-
 class IntermediateDecoder(nn.Module):
     def __init__(self, in_ch, out_ch, skip_ch) -> None:
         super().__init__()
         self.convblock = nn.Sequential(
             convrelu(in_ch*3+4, in_ch*3), 
             ResBlock(in_ch*3, skip_ch), 
-            nn.ConvTranspose2d(in_ch * 3, out_ch + 4, 4, 2, 1, bias=True),
+            nn.ConvTranspose2d(in_ch*3, out_ch+4, 4, 2, 1, bias=True)
         )
-
     def forward(self, ft_, f0, f1, flow0_in, flow1_in):
         f0_warp = warp(f0, flow0_in)
         f1_warp = warp(f1, flow1_in)

src/networks/blocks/raft.py

@@ -4,12 +4,10 @@ import torch.nn.functional as F
 
 
 def resize(x, scale_factor):
-    return F.interpolate(
-        x, scale_factor=scale_factor, mode="bilinear", align_corners=False
-    )
+    return F.interpolate(x, scale_factor=scale_factor, mode="bilinear", align_corners=False)
 
 
-def bilinear_sampler(img: torch.Tensor, coords: torch.Tensor, mask=False):
+def bilinear_sampler(img, coords, mask=False):
     """ Wrapper for grid_sample, uses pixel coordinates """
     H, W = img.shape[-2:]
     xgrid, ygrid = coords.split([1,1], dim=-1)
@@ -27,25 +25,16 @@ def bilinear_sampler(img: torch.Tensor, coords: torch.Tensor, mask=False):
 
 
 def coords_grid(batch, ht, wd, device):
-    coords = torch.meshgrid(
-        torch.arange(ht, device=device), torch.arange(wd, device=device), indexing="ij"
-    )
+    coords = torch.meshgrid(torch.arange(ht, device=device), 
+                            torch.arange(wd, device=device), 
+                            indexing='ij')
     coords = torch.stack(coords[::-1], dim=0).float()
     return coords[None].repeat(batch, 1, 1, 1)
 
 
 class SmallUpdateBlock(nn.Module):
-    def __init__(
-        self,
-        cdim,
-        hidden_dim,
-        flow_dim,
-        corr_dim,
-        fc_dim,
-        corr_levels=4,
-        radius=3,
-        scale_factor=None,
-    ):
+    def __init__(self, cdim, hidden_dim, flow_dim, corr_dim, fc_dim,
+                 corr_levels=4, radius=3, scale_factor=None):
         super(SmallUpdateBlock, self).__init__()
         cor_planes = corr_levels * (2 * radius + 1) **2
         self.scale_factor = scale_factor
@@ -76,9 +65,8 @@ class SmallUpdateBlock(nn.Module):
         self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True)
             
     def forward(self, net, flow, corr):
-        net = (
-            resize(net, 1 / self.scale_factor) if self.scale_factor is not None else net
-        )
+        net = resize(net, 1 / self.scale_factor
+                      ) if self.scale_factor is not None else net
         cor = self.lrelu(self.convc1(corr))
         flo = self.lrelu(self.convf1(flow))
         flo = self.lrelu(self.convf2(flo))
@@ -92,27 +80,14 @@ class SmallUpdateBlock(nn.Module):
         
         if self.scale_factor is not None:
             delta_net = resize(delta_net, scale_factor=self.scale_factor)
-            delta_flow = self.scale_factor * resize(
-                delta_flow, scale_factor=self.scale_factor
-            )
+            delta_flow = self.scale_factor * resize(delta_flow, scale_factor=self.scale_factor)
         
         return delta_net, delta_flow
 
 
 class BasicUpdateBlock(nn.Module):
-    def __init__(
-        self,
-        cdim,
-        hidden_dim,
-        flow_dim,
-        corr_dim,
-        corr_dim2,
-        fc_dim,
-        corr_levels=4,
-        radius=3,
-        scale_factor=None,
-        out_num=1,
-    ):
+    def __init__(self, cdim, hidden_dim, flow_dim, corr_dim, corr_dim2, 
+                 fc_dim, corr_levels=4, radius=3, scale_factor=None, out_num=1):
         super(BasicUpdateBlock, self).__init__()
         cor_planes = corr_levels * (2 * radius + 1) **2
 
@@ -144,9 +119,8 @@ class BasicUpdateBlock(nn.Module):
         self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True)
             
     def forward(self, net, flow, corr):
-        net = (
-            resize(net, 1 / self.scale_factor) if self.scale_factor is not None else net
-        )
+        net = resize(net, 1 / self.scale_factor
+                      ) if self.scale_factor is not None else net
         cor = self.lrelu(self.convc1(corr))
         cor = self.lrelu(self.convc2(cor))
         flo = self.lrelu(self.convf1(flow))
@@ -161,20 +135,16 @@ class BasicUpdateBlock(nn.Module):
         
         if self.scale_factor is not None:
             delta_net = resize(delta_net, scale_factor=self.scale_factor)
-            delta_flow = self.scale_factor * resize(
-                delta_flow, scale_factor=self.scale_factor
-            )
+            delta_flow = self.scale_factor * resize(delta_flow, scale_factor=self.scale_factor)
         return delta_net, delta_flow
 
 
 class BidirCorrBlock:
-    def __init__(
-        self, fmap1: torch.Tensor, fmap2: torch.Tensor, num_levels=4, radius=4
-    ):
+    def __init__(self, fmap1, fmap2, num_levels=4, radius=4):
         self.num_levels = num_levels
         self.radius = radius
-        self.corr_pyramid: list[torch.Tensor] = []
-        self.corr_pyramid_T: list[torch.Tensor] = []
+        self.corr_pyramid = []
+        self.corr_pyramid_T = []
 
         corr = BidirCorrBlock.corr(fmap1, fmap2)
         batch, h1, w1, dim, h2, w2 = corr.shape
@@ -192,13 +162,11 @@ class BidirCorrBlock:
             self.corr_pyramid.append(corr)
             self.corr_pyramid_T.append(corr_T)
 
