Попытка добавить onnx в работе с nvidia

main.py

@@ -2,6 +2,7 @@ import logging
 from pathlib import Path
 from typing import TYPE_CHECKING
 
+from cv2 import imwrite
 import tqdm
 
 from src.config import presets
@@ -18,6 +19,7 @@ from src.interpolator import (
 
 if TYPE_CHECKING:
     import torch
+    import numpy as np
 
 
 def performing_warning_message(device: "torch.device"):
@@ -53,7 +55,7 @@ def init_device() -> "torch.device":
     device = get_device()
     performing_warning_message(device)
     vram_available = get_vram_available(device)
-    logging.info(f"Available VRAM: {vram_available / (1024 ** 3):.2f} GB")
+    logging.info(f"Available VRAM: {vram_available / (1024**3):.2f} GB")
     return device
 
 
@@ -68,10 +70,8 @@ def init_anchor(device: "torch.device") -> Anchor:
         raise Exception(f"Unsupported device type: {device.type}")
 
 
-def init_model_runner(
-    config: Path, checkpoint_path: Path, device: "torch.device"
-) -> ModelRunner:
-    return ModelRunner(config, checkpoint_path, device)
+def init_model_runner(preset: presets.Preset, device: "torch.device") -> ModelRunner:
+    return ModelRunner(preset, device)
 
 
 def init_interpolator(
@@ -84,63 +84,58 @@ def init_interpolator(
 class InterpolationPipeline:
     def __init__(
         self,
-        config: Path,
-        checkpoint_path: Path,
+        preset: presets.Preset,
         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(config, checkpoint_path, self.device)
+        self.model_runner = init_model_runner(preset, self.device)
         self.interpolator = init_interpolator(self.model_runner, self.device)
 
     def run(self, video_path: Path, output_video: str):
-        prev_frame_path = None
-        frame_count = 0
+        prev_frames = tuple()
+        interpolated_frames: list["np.ndarray"] = []
         part = 0
-        source_frame_length = 0
-        chunk_seconds = 10
+        chunk_seconds = 1
         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
         fps = self.video_maker.get_fps(video_path)
         logging.info(f"Video FPS: {fps}")
         fps *= 2  # Doubling FPS
-        for frame_paths in self.video_maker.video_to_frames_generator(
+        width, height = self.video_maker.get_size(video_path)
+        for frames in self.video_maker.video_to_frames_generator(
             video_path, self.fs.frames_path, chunk_seconds
         ):
-            logging.info(f"Processing frames: {len(frame_paths)}")
-            if prev_frame_path is not None:
-                img1 = prev_frame_path[-1]
-                img2 = frame_paths[0]
-                output_path = self.fs.interpolated_path / f"img_{frame_count:08d}.png"
-                self.interpolator.interpolate(img1, img2, output_path)
-                logging.debug(f"Interpolated image saved to: {output_path}")
-                self._merge_frames_to_video(
-                    self.fs.video_part_path / f"video_{part:08d}.mp4",
-                    fps,
-                    source_frame_length=source_frame_length,
+            logging.info(f"Processing frames: {len(frames)}")
+            if prev_frames:
+                img1 = prev_frames[-1]
+                img2 = frames[0]
+                img1_2 = self.interpolator.interpolate(img1, img2)
+                interpolated_frames.append(img1_2)
+                generator = self._frame_generator(prev_frames, interpolated_frames)
+                part_path = self.fs.video_part_path / f"video_{part:08d}.mp4"
+                self.video_maker.images_to_video_pipeline(
+                    generator, part_path, width, height, fps
                 )
+                interpolated_frames = []
                 logging.info(f"Finished processing part {part:08d}")
-                frame_count += 1
                 part += 1
             for i in tqdm.tqdm(
-                range(len(frame_paths) - 1),
+                range(len(frames) - 1),
                 desc=f"Processing video frames {part + 1} / {total_parts}",
             ):
-                img1 = frame_paths[i]
-                img2 = frame_paths[i + 1]
-                output_path = self.fs.interpolated_path / f"img_{i:08d}.png"
-                self.interpolator.interpolate(img1, img2, output_path)
-                logging.debug(f"Interpolated image saved to: {output_path}")
-                frame_count += 1
-            source_frame_length = len(frame_paths)
-            prev_frame_path = frame_paths
+                img1 = frames[i]
+                img2 = frames[i + 1]
+                img1_2 = self.interpolator.interpolate(img1, img2)
+                interpolated_frames.append(img1_2)
+            prev_frames = frames
 
-        self._merge_frames_to_video(
-            self.fs.video_part_path / f"video_{part:08d}.mp4",
-            fps,
-            source_frame_length=source_frame_length,
+        generator = self._frame_generator(prev_frames, interpolated_frames)
+        part_path = self.fs.video_part_path / f"video_{part:08d}.mp4"
+        self.video_maker.images_to_video_pipeline(
+            generator, part_path, width, height, fps
         )
         logging.info(f"Finished processing part {part:08d}")
         self._merge_video_parts(self.fs.output_path / output_video)
@@ -148,32 +143,40 @@ class InterpolationPipeline:
             f"Video interpolation completed. Output saved to: {self.fs.output_path / output_video}"
         )
 
-    def _merge_frames_to_video(
-        self, output_video: Path, fps: float, source_frame_length: int = 0
+    def _save_images(
+        self,
+        source: tuple["np.ndarray", ...],
+        interpolated: list["np.ndarray"],
     ):
-        self._move_frames(source_frame_length)
+        logging.info("Saving images...")
+        self.fs.clear_directory(self.fs.moved_path)
+        index = 0
+        for i, frame in enumerate(source):
+            name = self.fs.moved_path / f"img_{index:08d}.png"
+            index += 1
+            imwrite(name, frame)
+            if i < len(interpolated):
+                name = self.fs.moved_path / f"img_{index:08d}.png"
+                index += 1
+                imwrite(name, interpolated[i])
+        logging.info("Success...")
+
+    def _merge_frames_to_video(self, output_video: Path, fps: float):
         self.video_maker.images_to_video(self.fs.moved_path, output_video, fps)
 
     def _merge_video_parts(self, output_video: Path):
         self.video_maker.concatenate_videos(self.fs.video_part_path, output_video)
         self.fs.clear_directory(self.fs.video_part_path)
 
-    def _move_frames(self, source_frame_length: int = 0):
-        self.fs.clear_directory(self.fs.moved_path)
-        src_frames = sorted(self.fs.frames_path.glob("*.png"))
-        interpolated_frames = sorted(self.fs.interpolated_path.glob("*.png"))
-        index = 0
-        for i in range(source_frame_length):
-            moved_frame_path = self.fs.moved_path / f"img_{index:08d}.png"
-            src_frames[i].rename(moved_frame_path)
-            index += 1
-            if i < len(interpolated_frames):
-                moved_interpolated_path = self.fs.moved_path / f"img_{index:08d}.png"
-                interpolated_frames[i].rename(moved_interpolated_path)
-                index += 1
-        logging.info(
-            f"Moved {len(src_frames)} source frames and {len(interpolated_frames)} interpolated frames to {self.fs.moved_path}"
-        )
+    def _frame_generator(
+        self,
+        source: tuple["np.ndarray", ...],
+        interpolated: list["np.ndarray"],
+    ):
+        for i, frame in enumerate(source):
+            yield frame
+            if i < len(interpolated):
+                yield interpolated[i]
 
 
 def runner(
@@ -183,8 +186,7 @@ def runner(
     preset: presets.Preset = presets.LARGE,
 ):
     pipeline = InterpolationPipeline(
-        config=preset.config,
-        checkpoint_path=preset.checkpoint,
+        preset=preset,
         base_path=base_path,
     )
     pipeline.run(video_path, output_video)
@@ -220,7 +222,7 @@ def main():
         base_path=Path(args.base_path),
         video_path=Path(args.video_path),
         output_video=args.output,
-        preset=getattr(presets, args.preset.upper())
+        preset=getattr(presets, args.preset.upper()),
     )
 
 

onnx_export.py

@@ -0,0 +1,8 @@
+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

@@ -7,8 +7,14 @@ 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",
-    "torch>=2.11.0",
+    "tensorrt>=10.16.1.11",
+    "torch==2.5.1",
+    "torch-tensorrt>=2.5.0",
+    "torchvision>=0.20.1",
     "tqdm>=4.67.3",
 ]

src/config/presets.py

@@ -1,3 +1,4 @@
+from typing import Literal
 from pathlib import Path
 from dataclasses import dataclass
 
@@ -6,6 +7,7 @@ from dataclasses import dataclass
 class Preset:
     config: Path
     checkpoint: Path
+    onnx: Path | None = None
 
 
 SMALL = Preset(
@@ -19,6 +21,6 @@ LARGE = Preset(
 )
 
 GLOBAL = Preset(
-    config=Path("src/config/AMT-g.yaml"),
+    config=Path("src/config/AMT-G.yaml"),
     checkpoint=Path("src/pretrained/amt-g.pth"),
 )

src/export_to_onnx.py

@@ -0,0 +1,29 @@
+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

@@ -1,14 +1,16 @@
 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 imageio import imread, imwrite
 
-from src.utils.torch import img2tensor, check_dim_and_resize, tensor2img
+from src.config.presets import Preset
+from src.utils.torch import img2tensor, tensor2img
 from src.utils.build import build_from_cfg
-from src.utils.padder import InputPadder
 
 
 class Anchor:
@@ -21,8 +23,27 @@ 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, config: Path, ckpt_path: Path, device: torch.device) -> None:
+    def __init__(self, preset: Preset, device: torch.device) -> None:
         """Initializes the ModelRunner with configuration and checkpoint.
 
         Args:
@@ -30,17 +51,73 @@ class ModelRunner:
             ckpt_path (Path): Path to model checkpoint in .pth format
             device (torch.device): Device to load the model on
         """
-        omega_config = OmegaConf.load(config)
+        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)
         network_config: DictConfig = omega_config.network
         logging.info(
-            f"Loaded network configuration: {network_config} from [{ckpt_path}]"
+            f"Loaded network configuration: {network_config} from [{preset.checkpoint}]"
         )
         model = build_from_cfg(network_config)
-        checkpoint = torch.load(ckpt_path, map_location=device, weights_only=False)
+        checkpoint = torch.load(
+            preset.checkpoint, map_location=device, weights_only=False
+        )
         model.load_state_dict(checkpoint["state_dict"])
-        model = model.to(get_device())
+        model = model.to(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:
@@ -77,13 +154,12 @@ 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"
         )
 
-    def interpolate(self, image1: Path, image2: Path, output_path: Path):
+    def interpolate(self, image1: np.ndarray, image2: np.ndarray) -> np.ndarray:
         """
         Interpolates between two images and saves the result.
         Args:
@@ -91,39 +167,7 @@ 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)
         """
-        logging.debug(f"Reading images: {image1} and {image2}")
-        tensor1 = img2tensor(imread(image1)).to(self.device)
-        tensor2 = img2tensor(imread(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}")
-        imwrite(output_path, tensor2img(interpolated.cpu()))
-        logging.debug(f"Saved interpolated image to: {output_path}")
+        return self.model_runner.run(image1, image2)
 
     def scale(self, height: int, width: int) -> float:
         scale = (

src/networks/AMT-S.py

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

src/networks/blocks/feat_enc.py

@@ -1,40 +1,44 @@
+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.conv3 = nn.Conv2d(planes//4, planes, kernel_size=1, padding=0)
+        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.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':
-            self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes//4)
-            self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes//4)
+        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':
-            self.norm1 = nn.BatchNorm2d(planes//4)
-            self.norm2 = nn.BatchNorm2d(planes//4)
+        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':
-            self.norm1 = nn.InstanceNorm2d(planes//4)
-            self.norm2 = nn.InstanceNorm2d(planes//4)
+        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()
@@ -46,8 +50,8 @@ 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
@@ -58,38 +62,40 @@ class BottleneckBlock(nn.Module):
         if self.downsample is not None:
             x = self.downsample(x)
 
-        return self.relu(x+y)
+        return self.relu(x + y)
 
 
 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:
@@ -100,8 +106,8 @@ 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
@@ -111,31 +117,31 @@ class ResidualBlock(nn.Module):
         if self.downsample is not None:
             x = self.downsample(x)
 
-        return self.relu(x+y)
+        return self.relu(x + y)
 
 
 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)
         self.relu1 = nn.ReLU(inplace=True)
 
         self.in_planes = 32
-        self.layer1 = self._make_layer(32,  stride=1)
+        self.layer1 = self._make_layer(32, stride=1)
         self.layer2 = self._make_layer(64, stride=2)
         self.layer3 = self._make_layer(96, stride=2)
 
@@ -147,7 +153,7 @@ 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)
@@ -162,14 +168,15 @@ class SmallEncoder(nn.Module):
         self.in_planes = dim
         return nn.Sequential(*layers)
 
-
-    def forward(self, x):
+    def forward(
+        self, x: torch.Tensor | list[torch.Tensor] | tuple[torch.Tensor, ...]
+    ):
 
         # if input is list, combine batch dimension
-        is_list = isinstance(x, tuple) or isinstance(x, list)
-        if is_list:
+        batch_dim = None
+        if is_list := isinstance(x, tuple) or isinstance(x, list):
             batch_dim = x[0].shape[0]
-            x = torch.cat(x, dim=0)
+            x: torch.Tensor = torch.cat(x, dim=0)
 
         x = self.conv1(x)
         x = self.norm1(x)
@@ -183,33 +190,37 @@ class SmallEncoder(nn.Module):
         if self.training and self.dropout is not None:
             x = self.dropout(x)
 
-        if is_list:
-            x = torch.split(x, [batch_dim, batch_dim], dim=0)
+        if is_list and batch_dim is not None:
+            return 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)
         self.relu1 = nn.ReLU(inplace=True)
 
         self.in_planes = 64
-        self.layer1 = self._make_layer(64,  stride=1)
+        self.layer1 = self._make_layer(64, stride=1)
         self.layer2 = self._make_layer(72, stride=2)
         self.layer3 = self._make_layer(128, stride=2)
 
@@ -222,7 +233,7 @@ 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)
@@ -237,7 +248,6 @@ class BasicEncoder(nn.Module):
         self.in_planes = dim
         return nn.Sequential(*layers)
 
-
     def forward(self, x):
 
         # if input is list, combine batch dimension
@@ -264,21 +274,22 @@ 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)
@@ -299,7 +310,7 @@ 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)
@@ -314,7 +325,6 @@ 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

@@ -4,54 +4,97 @@ import torch.nn.functional as F
 from src.utils.flow_utils import warp
 
 
-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):
-    return nn.Sequential(
-        nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias=bias), 
-        nn.PReLU(out_channels)
+def resize(x: torch.Tensor, scale_factor: float) -> torch.Tensor:
+    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,
+):
+    return nn.Sequential(
+        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)
+            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.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):
         out = self.conv1(x)
 
-        res_feat = out[:, :-self.side_channels, ...]
-        side_feat = out[:, -self.side_channels:, :, :]
+        res_feat = out[:, : -self.side_channels, ...]
+        side_feat = out[:, -self.side_channels :, :, :]
         side_feat = self.conv2(side_feat)
         out = self.conv3(torch.cat([res_feat, side_feat], 1))
 
-        res_feat = out[:, :-self.side_channels, ...]
-        side_feat = out[:, -self.side_channels:, :, :]
+        res_feat = out[:, : -self.side_channels, ...]
+        side_feat = out[:, -self.side_channels :, :, :]
         side_feat = self.conv4(side_feat)
         out = self.conv5(torch.cat([res_feat, side_feat], 1))
 
         out = self.prelu(x + out)
         return out
 
+
 class Encoder(nn.Module):
     def __init__(self, channels, large=False):
         super(Encoder, self).__init__()
@@ -59,30 +102,33 @@ class Encoder(nn.Module):
         prev_ch = 3
         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}', 
-            nn.Sequential(
-                convrelu(prev_ch, ch, k, 2, p),
-                convrelu(ch, ch, 3, 1, 1)
-            ))
+            p = 3 if k == 7 else 1
+            self.register_module(
+                f"pyramid{idx}",
+                nn.Sequential(
+                    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)
+            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),
         )
+
     def forward(self, f0, f1, embt):
         h, w = f0.shape[2:]
         embt = embt.repeat(1, 1, h, w)
@@ -91,14 +137,16 @@ 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)
+            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),
         )
+
     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

@@ -4,15 +4,17 @@ 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, coords, mask=False):
-    """ Wrapper for grid_sample, uses pixel coordinates """
+def bilinear_sampler(img: torch.Tensor, coords: torch.Tensor, mask=False):
+    """Wrapper for grid_sample, uses pixel coordinates"""
     H, W = img.shape[-2:]
-    xgrid, ygrid = coords.split([1,1], dim=-1)
-    xgrid = 2*xgrid/(W-1) - 1
-    ygrid = 2*ygrid/(H-1) - 1
+    xgrid, ygrid = coords.split([1, 1], dim=-1)
+    xgrid = 2 * xgrid / (W - 1) - 1
+    ygrid = 2 * ygrid / (H - 1) - 1
 
     grid = torch.cat([xgrid, ygrid], dim=-1)
     img = F.grid_sample(img, grid, align_corners=True)
@@ -25,27 +27,36 @@ def bilinear_sampler(img, coords, 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
+        cor_planes = corr_levels * (2 * radius + 1) ** 2
         self.scale_factor = scale_factor
 
         self.convc1 = nn.Conv2d(2 * cor_planes, corr_dim, 1, padding=0)
-        self.convf1 = nn.Conv2d(4, flow_dim*2, 7, padding=3)
-        self.convf2 = nn.Conv2d(flow_dim*2, flow_dim, 3, padding=1)
-        self.conv = nn.Conv2d(corr_dim+flow_dim, fc_dim, 3, padding=1)
+        self.convf1 = nn.Conv2d(4, flow_dim * 2, 7, padding=3)
+        self.convf2 = nn.Conv2d(flow_dim * 2, flow_dim, 3, padding=1)
+        self.conv = nn.Conv2d(corr_dim + flow_dim, fc_dim, 3, padding=1)
 
         self.gru = nn.Sequential(
-            nn.Conv2d(fc_dim+4+cdim, hidden_dim, 3, padding=1),
+            nn.Conv2d(fc_dim + 4 + cdim, hidden_dim, 3, padding=1),
             nn.LeakyReLU(negative_slope=0.1, inplace=True),
             nn.Conv2d(hidden_dim, hidden_dim, 3, padding=1),
         )
@@ -65,8 +76,9 @@ 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))
@@ -80,26 +92,39 @@ 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
+        cor_planes = corr_levels * (2 * radius + 1) ** 2
 
         self.scale_factor = scale_factor
         self.convc1 = nn.Conv2d(2 * cor_planes, corr_dim, 1, padding=0)
         self.convc2 = nn.Conv2d(corr_dim, corr_dim2, 3, padding=1)
-        self.convf1 = nn.Conv2d(4, flow_dim*2, 7, padding=3)
-        self.convf2 = nn.Conv2d(flow_dim*2, flow_dim, 3, padding=1)
-        self.conv = nn.Conv2d(flow_dim+corr_dim2, fc_dim, 3, padding=1)
+        self.convf1 = nn.Conv2d(4, flow_dim * 2, 7, padding=3)
+        self.convf2 = nn.Conv2d(flow_dim * 2, flow_dim, 3, padding=1)
+        self.conv = nn.Conv2d(flow_dim + corr_dim2, fc_dim, 3, padding=1)
 
         self.gru = nn.Sequential(
-            nn.Conv2d(fc_dim+4+cdim, hidden_dim, 3, padding=1),
+            nn.Conv2d(fc_dim + 4 + cdim, hidden_dim, 3, padding=1),
             nn.LeakyReLU(negative_slope=0.1, inplace=True),
             nn.Conv2d(hidden_dim, hidden_dim, 3, padding=1),
         )
@@ -113,14 +138,15 @@ class BasicUpdateBlock(nn.Module):
         self.flow_head = nn.Sequential(
             nn.Conv2d(hidden_dim, hidden_dim, 3, padding=1),
             nn.LeakyReLU(negative_slope=0.1, inplace=True),
-            nn.Conv2d(hidden_dim, 4*out_num, 3, padding=1),
+            nn.Conv2d(hidden_dim, 4 * out_num, 3, padding=1),
         )
 
         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))
@@ -135,38 +161,44 @@ 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, fmap2, num_levels=4, radius=4):
+    def __init__(
+        self, fmap1: torch.Tensor, fmap2: torch.Tensor, num_levels=4, radius=4
+    ):
         self.num_levels = num_levels
         self.radius = radius
-        self.corr_pyramid = []
-        self.corr_pyramid_T = []
+        self.corr_pyramid: list[torch.Tensor] = []
+        self.corr_pyramid_T: list[torch.Tensor] = []
 
         corr = BidirCorrBlock.corr(fmap1, fmap2)
         batch, h1, w1, dim, h2, w2 = corr.shape
         corr_T = corr.clone().permute(0, 4, 5, 3, 1, 2)
 
-        corr = corr.reshape(batch*h1*w1, dim, h2, w2)
-        corr_T = corr_T.reshape(batch*h2*w2, dim, h1, w1)
+        corr = corr.reshape(batch * h1 * w1, dim, h2, w2)
+        corr_T = corr_T.reshape(batch * h2 * w2, dim, h1, w1)
 
         self.corr_pyramid.append(corr)
         self.corr_pyramid_T.append(corr_T)
 
-        for _ in range(self.num_levels-1):
+        for _ in range(self.num_levels - 1):
             corr = F.avg_pool2d(corr, 2, stride=2)
             corr_T = F.avg_pool2d(corr_T, 2, stride=2)
             self.corr_pyramid.append(corr)
             self.corr_pyramid_T.append(corr_T)
 
-    def __call__(self, coords0, coords1):
+    def __call__(self, coords0: torch.Tensor, coords1: torch.Tensor):
         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 = []
@@ -175,15 +207,15 @@ class BidirCorrBlock:
             corr = self.corr_pyramid[i]
             corr_T = self.corr_pyramid_T[i]
 
-            dx = torch.linspace(-r, r, 2*r+1, device=coords0.device)
-            dy = torch.linspace(-r, r, 2*r+1, device=coords0.device)
-            delta = torch.stack(torch.meshgrid(dy, dx, indexing='ij'), axis=-1)
-            delta_lvl = delta.view(1, 2*r+1, 2*r+1, 2)
+            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)
 
-            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
+            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
 
             corr = bilinear_sampler(corr, coords_lvl_0)
             corr_T = bilinear_sampler(corr_T, coords_lvl_1)
@@ -194,14 +226,16 @@ 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, fmap2):
+    def corr(fmap1: torch.Tensor, fmap2: torch.Tensor):
         batch, dim, ht, wd = fmap1.shape
-        fmap1 = fmap1.view(batch, dim, ht*wd)
-        fmap2 = fmap2.view(batch, dim, ht*wd) 
+        fmap1 = fmap1.view(batch, dim, ht * wd)
+        fmap2 = fmap2.view(batch, dim, ht * wd)
 
-        corr = torch.matmul(fmap1.transpose(1,2), fmap2)
+        corr = torch.matmul(fmap1.transpose(1, 2), fmap2)
         corr = corr.view(batch, ht, wd, 1, ht, wd)
-        return corr  / torch.sqrt(torch.tensor(dim).float())
+        return corr * (dim**-0.5)

src/utils/torch.py

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

src/utils/video.py

@@ -2,9 +2,10 @@ import os
 import logging
 import subprocess
 from pathlib import Path
+from typing import Generator, Iterable
 
 import cv2
-from typing import Generator
+import numpy as np
 
 
 class VideoMaker:
@@ -35,7 +36,7 @@ class VideoMaker:
         with open(file, "w") as f:
             for video in videos:
                 f.write(f"file '{video}'\n")
-        cmd = f"ffmpeg -f concat -safe 0 -i {file} -c copy {output_path}"
+        cmd = f"ffmpeg -y -f concat -safe 0 -i {file} -c copy {output_path}"
         logging.info(f"Running command: {cmd}")
         result = self.run_command(cmd)
         if result != 0:
@@ -66,7 +67,13 @@ class VideoMaker:
 
     def run_command(self, cmd: str) -> int:
         try:
-            subprocess.run(cmd, shell=True, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+            subprocess.run(
+                cmd,
+                shell=True,
+                check=True,
+                stdout=subprocess.DEVNULL,
+                stderr=subprocess.DEVNULL,
+            )
             return 0
         except subprocess.CalledProcessError as e:
             logging.error(f"Command failed with error: {e}")
@@ -74,7 +81,7 @@ class VideoMaker:
 
     def video_to_frames_generator(
         self, video_path: Path, output_dir: Path, chunk_seconds: int = 10
-    ) -> Generator[tuple[Path, ...], None, None]:
+    ) -> Generator[tuple[np.ndarray, ...], None, None]:
         """Extracts frames from a video and saves them to disk, yielding paths to the saved frames."""
 
         cap = cv2.VideoCapture(str(video_path))
@@ -85,21 +92,56 @@ class VideoMaker:
         fps = cap.get(cv2.CAP_PROP_FPS)
         frames_per_chunk = int(fps * chunk_seconds)
 
-        frame_index = 0
-
         while True:
             paths = []
-
             for _ in range(frames_per_chunk):
                 ret, frame = cap.read()
                 if not ret:
                     cap.release()
                     return
-
-                frame_path = output_dir / f"img_{frame_index:08d}.png"
-                cv2.imwrite(str(frame_path), frame)
-
-                paths.append(frame_path)
-                frame_index += 1
-
+                paths.append(frame)
             yield tuple(paths)
+
+    def images_to_video_pipeline(
+        self,
+        frames: Iterable[np.ndarray],
+        output_path: Path,
+        width: int,
+        height: int,
+        fps: float,
+    ):
+        pipeline = subprocess.Popen(
+            [
+                "ffmpeg",
+                "-y",
+                "-f", "rawvideo",
+                "-vcodec", "rawvideo",
+                "-pix_fmt", "bgr24",
+                "-s", f"{width}x{height}",
+                "-r", str(fps),
+                "-i", "-",
+                "-an",
+                "-vcodec", "libx264",
+                "-pix_fmt", "yuv420p",
+                str(output_path),
+            ],
+            stdin=subprocess.PIPE,
+            stderr=subprocess.DEVNULL
+        )
+        if pipeline.stdin is None:
+            raise Exception("STDIN closed")
+        for frame in frames:
+            pipeline.stdin.write(frame.tobytes())
+
+        pipeline.stdin.close()
+        pipeline.wait()
+
+    def get_size(self, video_path: Path) -> tuple[int, int]:
+        cap = cv2.VideoCapture(str(video_path))
+        if not cap.isOpened():
+            raise ValueError(f"Cannot open video: {video_path}")
+        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        
+        cap.release()
+        return width, height