-    def __call__(self, coords0: torch.Tensor, coords1: torch.Tensor):
+    def __call__(self, coords0, coords1):
         r = self.radius
         coords0 = coords0.permute(0, 2, 3, 1)
         coords1 = coords1.permute(0, 2, 3, 1)
-        assert coords0.shape == coords1.shape, (
-            f"coords0 shape: [{coords0.shape}] is not equal to [{coords1.shape}]"
-        )
+        assert coords0.shape == coords1.shape, f"coords0 shape: [{coords0.shape}] is not equal to [{coords1.shape}]"
         batch, h1, w1, _ = coords0.shape
 
         out_pyramid = []
@@ -209,13 +177,13 @@ class BidirCorrBlock:
 
             dx = torch.linspace(-r, r, 2*r+1, device=coords0.device)
             dy = torch.linspace(-r, r, 2*r+1, device=coords0.device)
-            delta: torch.Tensor = torch.stack(torch.meshgrid(dy, dx, indexing="ij"), axis=-1)
-            delta_lvl: torch.Tensor = delta.view(1, 2 * r + 1, 2 * r + 1, 2)
+            delta = torch.stack(torch.meshgrid(dy, dx, indexing='ij'), axis=-1)
+            delta_lvl = delta.view(1, 2*r+1, 2*r+1, 2)
 
-            centroid_lvl_0: torch.Tensor = coords0.reshape(batch * h1 * w1, 1, 1, 2) / 2**i
-            centroid_lvl_1: torch.Tensor = coords1.reshape(batch * h1 * w1, 1, 1, 2) / 2**i
-            coords_lvl_0: torch.Tensor = centroid_lvl_0 + delta_lvl
-            coords_lvl_1: torch.Tensor = centroid_lvl_1 + delta_lvl
+            centroid_lvl_0 = coords0.reshape(batch*h1*w1, 1, 1, 2) / 2**i
+            centroid_lvl_1 = coords1.reshape(batch*h1*w1, 1, 1, 2) / 2**i
+            coords_lvl_0 = centroid_lvl_0 + delta_lvl
+            coords_lvl_1 = centroid_lvl_1 + delta_lvl
 
             corr = bilinear_sampler(corr, coords_lvl_0)
             corr_T = bilinear_sampler(corr_T, coords_lvl_1)
@@ -226,16 +194,14 @@ class BidirCorrBlock:
 
         out = torch.cat(out_pyramid, dim=-1)
         out_T = torch.cat(out_pyramid_T, dim=-1)
-        return out.permute(0, 3, 1, 2).contiguous().float(), out_T.permute(
-            0, 3, 1, 2
-        ).contiguous().float()
+        return out.permute(0, 3, 1, 2).contiguous().float(), out_T.permute(0, 3, 1, 2).contiguous().float()
 
     @staticmethod
-    def corr(fmap1: torch.Tensor, fmap2: torch.Tensor):
+    def corr(fmap1, fmap2):
         batch, dim, ht, wd = fmap1.shape
         fmap1 = fmap1.view(batch, dim, ht*wd)
         fmap2 = fmap2.view(batch, dim, ht*wd) 
         
         corr = torch.matmul(fmap1.transpose(1,2), fmap2)
         corr = corr.view(batch, ht, wd, 1, ht, wd)
-        return corr * (dim**-0.5)
+        return corr  / torch.sqrt(torch.tensor(dim).float())

src/utils/torch.py

@@ -5,26 +5,23 @@ import numpy as np
 
 
 def tensor2img(tensor: torch.Tensor):
-    tensor = (
-        tensor.mul(255.0)
-        .clamp_(0, 255)
-        .to(torch.uint8)
+    return (
+        (tensor * 255.0)
+        .detach()
         .squeeze(0)
         .permute(1, 2, 0)
+        .cpu()
+        .numpy()
+        .clip(0, 255)
+        .astype(np.uint8)
     )
 
-    return tensor.cpu().numpy()
 
-
-def img2tensor(img: np.ndarray, device: torch.device) -> torch.Tensor:
+def img2tensor(img: np.ndarray) -> torch.Tensor:
     logging.debug(f"Converting image of shape {img.shape} to tensor")
     if img.shape[-1] > 3:
         img = img[:, :, :3]
-    tensor = torch.from_numpy(img).permute(2, 0, 1).unsqueeze(0)
-    if device.type != "cuda":
-        return tensor.float() / 255.0
-    
-    return tensor.cuda(non_blocking=True).float().div_(255.0)
+    return torch.tensor(img).permute(2, 0, 1).unsqueeze(0) / 255.0
 
 
 def check_dim_and_resize(*args: torch.Tensor) -> list[torch.Tensor]: