Попытка оптимизировать модель для более быстрого расчёта

Reviewed-on: #2

.gitignore

@@ -175,5 +175,6 @@ cython_debug/
 .pypirc
 
 
+.DS_Store
 source/
 output/

main.py

@@ -1,239 +1,42 @@
 import logging
-import subprocess
 from pathlib import Path
-
-import cv2
-from tqdm import tqdm
-from time import perf_counter
-from decimal import Decimal
-
-from interpolator import get_device
-from interpolator import ImageInterpolator
-from interpolator import ModelRunner, Anchor
-
-
-logging.basicConfig(
-    level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
-)
-
-from pathlib import Path
-
-
-def move_images(src_dir: str, interpolated_dir: str, output_dir: str):
-    src_dir = Path(src_dir)
-    interpolated_dir = Path(interpolated_dir)
-    output_dir = Path(output_dir)
-    output_dir.mkdir(parents=True, exist_ok=True)
-
-    index = 0
-    src_frames = sorted(src_dir.glob("img_*.png"))
-    interp_frames = sorted(interpolated_dir.glob("img_*.png"))
-    for i in range(len(src_frames)):
-        output_frame = output_dir / f"img_{index:08d}.png"
-        src_frames[i].rename(output_frame)
-        index += 1
-
-        if i < len(interp_frames):
-            output_interp = output_dir / f"img_{index:08d}.png"
-            interp_frames[i].rename(output_interp)
-            index += 1
-
-
-def build_file_list(moved_dir: str, list_path: str):
-    import os
-    moved_dir = Path(moved_dir)
-    frames = sorted(moved_dir.glob("img_*.png"))
-    print(frames[0])
-
-    with open(list_path, "w") as f:
-        for frame in frames:
-            f.write(f"file '{os.path.abspath(frame)}'\n")
-
-
-def build_ffmpeg_file_list(frames_dir: str, interpolated_dir: str, list_path: str):
-    frames = sorted(Path(frames_dir).glob("img_*.png"))
-    interps = sorted(Path(interpolated_dir).glob("img_*.png"))
-
-    if len(interps) != len(frames) - 1:
-        raise ValueError("Interpolated frames must be N-1")
-
-    with open(list_path, "w") as f:
-        for i in range(len(frames)):
-            f.write(f"file '{frames[i].resolve().as_posix()}'\n")
-
-            if i < len(interps):
-                f.write(f"file '{interps[i].resolve().as_posix()}'\n")
-
-
-def merge_with_ffmpeg(
-    original_video: str,
-    file_list: str,
-    output_video: str,
-):
-    cap = cv2.VideoCapture(original_video)
-
-    if not cap.isOpened():
-        raise ValueError("Cannot open original video")
-
-    fps = cap.get(cv2.CAP_PROP_FPS)
-    cap.release()
-
-    new_fps = Decimal(fps * 2)
-
-    cmd = [
-        "ffmpeg",
-        "-y",
-        "-r", str(new_fps.quantize(Decimal("1.0000000000"))),
-        "-f", "concat",
-        "-safe", "0",
-        "-i", file_list,
-        "-c:v", "libx264rgb",
-        output_video,
-    ]
-    print("Running ffmpeg command:", " ".join(cmd))
-
-    subprocess.run(cmd, check=True)
-
-
-
-def video_frames_to_disk_generator(
-    video_path: str | Path,
-    output_dir: str | Path,
-    chunk_seconds: int = 10
-):
-    output_dir = Path(output_dir)
-    output_dir.mkdir(parents=True, exist_ok=True)
-
-    cap = cv2.VideoCapture(str(video_path))
-
-    if not cap.isOpened():
-        raise ValueError(f"Cannot open video: {video_path}")
-
-    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
-
-        yield tuple(paths)
+from src.runner import run
+from src.config import presets
 
 
 def main():
-    start = perf_counter()
-    logging.info("Starting video interpolation process")
-    config_path = Path("src/config/AMT-G.yaml")
-    ckpt_path = Path("src/pretrained/amt-g.pth")
-    video_path = Path("example/video.mp4")
-    output_dir = Path("output/frames")
-    output_interpolated_dir = Path("output/interpolated")
-    output_interpolated_dir.mkdir(parents=True, exist_ok=True)
+    import argparse
 
-    device = get_device()
-    model_runner = ModelRunner(config_path, ckpt_path, device)
-    if device.type in ("cpu", "mps"):
-        if device.type == "mps":
-            logging.warning(
-                "Running on Apple Silicon GPU (MPS) may have limited performance. Consider using a CUDA-enabled GPU for better performance."
-            )
-        else:
-            logging.warning(
-                "Running on CPU may be very slow. Consider using a GPU for better performance."
-            )
-        anchor = Anchor(resolution=8192 * 8192, memory=1, memory_bias=0)
-    elif device.type == "cuda":
-        anchor = Anchor(
-            resolution=1024 * 512, memory=1500 * 1024**2, memory_bias=2500 * 1024**2
-        )
-    else:
-        raise Exception(f"Unsupported device type: {device.type}")
+    logging.basicConfig(
+        level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
+    )
 
-    interpolator = ImageInterpolator(device, anchor, model_runner)
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-b", "--base_path", help="Base path", default="output")
+    parser.add_argument(
+        "-v", "--video_path", help="Video path", default="example/video.mp4"
+    )
+    parser.add_argument(
+        "-o",
+        "--output",
+        help="Output video name (example: 'interpolated_video.mp4')",
+        default="interpolated_video.mp4",
+    )
+    parser.add_argument(
+        "-p",
+        "--preset",
+        help="Model preset",
+        choices=["small", "large", "global"],
+        default="global",
+    )
+    args = parser.parse_args()
+    run(
+        base_path=Path(args.base_path),
+        video_path=Path(args.video_path),
+        output_video=args.output,
+        preset=getattr(presets, args.preset.upper()),
+    )
 
-    loaded_time = perf_counter() - start
-    logging.info(f"Model loaded and initialized in {loaded_time:.2f} seconds")
-    
-    prev_frame_path = None
-    frame_count = 0
-    for frame_paths in video_frames_to_disk_generator(video_path, output_dir):
-        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 = output_interpolated_dir / f"img_{frame_count:08d}.png"
-            interpolator.interpolate(img1, img2, output_path)
-            logging.debug(f"Interpolated image saved to: {output_path}")
-            frame_count += 1
-        for i in tqdm(range(len(frame_paths) - 1), desc="Interpolating frames"):
-            img1 = frame_paths[i]
-            img2 = frame_paths[i + 1]
-            output_path = output_interpolated_dir / f"img_{frame_count:08d}.png"
-            interpolator.interpolate(img1, img2, output_path)
-            logging.debug(f"Interpolated image saved to: {output_path}")
-            frame_count += 1
-        prev_frame_path = frame_paths
-    total_time = perf_counter() - start
-    logging.info(f"Video interpolation completed in {total_time:.2f} seconds")
-
-
-def builder():
-    frames_dir = "output/frames"
-    interpolated_dir = "output/interpolated"
-    moved_dir = "output/moved"
-    video_path = "example/video.mp4"
-    output_video = "output/interpolated_video.mp4"
-    move_images(frames_dir, interpolated_dir, moved_dir)
-
-    cap = cv2.VideoCapture(video_path)
-
-    if not cap.isOpened():
-        raise ValueError("Cannot open original video")
-
-    fps = cap.get(cv2.CAP_PROP_FPS)
-    cmd = [
-        "ffmpeg",
-        "-y",
-        "-framerate", str(fps * 2),
-        "-i", f"{moved_dir}/img_%08d.png",
-        "-i", video_path,
-        "-c:v", "libx264",
-        "-c:a", "copy",
-        "-shortest",
-        output_video,
-    ]
-    logging.info("Running ffmpeg command to build final video: " + " ".join(cmd))
-    subprocess.run(cmd, check=True)
-
-
-def cleanup():
-    import os
-    import shutil
-    frames_dir = "output/frames"
-    interpolated_dir = "output/interpolated"
-    moved_dir = "output/moved"
-    os.makedirs(frames_dir, exist_ok=True)
-    os.makedirs(interpolated_dir, exist_ok=True)
-    os.makedirs(moved_dir, exist_ok=True)
-    shutil.rmtree(frames_dir)
-    shutil.rmtree(interpolated_dir)
-    shutil.rmtree(moved_dir)
 
 if __name__ == "__main__":
-    cleanup()
     main()
-    builder()
-    cleanup()

networks/blocks/multi_flow.py

@@ -1,69 +0,0 @@
-import torch
-import torch.nn as nn
-from src.utils.flow_utils import warp
-from networks.blocks.ifrnet import (
-    convrelu, resize,
-    ResBlock,
-)
-
-
-def multi_flow_combine(comb_block, img0, img1, flow0, flow1, 
-                       mask=None, img_res=None, mean=None):
-        '''
-            A parallel implementation of multiple flow field warping 
-            comb_block: An nn.Seqential object.
-            img shape: [b, c, h, w]
-            flow shape: [b, 2*num_flows, h, w]
-            mask (opt):
-                If 'mask' is None, the function conduct a simple average.
-            img_res (opt):
-                If 'img_res' is None, the function adds zero instead. 
-            mean (opt):
-                If 'mean' is None, the function adds zero instead.       
-        '''
-        b, c, h, w = flow0.shape
-        num_flows = c // 2
-        flow0   =   flow0.reshape(b, num_flows, 2, h, w).reshape(-1, 2, h, w)
-        flow1   =   flow1.reshape(b, num_flows, 2, h, w).reshape(-1, 2, h, w)
-        
-        mask    =    mask.reshape(b, num_flows, 1, h, w
-                            ).reshape(-1, 1, h, w) if mask is not None else None
-        img_res = img_res.reshape(b, num_flows, 3, h, w
-                            ).reshape(-1, 3, h, w)  if img_res is not None else 0
-        img0 = torch.stack([img0] * num_flows, 1).reshape(-1, 3, h, w)
-        img1 = torch.stack([img1] * num_flows, 1).reshape(-1, 3, h, w)
-        mean = torch.stack([mean] * num_flows, 1).reshape(-1, 1, 1, 1
-                                                    ) if mean is not None else 0
-        
-        img0_warp = warp(img0, flow0)
-        img1_warp = warp(img1, flow1)
-        img_warps = mask * img0_warp + (1 - mask) * img1_warp + mean + img_res
-        img_warps = img_warps.reshape(b, num_flows, 3, h, w)
-        imgt_pred = img_warps.mean(1) + comb_block(img_warps.view(b, -1, h, w))
-        return imgt_pred
-
-
-class MultiFlowDecoder(nn.Module):
-    def __init__(self, in_ch, skip_ch, num_flows=3):
-        super(MultiFlowDecoder, self).__init__()
-        self.num_flows = num_flows
-        self.convblock = nn.Sequential(
-            convrelu(in_ch*3+4, in_ch*3), 
-            ResBlock(in_ch*3, skip_ch), 
-            nn.ConvTranspose2d(in_ch*3, 8*num_flows, 4, 2, 1, bias=True)
-        )
-        
-    def forward(self, ft_, f0, f1, flow0, flow1):
-        n = self.num_flows
-        f0_warp = warp(f0, flow0)
-        f1_warp = warp(f1, flow1)
-        out = self.convblock(torch.cat([ft_, f0_warp, f1_warp, flow0, flow1], 1))
-        delta_flow0, delta_flow1, mask, img_res = torch.split(out, [2*n, 2*n, n, 3*n], 1)
-        mask = torch.sigmoid(mask)
-        
-        flow0 = delta_flow0 + 2.0 * resize(flow0, scale_factor=2.0
-                                           ).repeat(1, self.num_flows, 1, 1)
-        flow1 = delta_flow1 + 2.0 * resize(flow1, scale_factor=2.0
-                                           ).repeat(1, self.num_flows, 1, 1)
-        
-        return flow0, flow1, mask, img_res

pyproject.toml

@@ -5,7 +5,6 @@ description = "Add your description here"
 readme = "README.md"
 requires-python = ">=3.12"
 dependencies = [
-    "imageio>=2.37.3",
     "numpy>=2.4.4",
     "omegaconf>=2.3.0",
     "opencv-python>=4.13.0.92",

src/config/AMT-G.yaml

@@ -10,7 +10,7 @@ save_dir: work_dir
 eval_interval: 1
 
 network:
-  name: networks.AMT-G.Model
+  name: AMT-G.Model
   params:
     corr_radius: 3
     corr_lvls: 4

src/config/AMT-L.yaml

@@ -0,0 +1,62 @@
+exp_name: floloss1e-2_300epoch_bs24_lr2e-4
+seed: 2023
+epochs: 300
+distributed: true
+lr: 2e-4
+lr_min: 2e-5
+weight_decay: 0.0
+resume_state: null
+save_dir: work_dir
+eval_interval: 1
+
+network:
+  name: AMT-L.Model
+  params:
+    corr_radius: 3
+    corr_lvls: 4
+    num_flows: 5
+data:
+  train: 
+    name: datasets.vimeo_datasets.Vimeo90K_Train_Dataset
+    params: 
+      dataset_dir: data/vimeo_triplet
+  val:
+    name: datasets.vimeo_datasets.Vimeo90K_Test_Dataset
+    params: 
+      dataset_dir: data/vimeo_triplet
+  train_loader:
+    batch_size: 24
+    num_workers: 12
+  val_loader:
+    batch_size: 24
+    num_workers: 3
+
+logger:
+  use_wandb: true  
+  resume_id: null
+
+losses:
+  - {
+    name: losses.loss.CharbonnierLoss,
+    nickname: l_rec,
+    params: {
+      loss_weight: 1.0,
+      keys: [imgt_pred, imgt]
+    }
+  }
+  - {
+    name: losses.loss.TernaryLoss,
+    nickname: l_ter,
+    params: {
+      loss_weight: 1.0,
+      keys: [imgt_pred, imgt]
+    }
+  }
+  - {
+    name: losses.loss.MultipleFlowLoss,
+    nickname: l_flo,
+    params: {
+      loss_weight: 0.002,
+      keys: [flow0_pred, flow1_pred, flow]
+    }
+  }

src/config/AMT-S.yaml

@@ -10,7 +10,7 @@ save_dir: work_dir
 eval_interval: 1
 
 network:
-  name: networks.AMT-S.Model
+  name: AMT-S.Model
   params:
     corr_radius: 3
     corr_lvls: 4

src/config/presets.py

@@ -0,0 +1,24 @@
+from pathlib import Path
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class Preset:
+    config: Path
+    checkpoint: Path
+
+
+SMALL = Preset(
+    config=Path("src/config/AMT-S.yaml"),
+    checkpoint=Path("src/pretrained/amt-s.pth"),
+)
+
+LARGE = Preset(
+    config=Path("src/config/AMT-L.yaml"),
+    checkpoint=Path("src/pretrained/amt-l.pth"),
+)
+
+GLOBAL = Preset(
+    config=Path("src/config/AMT-G.yaml"),
+    checkpoint=Path("src/pretrained/amt-g.pth"),
+)

src/interpolator.py

@@ -5,10 +5,9 @@ import torch
 import numpy as np
 from omegaconf import OmegaConf, DictConfig
 
-from src.utils import utils
-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
+from .utils.torch import img2tensor, check_dim_and_resize, tensor2img
+from .utils.build import build_from_cfg
+from .utils.padder import InputPadder
 
 
 class Anchor:
@@ -30,6 +29,7 @@ class ModelRunner:
             ckpt_path (Path): Path to model checkpoint in .pth format
             device (torch.device): Device to load the model on
         """
+        torch.set_float32_matmul_precision("high")
         omega_config = OmegaConf.load(config)
         network_config: DictConfig = omega_config.network
         logging.info(
@@ -40,7 +40,7 @@ class ModelRunner:
         model.load_state_dict(checkpoint["state_dict"])
         model = model.to(get_device())
         model.eval()
-        self.model = model
+        self.model = torch.compile(model, mode="max-autotune")
 
 
 def get_vram_available(device: torch.device) -> int:
@@ -77,59 +77,46 @@ class ImageInterpolator:
         self.device = device
         self.anchor = anchor
         self.vram_available = get_vram_available(device)
+        self._scale = None
+        self._padder = None
         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):
-        logging.debug(f"Reading images: {image1} and {image2}")
-        tensor1 = img2tensor(utils.read(image1)).to(self.device)
-        tensor2 = img2tensor(utils.read(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}")
+    def interpolate(self, image1: torch.Tensor, image2: torch.Tensor) -> torch.Tensor:
+        interpolated = self.model_runner.model(
+            image1, image2, self.embt, scale_factor=self._scale, eval=True
+        )["imgt_pred"]
+        if not self._padder:
+            raise NotImplemented("Padder not implemented")
+        return self._padder.unpad(interpolated)[0]
 
+    def make_tensor(self, img: np.ndarray) -> torch.Tensor:
+        tensor = img2tensor(img).to(self.device)
+        h, w = tensor.shape[2], tensor.shape[3]
         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}")
-        utils.write(output_path, tensor2img(interpolated.cpu()))
-        logging.debug(f"Saved interpolated image to: {output_path}")
+        if self._padder is None:
+            self._padder = InputPadder(tensor.shape, padding)
+        return self._padder.pad(tensor)[0]
 
     def scale(self, height: int, width: int) -> float:
-        scale = (
-            self.anchor.resolution
-            / (height * width)
-            * np.sqrt(
-                (self.vram_available - self.anchor.memory_bias) / self.anchor.memory
+        if self._scale is None:
+            scale = (
+                self.anchor.resolution
+                / (height * width)
+                * np.sqrt(
+                    (self.vram_available - self.anchor.memory_bias) / self.anchor.memory
+                )
             )
-        )
-        scale = 1 if scale > 1 else scale
-        scale = 1 / np.floor(1 / np.sqrt(scale) * 16) * 16
-        if scale < 1:
-            logging.info(
-                f"Due to the limited VRAM, the video will be scaled by {scale:.2f}"
-            )
-        return scale
+            scale = 1 if scale > 1 else scale
+            scale = 1 / np.floor(1 / np.sqrt(scale) * 16) * 16
+            if scale < 1:
+                logging.debug(
+                    f"Due to the limited VRAM, the video will be scaled by {scale:.2f}"
+                )
+            self._scale = float(scale)
+            logging.info(f"Calculated scale factor: {self._scale:.2f}")
+        return self._scale

src/networks/AMT_G.py

@@ -1,9 +1,11 @@
+from typing import Optional
+
 import torch
 import torch.nn as nn
-from networks.blocks.raft import coords_grid, BasicUpdateBlock, BidirCorrBlock
-from networks.blocks.feat_enc import LargeEncoder
-from networks.blocks.ifrnet import resize, Encoder, InitDecoder, IntermediateDecoder
-from networks.blocks.multi_flow import multi_flow_combine, MultiFlowDecoder
+from .blocks.raft import coords_grid, BasicUpdateBlock, BidirCorrBlock
+from .blocks.feat_enc import LargeEncoder
+from .blocks.ifrnet import resize, Encoder, InitDecoder, IntermediateDecoder
+from .blocks.multi_flow import multi_flow_combine, MultiFlowDecoder
 
 
 class Model(nn.Module):
@@ -42,7 +44,7 @@ class Model(nn.Module):
             nn.Conv2d(6 * self.num_flows, 3, 7, 1, 3),
         )
 
-    def _get_updateblock(self, cdim, scale_factor=None):
+    def _get_updateblock(self, cdim: int, scale_factor: Optional[float] = None):
         return BasicUpdateBlock(
             cdim=cdim,
             hidden_dim=192,
@@ -55,7 +57,15 @@ class Model(nn.Module):
             radius=self.radius,
         )
 
-    def _corr_scale_lookup(self, corr_fn, coord, flow0, flow1, embt, downsample=1):
+    def _corr_scale_lookup(
+        self,
+        corr_fn: BidirCorrBlock,
+        coord: torch.Tensor,
+        flow0: torch.Tensor,
+        flow1: torch.Tensor,
+        embt: torch.Tensor,
+        downsample: int = 1,
+    ):
         # convert t -> 0 to 0 -> 1 | convert t -> 1 to 1 -> 0
         # based on linear assumption
         t1_scale = 1.0 / embt
@@ -70,7 +80,15 @@ 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: float = 1.0,
+        eval: bool = False,
+        **kwargs,
+    ):
         mean_ = (
             torch.cat([img0, img1], 2)
             .mean(1, keepdim=True)
@@ -159,14 +177,11 @@ class Model(nn.Module):
         )
 
         if scale_factor != 1.0:
-            up_flow0_1 = resize(up_flow0_1, scale_factor=(1.0 / scale_factor)) * (
-                1.0 / scale_factor
-            )
-            up_flow1_1 = resize(up_flow1_1, scale_factor=(1.0 / scale_factor)) * (
-                1.0 / scale_factor
-            )
-            mask = resize(mask, scale_factor=(1.0 / scale_factor))
-            img_res = resize(img_res, scale_factor=(1.0 / scale_factor))
+            factor = 1.0 / scale_factor
+            up_flow0_1 = resize(up_flow0_1, factor) * factor
+            up_flow1_1 = resize(up_flow1_1, factor) * factor
+            mask = resize(mask, factor)
+            img_res = resize(img_res, factor)
 
         # Merge multiple predictions
         imgt_pred = multi_flow_combine(

src/networks/AMT_L.py

@@ -1,38 +1,29 @@
 import torch
 import torch.nn as nn
-from networks.blocks.raft import (
-    coords_grid,
-    BasicUpdateBlock, BidirCorrBlock
-)
-from networks.blocks.feat_enc import (
-    BasicEncoder
-)
-from networks.blocks.ifrnet import (
-    resize,
-    Encoder,
-    InitDecoder,
-    IntermediateDecoder
-)
-from networks.blocks.multi_flow import (
-    multi_flow_combine,
-    MultiFlowDecoder
-)
+from .blocks.raft import coords_grid, BasicUpdateBlock, BidirCorrBlock
+from .blocks.feat_enc import BasicEncoder
+from .blocks.ifrnet import resize, Encoder, InitDecoder, IntermediateDecoder
+
+from .blocks.multi_flow import multi_flow_combine, MultiFlowDecoder
 
 
 class Model(nn.Module):
-    def __init__(self, 
-                 corr_radius=3, 
-                 corr_lvls=4, 
-                 num_flows=5,
-                 channels=[48, 64, 72, 128], 
-                 skip_channels=48
-                 ):
+    def __init__(
+        self,
+        corr_radius=3,
+        corr_lvls=4,
+        num_flows=5,
+        channels=[48, 64, 72, 128],
+        skip_channels=48,
+    ):
         super(Model, self).__init__()
         self.radius = corr_radius
         self.corr_levels = corr_lvls
         self.num_flows = num_flows
 
-        self.feat_encoder = BasicEncoder(output_dim=128, norm_fn='instance', dropout=0.)
+        self.feat_encoder = BasicEncoder(
+            output_dim=128, norm_fn="instance", dropout=0.0
+        )
         self.encoder = Encoder([48, 64, 72, 128], large=True)
 
         self.decoder4 = InitDecoder(channels[3], channels[2], skip_channels)
@@ -45,22 +36,29 @@ class Model(nn.Module):
         self.update2 = self._get_updateblock(48, 4.0)
 
         self.comb_block = nn.Sequential(
-            nn.Conv2d(3*self.num_flows, 6*self.num_flows, 7, 1, 3),
-            nn.PReLU(6*self.num_flows),
-            nn.Conv2d(6*self.num_flows, 3, 7, 1, 3),
+            nn.Conv2d(3 * self.num_flows, 6 * self.num_flows, 7, 1, 3),
+            nn.PReLU(6 * self.num_flows),
+            nn.Conv2d(6 * self.num_flows, 3, 7, 1, 3),
         )
 
     def _get_updateblock(self, cdim, scale_factor=None):
-        return BasicUpdateBlock(cdim=cdim, hidden_dim=128, flow_dim=48, 
-                                corr_dim=256, corr_dim2=160, fc_dim=124, 
-                                scale_factor=scale_factor, corr_levels=self.corr_levels, 
-                                radius=self.radius)
+        return BasicUpdateBlock(
+            cdim=cdim,
+            hidden_dim=128,
+            flow_dim=48,
+            corr_dim=256,
+            corr_dim2=160,
+            fc_dim=124,
+            scale_factor=scale_factor,
+            corr_levels=self.corr_levels,
+            radius=self.radius,
+        )
 
     def _corr_scale_lookup(self, corr_fn, coord, flow0, flow1, embt, downsample=1):
         # convert t -> 0 to 0 -> 1 | convert t -> 1 to 1 -> 0
         # based on linear assumption
-        t1_scale = 1. / embt
-        t0_scale = 1. / (1. - embt)
+        t1_scale = 1.0 / embt
+        t0_scale = 1.0 / (1.0 - embt)
         if downsample != 1:
             inv = 1 / downsample
             flow0 = inv * resize(flow0, scale_factor=inv)
@@ -72,7 +70,12 @@ class Model(nn.Module):
         return corr, flow
 
     def forward(self, img0, img1, embt, scale_factor=1.0, eval=False, **kwargs):
-        mean_ = torch.cat([img0, img1], 2).mean(1, keepdim=True).mean(2, keepdim=True).mean(3, keepdim=True)
+        mean_ = (
+            torch.cat([img0, img1], 2)
+            .mean(1, keepdim=True)
+            .mean(2, keepdim=True)
+            .mean(3, keepdim=True)
+        )
         img0 = img0 - mean_
         img1 = img1 - mean_
         img0_ = resize(img0, scale_factor) if scale_factor != 1.0 else img0
@@ -80,8 +83,10 @@ class Model(nn.Module):
         b, _, h, w = img0_.shape
         coord = coords_grid(b, h // 8, w // 8, img0.device)
 
-        fmap0, fmap1 = self.feat_encoder([img0_, img1_]) # [1, 128, H//8, W//8]
-        corr_fn = BidirCorrBlock(fmap0, fmap1, radius=self.radius, num_levels=self.corr_levels)
+        fmap0, fmap1 = self.feat_encoder([img0_, img1_])  # [1, 128, H//8, W//8]
+        corr_fn = BidirCorrBlock(
+            fmap0, fmap1, radius=self.radius, num_levels=self.corr_levels
+        )
 
         # f0_1: [1, c0, H//2, W//2] | f0_2: [1, c1, H//4, W//4]
         # f0_3: [1, c2, H//8, W//8] | f0_4: [1, c3, H//16, W//16]
@@ -90,9 +95,9 @@ class Model(nn.Module):
 
         ######################################### the 4th decoder #########################################
         up_flow0_4, up_flow1_4, ft_3_ = self.decoder4(f0_4, f1_4, embt)
-        corr_4, flow_4 = self._corr_scale_lookup(corr_fn, coord, 
-                                                 up_flow0_4, up_flow1_4, 
-                                                 embt, downsample=1)
+        corr_4, flow_4 = self._corr_scale_lookup(
+            corr_fn, coord, up_flow0_4, up_flow1_4, embt, downsample=1
+        )
 
         # residue update with lookup corr
         delta_ft_3_, delta_flow_4 = self.update4(ft_3_, flow_4, corr_4)
@@ -102,10 +107,12 @@ class Model(nn.Module):
         ft_3_ = ft_3_ + delta_ft_3_
 
         ######################################### the 3rd decoder #########################################
-        up_flow0_3, up_flow1_3, ft_2_ = self.decoder3(ft_3_, f0_3, f1_3, up_flow0_4, up_flow1_4)
-        corr_3, flow_3 = self._corr_scale_lookup(corr_fn, 
-                                                 coord, up_flow0_3, up_flow1_3, 
-                                                 embt, downsample=2)
+        up_flow0_3, up_flow1_3, ft_2_ = self.decoder3(
+            ft_3_, f0_3, f1_3, up_flow0_4, up_flow1_4
+        )
+        corr_3, flow_3 = self._corr_scale_lookup(
+            corr_fn, coord, up_flow0_3, up_flow1_3, embt, downsample=2
+        )
 
         # residue update with lookup corr
         delta_ft_2_, delta_flow_3 = self.update3(ft_2_, flow_3, corr_3)
@@ -115,10 +122,12 @@ class Model(nn.Module):
         ft_2_ = ft_2_ + delta_ft_2_
 
         ######################################### the 2nd decoder #########################################
-        up_flow0_2, up_flow1_2, ft_1_  = self.decoder2(ft_2_, f0_2, f1_2, up_flow0_3, up_flow1_3)
-        corr_2, flow_2 = self._corr_scale_lookup(corr_fn, 
-                                                 coord, up_flow0_2, up_flow1_2, 
-                                                 embt, downsample=4)
+        up_flow0_2, up_flow1_2, ft_1_ = self.decoder2(
+            ft_2_, f0_2, f1_2, up_flow0_3, up_flow1_3
+        )
+        corr_2, flow_2 = self._corr_scale_lookup(
+            corr_fn, coord, up_flow0_2, up_flow1_2, embt, downsample=4
+        )
 
         # residue update with lookup corr
         delta_ft_1_, delta_flow_2 = self.update2(ft_1_, flow_2, corr_2)
@@ -128,28 +137,33 @@ class Model(nn.Module):
         ft_1_ = ft_1_ + delta_ft_1_
 
         ######################################### the 1st decoder #########################################
-        up_flow0_1, up_flow1_1, mask, img_res = self.decoder1(ft_1_, f0_1, f1_1, up_flow0_2, up_flow1_2)
+        up_flow0_1, up_flow1_1, mask, img_res = self.decoder1(
+            ft_1_, f0_1, f1_1, up_flow0_2, up_flow1_2
+        )
 
         if scale_factor != 1.0:
-            up_flow0_1 = resize(up_flow0_1, scale_factor=(1.0/scale_factor)) * (1.0/scale_factor)
-            up_flow1_1 = resize(up_flow1_1, scale_factor=(1.0/scale_factor)) * (1.0/scale_factor)
-            mask = resize(mask, scale_factor=(1.0/scale_factor))
-            img_res = resize(img_res, scale_factor=(1.0/scale_factor))
+            factor = 1.0 / scale_factor
+            up_flow0_1 = resize(up_flow0_1, factor) * factor
+            up_flow1_1 = resize(up_flow1_1, factor) * factor
+            mask = resize(mask, factor)
+            img_res = resize(img_res, factor)
 
         # Merge multiple predictions
-        imgt_pred = multi_flow_combine(self.comb_block, img0, img1, up_flow0_1, up_flow1_1, 
-                                                                        mask, img_res, mean_)
+        imgt_pred = multi_flow_combine(
+            self.comb_block, img0, img1, up_flow0_1, up_flow1_1, mask, img_res, mean_
+        )
         imgt_pred = torch.clamp(imgt_pred, 0, 1)
 
         if eval:
-            return  { 'imgt_pred': imgt_pred, }
+            return {
+                "imgt_pred": imgt_pred,
+            }
         else:
             up_flow0_1 = up_flow0_1.reshape(b, self.num_flows, 2, h, w)
             up_flow1_1 = up_flow1_1.reshape(b, self.num_flows, 2, h, w)
             return {
-                'imgt_pred': imgt_pred,
-                'flow0_pred': [up_flow0_1, up_flow0_2, up_flow0_3, up_flow0_4],
-                'flow1_pred': [up_flow1_1, up_flow1_2, up_flow1_3, up_flow1_4],
-                'ft_pred': [ft_1_, ft_2_, ft_3_],
+                "imgt_pred": imgt_pred,
+                "flow0_pred": [up_flow0_1, up_flow0_2, up_flow0_3, up_flow0_4],
+                "flow1_pred": [up_flow1_1, up_flow1_2, up_flow1_3, up_flow1_4],
+                "ft_pred": [ft_1_, ft_2_, ft_3_],
             }
-    

src/networks/AMT_S.py

@@ -1,31 +1,20 @@
 import torch
 import torch.nn as nn
-from networks.blocks.raft import (
-    coords_grid,
-    SmallUpdateBlock, BidirCorrBlock
-)
-from networks.blocks.feat_enc import (
-    SmallEncoder
-)
-from networks.blocks.ifrnet import (
-    resize,
-    Encoder,
-    InitDecoder,
-    IntermediateDecoder
-)
-from networks.blocks.multi_flow import (
-    multi_flow_combine,
-    MultiFlowDecoder
-)
+from .blocks.raft import coords_grid, SmallUpdateBlock, BidirCorrBlock
+from .blocks.feat_enc import SmallEncoder
+from .blocks.ifrnet import resize, Encoder, InitDecoder, IntermediateDecoder
+from .blocks.multi_flow import multi_flow_combine, MultiFlowDecoder
 
 
 class Model(nn.Module):
-    def __init__(self, 
-                 corr_radius=3, 
-                 corr_lvls=4, 
-                 num_flows=3, 
-                 channels=[20, 32, 44, 56], 
-                 skip_channels=20):
+    def __init__(
+        self,
+        corr_radius=3,
+        corr_lvls=4,
+        num_flows=3,
+        channels=[20, 32, 44, 56],
+        skip_channels=20,
+    ):
         super(Model, self).__init__()
         self.radius = corr_radius
         self.corr_levels = corr_lvls
@@ -33,7 +22,7 @@ class Model(nn.Module):
         self.channels = channels
         self.skip_channels = skip_channels
 
-        self.feat_encoder = SmallEncoder(output_dim=84, norm_fn='instance', dropout=0.)
+        self.feat_encoder = SmallEncoder(output_dim=84, norm_fn="instance", dropout=0.0)
         self.encoder = Encoder(channels)
 
         self.decoder4 = InitDecoder(channels[3], channels[2], skip_channels)
@@ -46,21 +35,28 @@ class Model(nn.Module):
         self.update2 = self._get_updateblock(20, 4)
 
         self.comb_block = nn.Sequential(
-            nn.Conv2d(3*num_flows, 6*num_flows, 3, 1, 1),
-            nn.PReLU(6*num_flows),
-            nn.Conv2d(6*num_flows, 3, 3, 1, 1),
+            nn.Conv2d(3 * num_flows, 6 * num_flows, 3, 1, 1),
+            nn.PReLU(6 * num_flows),
+            nn.Conv2d(6 * num_flows, 3, 3, 1, 1),
         )
 
     def _get_updateblock(self, cdim, scale_factor=None):
-        return SmallUpdateBlock(cdim=cdim, hidden_dim=76, flow_dim=20, corr_dim=64, 
-                                fc_dim=68, scale_factor=scale_factor, 
-                                corr_levels=self.corr_levels, radius=self.radius)
+        return SmallUpdateBlock(
+            cdim=cdim,
+            hidden_dim=76,
+            flow_dim=20,
+            corr_dim=64,
+            fc_dim=68,
+            scale_factor=scale_factor,
+            corr_levels=self.corr_levels,
+            radius=self.radius,
+        )
 
     def _corr_scale_lookup(self, corr_fn, coord, flow0, flow1, embt, downsample=1):
         # convert t -> 0 to 0 -> 1 | convert t -> 1 to 1 -> 0
         # based on linear assumption
-        t1_scale = 1. / embt
-        t0_scale = 1. / (1. - embt)
+        t1_scale = 1.0 / embt
+        t0_scale = 1.0 / (1.0 - embt)
         if downsample != 1:
             inv = 1 / downsample
             flow0 = inv * resize(flow0, scale_factor=inv)
@@ -72,7 +68,12 @@ class Model(nn.Module):
         return corr, flow
 
     def forward(self, img0, img1, embt, scale_factor=1.0, eval=False, **kwargs):
-        mean_ = torch.cat([img0, img1], 2).mean(1, keepdim=True).mean(2, keepdim=True).mean(3, keepdim=True)
+        mean_ = (
+            torch.cat([img0, img1], 2)
+            .mean(1, keepdim=True)
+            .mean(2, keepdim=True)
+            .mean(3, keepdim=True)
+        )
         img0 = img0 - mean_
         img1 = img1 - mean_
         img0_ = resize(img0, scale_factor) if scale_factor != 1.0 else img0
@@ -80,8 +81,10 @@ class Model(nn.Module):
         b, _, h, w = img0_.shape
         coord = coords_grid(b, h // 8, w // 8, img0.device)
 
-        fmap0, fmap1 = self.feat_encoder([img0_, img1_]) # [1, 128, H//8, W//8]
-        corr_fn = BidirCorrBlock(fmap0, fmap1, radius=self.radius, num_levels=self.corr_levels)
+        fmap0, fmap1 = self.feat_encoder([img0_, img1_])  # [1, 128, H//8, W//8]
+        corr_fn = BidirCorrBlock(
+            fmap0, fmap1, radius=self.radius, num_levels=self.corr_levels
+        )
 
         # f0_1: [1, c0, H//2, W//2] | f0_2: [1, c1, H//4, W//4]
         # f0_3: [1, c2, H//8, W//8] | f0_4: [1, c3, H//16, W//16]
@@ -90,9 +93,9 @@ class Model(nn.Module):
 
         ######################################### the 4th decoder #########################################
         up_flow0_4, up_flow1_4, ft_3_ = self.decoder4(f0_4, f1_4, embt)
-        corr_4, flow_4 = self._corr_scale_lookup(corr_fn, coord, 
-                                                 up_flow0_4, up_flow1_4, 
-                                                 embt, downsample=1)
+        corr_4, flow_4 = self._corr_scale_lookup(
+            corr_fn, coord, up_flow0_4, up_flow1_4, embt, downsample=1
+        )
 
         # residue update with lookup corr
         delta_ft_3_, delta_flow_4 = self.update4(ft_3_, flow_4, corr_4)
@@ -102,10 +105,12 @@ class Model(nn.Module):
         ft_3_ = ft_3_ + delta_ft_3_
 
         ######################################### the 3rd decoder #########################################
-        up_flow0_3, up_flow1_3, ft_2_ = self.decoder3(ft_3_, f0_3, f1_3, up_flow0_4, up_flow1_4)
-        corr_3, flow_3 = self._corr_scale_lookup(corr_fn, 
-                                                 coord, up_flow0_3, up_flow1_3, 
-                                                 embt, downsample=2)
+        up_flow0_3, up_flow1_3, ft_2_ = self.decoder3(
+            ft_3_, f0_3, f1_3, up_flow0_4, up_flow1_4
+        )
+        corr_3, flow_3 = self._corr_scale_lookup(
+            corr_fn, coord, up_flow0_3, up_flow1_3, embt, downsample=2
+        )
 
         # residue update with lookup corr
         delta_ft_2_, delta_flow_3 = self.update3(ft_2_, flow_3, corr_3)
@@ -115,10 +120,12 @@ class Model(nn.Module):
         ft_2_ = ft_2_ + delta_ft_2_
 
         ######################################### the 2nd decoder #########################################
-        up_flow0_2, up_flow1_2, ft_1_  = self.decoder2(ft_2_, f0_2, f1_2, up_flow0_3, up_flow1_3)
-        corr_2, flow_2 = self._corr_scale_lookup(corr_fn, 
-                                                 coord, up_flow0_2, up_flow1_2, 
-                                                 embt, downsample=4)
+        up_flow0_2, up_flow1_2, ft_1_ = self.decoder2(
+            ft_2_, f0_2, f1_2, up_flow0_3, up_flow1_3
+        )
+        corr_2, flow_2 = self._corr_scale_lookup(
+            corr_fn, coord, up_flow0_2, up_flow1_2, embt, downsample=4
+        )
 
         # residue update with lookup corr
         delta_ft_1_, delta_flow_2 = self.update2(ft_1_, flow_2, corr_2)
@@ -128,27 +135,33 @@ class Model(nn.Module):
         ft_1_ = ft_1_ + delta_ft_1_
 
         ######################################### the 1st decoder #########################################
-        up_flow0_1, up_flow1_1, mask, img_res = self.decoder1(ft_1_, f0_1, f1_1, up_flow0_2, up_flow1_2)
+        up_flow0_1, up_flow1_1, mask, img_res = self.decoder1(
+            ft_1_, f0_1, f1_1, up_flow0_2, up_flow1_2
+        )
 
         if scale_factor != 1.0:
-            up_flow0_1 = resize(up_flow0_1, scale_factor=(1.0/scale_factor)) * (1.0/scale_factor)
-            up_flow1_1 = resize(up_flow1_1, scale_factor=(1.0/scale_factor)) * (1.0/scale_factor)
-            mask = resize(mask, scale_factor=(1.0/scale_factor))
-            img_res = resize(img_res, scale_factor=(1.0/scale_factor))
+            factor = 1.0 / scale_factor
+            up_flow0_1 = resize(up_flow0_1, factor) * factor
+            up_flow1_1 = resize(up_flow1_1, factor) * factor
+            mask = resize(mask, factor)
+            img_res = resize(img_res, factor)
 
         # Merge multiple predictions
-        imgt_pred = multi_flow_combine(self.comb_block, img0, img1, up_flow0_1, up_flow1_1, 
-                                                                        mask, img_res, mean_)
+        imgt_pred = multi_flow_combine(
+            self.comb_block, img0, img1, up_flow0_1, up_flow1_1, mask, img_res, mean_
+        )
         imgt_pred = torch.clamp(imgt_pred, 0, 1)
 
         if eval:
-            return  { 'imgt_pred': imgt_pred, }
+            return {
+                "imgt_pred": imgt_pred,
+            }
         else:
             up_flow0_1 = up_flow0_1.reshape(b, self.num_flows, 2, h, w)
             up_flow1_1 = up_flow1_1.reshape(b, self.num_flows, 2, h, w)
             return {
-                'imgt_pred': imgt_pred,
-                'flow0_pred': [up_flow0_1, up_flow0_2, up_flow0_3, up_flow0_4],
-                'flow1_pred': [up_flow1_1, up_flow1_2, up_flow1_3, up_flow1_4],
-                'ft_pred': [ft_1_, ft_2_, ft_3_],
+                "imgt_pred": imgt_pred,
+                "flow0_pred": [up_flow0_1, up_flow0_2, up_flow0_3, up_flow0_4],
+                "flow1_pred": [up_flow1_1, up_flow1_2, up_flow1_3, up_flow1_4],
+                "ft_pred": [ft_1_, ft_2_, ft_3_],
             }

src/networks/IFRNet.py

@@ -1,30 +1,23 @@
 import torch
 import torch.nn as nn
-from src.utils.flow_utils import warp
-from networks.blocks.ifrnet import (
-    convrelu, resize,
-    ResBlock, 
-)
+from ..utils.flow_utils import warp
+from .blocks.ifrnet import convrelu, resize, ResBlock
 
 
 class Encoder(nn.Module):
     def __init__(self):
         super(Encoder, self).__init__()
         self.pyramid1 = nn.Sequential(
-            convrelu(3, 32, 3, 2, 1), 
-            convrelu(32, 32, 3, 1, 1)
+            convrelu(3, 32, 3, 2, 1), convrelu(32, 32, 3, 1, 1)
         )
         self.pyramid2 = nn.Sequential(
-            convrelu(32, 48, 3, 2, 1), 
-            convrelu(48, 48, 3, 1, 1)
+            convrelu(32, 48, 3, 2, 1), convrelu(48, 48, 3, 1, 1)
         )
         self.pyramid3 = nn.Sequential(
-            convrelu(48, 72, 3, 2, 1), 
-            convrelu(72, 72, 3, 1, 1)
+            convrelu(48, 72, 3, 2, 1), convrelu(72, 72, 3, 1, 1)
         )
         self.pyramid4 = nn.Sequential(
-            convrelu(72, 96, 3, 2, 1), 
-            convrelu(96, 96, 3, 1, 1)
+            convrelu(72, 96, 3, 2, 1), convrelu(96, 96, 3, 1, 1)
         )
 
     def forward(self, img):
@@ -39,9 +32,9 @@ class Decoder4(nn.Module):
     def __init__(self):
         super(Decoder4, self).__init__()
         self.convblock = nn.Sequential(
-            convrelu(192+1, 192), 
+            convrelu(192 + 1, 192),
             ResBlock(192, 32),
-            nn.ConvTranspose2d(192, 76, 4, 2, 1, bias=True)
+            nn.ConvTranspose2d(192, 76, 4, 2, 1, bias=True),
         )
 
     def forward(self, f0, f1, embt):
@@ -58,7 +51,7 @@ class Decoder3(nn.Module):
         self.convblock = nn.Sequential(
             convrelu(220, 216),
             ResBlock(216, 32),
-            nn.ConvTranspose2d(216, 52, 4, 2, 1, bias=True)
+            nn.ConvTranspose2d(216, 52, 4, 2, 1, bias=True),
         )
 
     def forward(self, ft_, f0, f1, up_flow0, up_flow1):
@@ -75,7 +68,7 @@ class Decoder2(nn.Module):
         self.convblock = nn.Sequential(
             convrelu(148, 144),
             ResBlock(144, 32),
-            nn.ConvTranspose2d(144, 36, 4, 2, 1, bias=True)
+            nn.ConvTranspose2d(144, 36, 4, 2, 1, bias=True),
         )
 
     def forward(self, ft_, f0, f1, up_flow0, up_flow1):
@@ -92,7 +85,7 @@ class Decoder1(nn.Module):
         self.convblock = nn.Sequential(
             convrelu(100, 96),
             ResBlock(96, 32),
-            nn.ConvTranspose2d(96, 8, 4, 2, 1, bias=True)
+            nn.ConvTranspose2d(96, 8, 4, 2, 1, bias=True),
         )
 
     def forward(self, ft_, f0, f1, up_flow0, up_flow1):
@@ -113,7 +106,12 @@ class Model(nn.Module):
         self.decoder1 = Decoder1()
 
     def forward(self, img0, img1, embt, scale_factor=1.0, eval=False, **kwargs):
-        mean_ = torch.cat([img0, img1], 2).mean(1, keepdim=True).mean(2, keepdim=True).mean(3, keepdim=True)
+        mean_ = (
+            torch.cat([img0, img1], 2)
+            .mean(1, keepdim=True)
+            .mean(2, keepdim=True)
+            .mean(3, keepdim=True)
+        )
         img0 = img0 - mean_
         img1 = img1 - mean_
 
@@ -145,10 +143,14 @@ class Model(nn.Module):
         up_res_1 = out1[:, 5:]
 
         if scale_factor != 1.0:
-            up_flow0_1 = resize(up_flow0_1, scale_factor=(1.0/scale_factor)) * (1.0/scale_factor)
-            up_flow1_1 = resize(up_flow1_1, scale_factor=(1.0/scale_factor)) * (1.0/scale_factor)
-            up_mask_1 = resize(up_mask_1, scale_factor=(1.0/scale_factor))
-            up_res_1 = resize(up_res_1, scale_factor=(1.0/scale_factor))
+            up_flow0_1 = resize(up_flow0_1, scale_factor=(1.0 / scale_factor)) * (
+                1.0 / scale_factor
+            )
+            up_flow1_1 = resize(up_flow1_1, scale_factor=(1.0 / scale_factor)) * (
+                1.0 / scale_factor
+            )
+            up_mask_1 = resize(up_mask_1, scale_factor=(1.0 / scale_factor))
+            up_res_1 = resize(up_res_1, scale_factor=(1.0 / scale_factor))
 
         img0_warp = warp(img0, up_flow0_1)
         img1_warp = warp(img1, up_flow1_1)
@@ -157,13 +159,15 @@ class Model(nn.Module):
         imgt_pred = torch.clamp(imgt_pred, 0, 1)
 
         if eval:
-            return  { 'imgt_pred': imgt_pred, }
+            return {
+                "imgt_pred": imgt_pred,
+            }
         else:
             return {
-                'imgt_pred': imgt_pred,
-                'flow0_pred': [up_flow0_1, up_flow0_2, up_flow0_3, up_flow0_4],
-                'flow1_pred': [up_flow1_1, up_flow1_2, up_flow1_3, up_flow1_4],
-                'ft_pred': [ft_1_, ft_2_, ft_3_],
-                'img0_warp': img0_warp,
-                'img1_warp': img1_warp
+                "imgt_pred": imgt_pred,
+                "flow0_pred": [up_flow0_1, up_flow0_2, up_flow0_3, up_flow0_4],
+                "flow1_pred": [up_flow1_1, up_flow1_2, up_flow1_3, up_flow1_4],
+                "ft_pred": [ft_1_, ft_2_, ft_3_],
+                "img0_warp": img0_warp,
+                "img1_warp": img1_warp,
             }

src/networks/blocks/ifrnet.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from src.utils.flow_utils import warp
+from ...utils.flow_utils import warp
 
 
 def resize(x, scale_factor):

src/networks/blocks/multi_flow.py

@@ -0,0 +1,80 @@
+import torch
+import torch.nn as nn
+from ...utils.flow_utils import warp
+from .ifrnet import convrelu, resize, ResBlock
+
+
+def multi_flow_combine(
+    comb_block, img0, img1, flow0, flow1, mask=None, img_res=None, mean=None
+):
+    """
+    A parallel implementation of multiple flow field warping
+    comb_block: An nn.Seqential object.
+    img shape: [b, c, h, w]
+    flow shape: [b, 2*num_flows, h, w]
+    mask (opt):
+        If 'mask' is None, the function conduct a simple average.
+    img_res (opt):
+        If 'img_res' is None, the function adds zero instead.
+    mean (opt):
+        If 'mean' is None, the function adds zero instead.
+    """
+    b, c, h, w = flow0.shape
+    num_flows = c // 2
+    flow0 = flow0.reshape(b, num_flows, 2, h, w).reshape(-1, 2, h, w)
+    flow1 = flow1.reshape(b, num_flows, 2, h, w).reshape(-1, 2, h, w)
+
+    mask = (
+        mask.reshape(b, num_flows, 1, h, w).reshape(-1, 1, h, w)
+        if mask is not None
+        else None
+    )
+    img_res = (
+        img_res.reshape(b, num_flows, 3, h, w).reshape(-1, 3, h, w)
+        if img_res is not None
+        else 0
+    )
+    img0 = torch.stack([img0] * num_flows, 1).reshape(-1, 3, h, w)
+    img1 = torch.stack([img1] * num_flows, 1).reshape(-1, 3, h, w)
+    mean = (
+        torch.stack([mean] * num_flows, 1).reshape(-1, 1, 1, 1)
+        if mean is not None
+        else 0
+    )
+
+    img0_warp = warp(img0, flow0)
+    img1_warp = warp(img1, flow1)
+    img_warps = mask * img0_warp + (1 - mask) * img1_warp + mean + img_res
+    img_warps = img_warps.reshape(b, num_flows, 3, h, w)
+    imgt_pred = img_warps.mean(1) + comb_block(img_warps.view(b, -1, h, w))
+    return imgt_pred
+
+
+class MultiFlowDecoder(nn.Module):
+    def __init__(self, in_ch, skip_ch, num_flows=3):
+        super(MultiFlowDecoder, self).__init__()
+        self.num_flows = num_flows
+        self.convblock = nn.Sequential(
+            convrelu(in_ch * 3 + 4, in_ch * 3),
+            ResBlock(in_ch * 3, skip_ch),
+            nn.ConvTranspose2d(in_ch * 3, 8 * num_flows, 4, 2, 1, bias=True),
+        )
+
+    def forward(self, ft_, f0, f1, flow0, flow1):
+        n = self.num_flows
+        f0_warp = warp(f0, flow0)
+        f1_warp = warp(f1, flow1)
+        out = self.convblock(torch.cat([ft_, f0_warp, f1_warp, flow0, flow1], 1))
+        delta_flow0, delta_flow1, mask, img_res = torch.split(
+            out, [2 * n, 2 * n, n, 3 * n], 1
+        )
+        mask = torch.sigmoid(mask)
+
+        flow0 = delta_flow0 + 2.0 * resize(flow0, scale_factor=2.0).repeat(
+            1, self.num_flows, 1, 1
+        )
+        flow1 = delta_flow1 + 2.0 * resize(flow1, scale_factor=2.0).repeat(
+            1, self.num_flows, 1, 1
+        )
+
+        return flow0, flow1, mask, img_res

src/runner.py

@@ -0,0 +1,179 @@
+import logging
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+from cv2 import imwrite
+import tqdm
+import torch
+
+from .config import presets
+from .utils.fs import FileSystem
+from .utils.video import VideoMaker
+from .utils.torch import tensor2img
+from .interpolator import (
+    ImageInterpolator,
+    Anchor,
+    get_device,
+    get_vram_available,
+    ModelRunner,
+)
+
+
+if TYPE_CHECKING:
+    import torch
+    import numpy as np
+
+
+def performing_warning_message(device: "torch.device"):
+    if device.type in ("cpu", "mps"):
+        if device.type == "mps":
+            logging.warning(
+                "Running on Apple Silicon GPU (MPS) may have limited performance. Consider using a CUDA-enabled GPU for better performance."
+            )
+        else:
+            logging.warning(
+                "Running on CPU may be very slow. Consider using a GPU for better performance."
+            )
+    elif device.type == "cuda":
+        pass
+    else:
+        raise Exception(f"Unsupported device type: {device.type}")
+
+
+def init_fs(base_path: Path) -> FileSystem:
+    fs = FileSystem(base_path)
+    fs.clear_directory(fs.frames_path)
+    fs.clear_directory(fs.interpolated_path)
+    fs.clear_directory(fs.moved_path)
+    fs.clear_directory(fs.video_part_path)
+    return fs
+
+
+def init_video_maker() -> VideoMaker:
+    return VideoMaker()
+
+
+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")
+    return device
+
+
+def init_anchor(device: "torch.device") -> Anchor:
+    if device.type in ("cpu", "mps"):
+        return Anchor(resolution=8192 * 8192, memory=1, memory_bias=0)
+    elif device.type == "cuda":
+        # return Anchor(
+        #     resolution=1024 * 512, memory=1500 * 1024**2, memory_bias=2500 * 1024**2
+        # )
+        return Anchor(
+            resolution=1280 * 720, memory=6500 * 1024**2, memory_bias=7500 * 1024**2
+        )
+    else:
+        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_interpolator(
+    model_runner: ModelRunner, device: "torch.device"
+) -> ImageInterpolator:
+    anchor = init_anchor(device)
+    return ImageInterpolator(device, anchor, model_runner)
+
+
+class InterpolationPipeline:
+    def __init__(
+        self,
+        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(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["np.ndarray", ...] = tuple()
+        interpolated_frames: list["np.ndarray"] = []
+        part = 0
+        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
+        fps = self.video_maker.get_fps(video_path)
+        logging.info(f"Video FPS: {fps}")
+        fps *= 2  # Doubling FPS
+        width, height = self.video_maker.get_size(video_path)
+        with torch.autocast(self.device.type, torch.float16):
+            with torch.no_grad():
+                prev_tensor = None
+                for idx, frames in enumerate(
+                    self.video_maker.video_to_frames_generator(
+                        video_path, self.fs.frames_path, chunk_seconds
+                    )
+                ):
+                    interpolated_frames: list["np.ndarray"] = []
+                    for frame in tqdm.tqdm(frames):
+                        tensor = self.interpolator.make_tensor(frame)
+                        if prev_tensor is None:
+                            prev_tensor = tensor
+                            continue
+                        interpolated_frames.append(
+                            tensor2img(
+                                self.interpolator.interpolate(
+                                    prev_tensor, tensor
+                                )
+                            )
+                        )
+                        prev_tensor = tensor
+
+                    generator = self._frame_generator(frames, interpolated_frames)
+                    part_path = self.fs.video_part_path / f"video_{idx:08d}.mp4"
+                    self.video_maker.images_to_video_pipeline(
+                        generator, part_path, width, height, fps
+                    )
+                self._merge_video_parts(self.fs.output_path / output_video)
+
+    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 _frame_generator(
+        self,
+        source: tuple["np.ndarray", ...],
+        interpolated: list["np.ndarray"],
+    ):
+        if len(source) == len(interpolated):
+            first = interpolated
+            second = source
+        else:
+            first = source
+            second = interpolated
+
+        for i, frame in enumerate(first):
+            yield frame
+            if i < len(second):
+                yield second[i]
+
+
+def run(
+    base_path: Path,
+    video_path: Path,
+    output_video: str,
+    preset: presets.Preset = presets.LARGE,
+):
+    pipeline = InterpolationPipeline(
+        config=preset.config,
+        checkpoint_path=preset.checkpoint,
+        base_path=base_path,
+    )
+    pipeline.run(video_path, output_video)

src/utils/build.py

@@ -1,15 +1,19 @@
+
 from typing import TYPE_CHECKING
-import importlib
+from ..networks import AMT_G, AMT_L, AMT_S
 
 if TYPE_CHECKING:
     from omegaconf import DictConfig
 
 
-def base_build_fn(module: str, cls: str, params: dict):
-    return getattr(importlib.import_module(module, package=None), cls)(**params)
-
-
 def build_from_cfg(config: "DictConfig"):
+    packages = {
+        "AMT-G": AMT_G,
+        "AMT-L": AMT_L,
+        "AMT-S": AMT_S
+    }
+
     module, cls = config["name"].rsplit(".", 1)
     params: dict = config.get("params", {})
-    return base_build_fn(module, cls, params)
+    return getattr(packages[module], cls)(**params)
+

src/utils/fs.py

@@ -0,0 +1,53 @@
+from pathlib import Path
+
+
+class FileSystem:
+    SOURCE_PATH = "source"
+    OUTPUT_PATH = "output"
+    FRAMES_PATH = "frames"
+    INTERPOLATED_PATH = "interpolated"
+    MOVED_PATH = "moved"
+    VIDEO_PART_PATH = "video_parts"
+
+    def __init__(self, base_path: Path):
+        self.base_path = base_path
+        self.base_path.mkdir(parents=True, exist_ok=True)
+
+    def create_directory(self, dir_name: str) -> Path:
+        """Creates a directory under the base path."""
+        dir_path = self.base_path / dir_name
+        dir_path.mkdir(parents=True, exist_ok=True)
+        return dir_path
+
+    def clear_directory(self, dir_path: Path):
+        """Clears all files in the specified directory."""
+        for item in dir_path.iterdir():
+            if item.is_file():
+                item.unlink()
+            elif item.is_dir():
+                self.clear_directory(item)
+                item.rmdir()
+
+    @property
+    def source_path(self) -> Path:
+        return self.create_directory(self.SOURCE_PATH)
+
+    @property
+    def output_path(self) -> Path:
+        return self.create_directory(self.OUTPUT_PATH)
+
+    @property
+    def frames_path(self) -> Path:
+        return self.create_directory(self.FRAMES_PATH)
+
+    @property
+    def interpolated_path(self) -> Path:
+        return self.create_directory(self.INTERPOLATED_PATH)
+
+    @property
+    def moved_path(self) -> Path:
+        return self.create_directory(self.MOVED_PATH)
+
+    @property
+    def video_part_path(self) -> Path:
+        return self.create_directory(self.VIDEO_PART_PATH)

src/utils/padder.py

@@ -21,10 +21,7 @@ class InputPadder:
         ]
 
     def pad(self, *inputs: "torch.Tensor"):
-        if len(inputs) == 1:
-            return F.pad(inputs[0], self._pad, mode="replicate")
-        else:
-            return [F.pad(x, self._pad, mode="replicate") for x in inputs]
+        return [F.pad(x, self._pad, mode="replicate") for x in inputs]
 
     def unpad(self, *inputs: "torch.Tensor"):
         return [self._unpad(x) for x in inputs]

src/utils/utils.py

@@ -1,199 +0,0 @@
-import re
-import sys
-from pathlib import Path
-
-import numpy as np
-from imageio import imread, imwrite
-
-
-def read(file: Path) -> np.ndarray:
-    readers = {
-        ".float3": readFloat,
-        ".flo": readFlow,
-        ".ppm": readImage,
-        ".pgm": readImage,
-        ".png": readImage,
-        ".jpg": readImage,
-        ".pfm": lambda f: readPFM(f)[0],
-    }
-    func = readers.get(file.suffix.lower())
-    if func is None:
-        raise Exception("don't know how to read %s" % file)
-    return func(file)
-
-
-def write(file: Path, data: np.ndarray) -> None:
-    writers = {
-        ".float3": writeFloat,
-        ".flo": writeFlow,
-        ".ppm": writeImage,
-        ".pgm": writeImage,
-        ".png": writeImage,
-        ".jpg": writeImage,
-        ".pfm": writePFM,
-    }
-    func = writers.get(file.suffix.lower())
-    if func is None:
-        raise Exception("don't know how to write %s" % file)
-    return func(file, data)
-
-
-def readPFM(file: Path):
-    data = open(file, "rb")
-
-    color = None
-    width = None
-    height = None
-    scale = None
-    endian = None
-
-    header = data.readline().rstrip()
-    if header.decode("ascii") == "PF":
-        color = True
-    elif header.decode("ascii") == "Pf":
-        color = False
-    else:
-        raise Exception("Not a PFM file.")
-
-    dim_match = re.match(r"^(\d+)\s(\d+)\s$", data.readline().decode("ascii"))
-    if dim_match:
-        width, height = list(map(int, dim_match.groups()))
-    else:
-        raise Exception("Malformed PFM header.")
-
-    scale = float(data.readline().decode("ascii").rstrip())
-    if scale < 0:
-        endian = "<"
-        scale = -scale
-    else:
-        endian = ">"
-
-    result = np.fromfile(data, endian + "f")
-    shape = (height, width, 3) if color else (height, width)
-
-    result = np.reshape(result, shape)
-    result = np.flipud(result)
-    return result, scale
-
-
-def writePFM(file: Path, image: np.ndarray, scale=1):
-    data = open(file, "wb")
-
-    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 = True
-    elif len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1:
-        color = False
-    else:
-        raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.")
-
-    data.write("PF\n" if color else "Pf\n".encode())  # type: ignore
-    data.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
-
-    data.write("%f\n".encode() % scale)
-
-    image.tofile(data)
-
-
-def readFlow(file: Path):
-    if file.suffix.lower() == ".pfm":
-        return readPFM(file)[0][:, :, 0:2]
-
-    f = open(file, "rb")
-
-    header = f.read(4)
-    if header.decode("utf-8") != "PIEH":
-        raise Exception("Flow file header does not contain PIEH")
-
-    width = np.fromfile(f, np.int32, 1).squeeze()
-    height = np.fromfile(f, np.int32, 1).squeeze()
-
-    flow = np.fromfile(f, np.float32, width * height * 2).reshape((height, width, 2))
-
-    return flow.astype(np.float32)
-
-
-def readImage(file: Path):
-    if file.suffix.lower() == ".pfm":
-        data = readPFM(file)[0]
-        if len(data.shape) == 3:
-            return data[:, :, 0:3]
-        else:
-            return data
-    return imread(file)
-
-
-def writeImage(file: Path, data: np.ndarray):
-    if file.suffix.lower() == ".pfm":
-        return writePFM(file, data, 1)
-    return imwrite(file, data)
-
-
-def writeFlow(file: Path, flow: np.ndarray):
-    f = open(file, "wb")
-    f.write("PIEH".encode("utf-8"))
-    np.array([flow.shape[1], flow.shape[0]], dtype=np.int32).tofile(f)
-    flow = flow.astype(np.float32)
-    flow.tofile(f)
-
-
-def readFloat(file: Path):
-    f = open(file, "rb")
-
-    if (f.readline().decode("utf-8")) != "float\n":
-        raise Exception("float file %s did not contain <float> keyword" % file)
-
-    dim = int(f.readline())
-
-    dims = []
-    count = 1
-    for _ in range(0, dim):
-        d = int(f.readline())
-        dims.append(d)
-        count *= d
-
-    dims = list(reversed(dims))
-
-    data = np.fromfile(f, np.float32, count).reshape(dims)
-    if dim > 2:
-        data = np.transpose(data, (2, 1, 0))
-        data = np.transpose(data, (1, 0, 2))
-
-    return data
-
-
-def writeFloat(file: Path, data: np.ndarray):
-    f = open(file, "wb")
-
-    dim = len(data.shape)
-    if dim > 3:
-        raise Exception("bad float file dimension: %d" % dim)
-
-    f.write(("float\n").encode("ascii"))
-    f.write(("%d\n" % dim).encode("ascii"))
-
-    if dim == 1:
-        f.write(("%d\n" % data.shape[0]).encode("ascii"))
-    else:
-        f.write(("%d\n" % data.shape[1]).encode("ascii"))
-        f.write(("%d\n" % data.shape[0]).encode("ascii"))
-        for i in range(2, dim):
-            f.write(("%d\n" % data.shape[i]).encode("ascii"))
-
-    data = data.astype(np.float32)
-    if dim == 2:
-        data.tofile(f)
-
-    else:
-        np.transpose(data, (2, 0, 1)).tofile(f)

src/utils/video.py

@@ -0,0 +1,147 @@
+import os
+import logging
+import subprocess
+from pathlib import Path
+from typing import Generator, Iterable
+
+import cv2
+import numpy as np
+
+
+class VideoMaker:
+    def images_to_video(
+        self,
+        images_path: Path,
+        output_path: Path,
+        fps: float,
+        image_numerator: str = "img_%08d.png",
+    ):
+        """Converts a sequence of images to a video using ffmpeg."""
+        cmd = f"ffmpeg -framerate {fps} -i {images_path / image_numerator} -c:v libx264 -pix_fmt yuv420p {output_path}"
+        logging.info(f"Running command: {cmd}")
+        result = self.run_command(cmd)
+        if result != 0:
+            logging.error(f"Failed to create video. Command returned {result}")
+
+    def concatenate_videos(
+        self,
+        videos_path: Path,
+        output_path: Path,
+        video_numerator: str = "video_%08d.mp4",
+    ):
+        """Concatenates a sequence of videos using ffmpeg."""
+
+        videos = sorted(videos_path.glob("*.mp4"))
+        file = "file.txt"
+        with open(file, "w") as f:
+            for video in videos:
+                f.write(f"file '{video}'\n")
+        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:
+            logging.error(f"Failed to concatenate videos. Command returned {result}")
+        os.remove(file)
+
+    def get_fps(self, video_path: Path) -> float:
+        """Gets the frames per second (FPS) of a video."""
+        cap = cv2.VideoCapture(str(video_path))
+        if not cap.isOpened():
+            raise ValueError(f"Cannot open video: {video_path}")
+        fps = cap.get(cv2.CAP_PROP_FPS)
+        cap.release()
+        logging.debug(f"FPS of video {video_path}: {fps}")
+        return fps
+
+    def get_video_duration(self, video_path: Path) -> float:
+        """Gets the duration of a video in seconds."""
+        cap = cv2.VideoCapture(str(video_path))
+        if not cap.isOpened():
+            raise ValueError(f"Cannot open video: {video_path}")
+        fps = cap.get(cv2.CAP_PROP_FPS)
+        frame_count = cap.get(cv2.CAP_PROP_FRAME_COUNT)
+        cap.release()
+        duration = frame_count / fps
+        logging.debug(f"Duration of video {video_path}: {duration:.2f} seconds")
+        return duration
+
+    def run_command(self, cmd: str) -> int:
+        try:
+            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}")
+            return e.returncode
+
+    def video_to_frames_generator(
+        self, video_path: Path, output_dir: Path, chunk_seconds: int = 10
+    ) -> 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))
+
+        if not cap.isOpened():
+            raise ValueError(f"Cannot open video: {video_path}")
+
+        fps = cap.get(cv2.CAP_PROP_FPS)
+        frames_per_chunk = int(fps * chunk_seconds)
+
+        while True:
+            paths = []
+            for _ in range(frames_per_chunk):
+                ret, frame = cap.read()
+                if not ret:
+                    cap.release()
+                    return
+                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

uv.lock

@@ -1,5 +1,5 @@
 version = 1
-revision = 2
+revision = 3
 requires-python = ">=3.12"
 
 [[package]]
@@ -7,7 +7,6 @@ name = "amt-apple"
 version = "0.1.0"
 source = { virtual = "." }
 dependencies = [
-    { name = "imageio" },
     { name = "numpy" },
     { name = "omegaconf" },
     { name = "opencv-python" },
@@ -17,7 +16,6 @@ dependencies = [
 
 [package.metadata]
 requires-dist = [
-    { name = "imageio", specifier = ">=2.37.3" },
     { name = "numpy", specifier = ">=2.4.4" },
     { name = "omegaconf", specifier = ">=2.3.0" },
     { name = "opencv-python", specifier = ">=4.13.0.92" },
@@ -127,19 +125,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/d5/1f/5f4a3cd9e4440e9d9bc78ad0a91a1c8d46b4d429d5239ebe6793c9fe5c41/fsspec-2026.3.0-py3-none-any.whl", hash = "sha256:d2ceafaad1b3457968ed14efa28798162f1638dbb5d2a6868a2db002a5ee39a4", size = 202595, upload-time = "2026-03-27T19:11:13.595Z" },
 ]
 
-[[package]]
-name = "imageio"
-version = "2.37.3"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "numpy" },
-    { name = "pillow" },
-]
-sdist = { url = "https://files.pythonhosted.org/packages/b1/84/93bcd1300216ea50811cee96873b84a1bebf8d0489ffaf7f2a3756bab866/imageio-2.37.3.tar.gz", hash = "sha256:bbb37efbfc4c400fcd534b367b91fcd66d5da639aaa138034431a1c5e0a41451", size = 389673, upload-time = "2026-03-09T11:31:12.573Z" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/49/fa/391e437a34e55095173dca5f24070d89cbc233ff85bf1c29c93248c6588d/imageio-2.37.3-py3-none-any.whl", hash = "sha256:46f5bb8522cd421c0f5ae104d8268f569d856b29eb1a13b92829d1970f32c9f0", size = 317646, upload-time = "2026-03-09T11:31:10.771Z" },
-]
-
 [[package]]
 name = "jinja2"
 version = "3.1.6"
@@ -474,75 +459,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/e9/a5/1be1516390333ff9be3a9cb648c9f33df79d5096e5884b5df71a588af463/opencv_python-4.13.0.92-cp37-abi3-win_amd64.whl", hash = "sha256:423d934c9fafb91aad38edf26efb46da91ffbc05f3f59c4b0c72e699720706f5", size = 40212062, upload-time = "2026-02-05T07:02:12.724Z" },
 ]
 
-[[package]]
-name = "pillow"
-version = "12.1.1"
-source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/1f/42/5c74462b4fd957fcd7b13b04fb3205ff8349236ea74c7c375766d6c82288/pillow-12.1.1.tar.gz", hash = "sha256:9ad8fa5937ab05218e2b6a4cff30295ad35afd2f83ac592e68c0d871bb0fdbc4", size = 46980264, upload-time = "2026-02-11T04:23:07.146Z" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/07/d3/8df65da0d4df36b094351dce696f2989bec731d4f10e743b1c5f4da4d3bf/pillow-12.1.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:ab323b787d6e18b3d91a72fc99b1a2c28651e4358749842b8f8dfacd28ef2052", size = 5262803, upload-time = "2026-02-11T04:20:47.653Z" },
-    { url = "https://files.pythonhosted.org/packages/d6/71/5026395b290ff404b836e636f51d7297e6c83beceaa87c592718747e670f/pillow-12.1.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:adebb5bee0f0af4909c30db0d890c773d1a92ffe83da908e2e9e720f8edf3984", size = 4657601, upload-time = "2026-02-11T04:20:49.328Z" },
-    { url = "https://files.pythonhosted.org/packages/b1/2e/1001613d941c67442f745aff0f7cc66dd8df9a9c084eb497e6a543ee6f7e/pillow-12.1.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:bb66b7cc26f50977108790e2456b7921e773f23db5630261102233eb355a3b79", size = 6234995, upload-time = "2026-02-11T04:20:51.032Z" },
-    { url = "https://files.pythonhosted.org/packages/07/26/246ab11455b2549b9233dbd44d358d033a2f780fa9007b61a913c5b2d24e/pillow-12.1.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:aee2810642b2898bb187ced9b349e95d2a7272930796e022efaf12e99dccd293", size = 8045012, upload-time = "2026-02-11T04:20:52.882Z" },
-    { url = "https://files.pythonhosted.org/packages/b2/8b/07587069c27be7535ac1fe33874e32de118fbd34e2a73b7f83436a88368c/pillow-12.1.1-cp312-cp312-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a0b1cd6232e2b618adcc54d9882e4e662a089d5768cd188f7c245b4c8c44a397", size = 6349638, upload-time = "2026-02-11T04:20:54.444Z" },
-    { url = "https://files.pythonhosted.org/packages/ff/79/6df7b2ee763d619cda2fb4fea498e5f79d984dae304d45a8999b80d6cf5c/pillow-12.1.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7aac39bcf8d4770d089588a2e1dd111cbaa42df5a94be3114222057d68336bd0", size = 7041540, upload-time = "2026-02-11T04:20:55.97Z" },
-    { url = "https://files.pythonhosted.org/packages/2c/5e/2ba19e7e7236d7529f4d873bdaf317a318896bac289abebd4bb00ef247f0/pillow-12.1.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:ab174cd7d29a62dd139c44bf74b698039328f45cb03b4596c43473a46656b2f3", size = 6462613, upload-time = "2026-02-11T04:20:57.542Z" },
-    { url = "https://files.pythonhosted.org/packages/03/03/31216ec124bb5c3dacd74ce8efff4cc7f52643653bad4825f8f08c697743/pillow-12.1.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:339ffdcb7cbeaa08221cd401d517d4b1fe7a9ed5d400e4a8039719238620ca35", size = 7166745, upload-time = "2026-02-11T04:20:59.196Z" },
-    { url = "https://files.pythonhosted.org/packages/1f/e7/7c4552d80052337eb28653b617eafdef39adfb137c49dd7e831b8dc13bc5/pillow-12.1.1-cp312-cp312-win32.whl", hash = "sha256:5d1f9575a12bed9e9eedd9a4972834b08c97a352bd17955ccdebfeca5913fa0a", size = 6328823, upload-time = "2026-02-11T04:21:01.385Z" },
-    { url = "https://files.pythonhosted.org/packages/3d/17/688626d192d7261bbbf98846fc98995726bddc2c945344b65bec3a29d731/pillow-12.1.1-cp312-cp312-win_amd64.whl", hash = "sha256:21329ec8c96c6e979cd0dfd29406c40c1d52521a90544463057d2aaa937d66a6", size = 7033367, upload-time = "2026-02-11T04:21:03.536Z" },
-    { url = "https://files.pythonhosted.org/packages/ed/fe/a0ef1f73f939b0eca03ee2c108d0043a87468664770612602c63266a43c4/pillow-12.1.1-cp312-cp312-win_arm64.whl", hash = "sha256:af9a332e572978f0218686636610555ae3defd1633597be015ed50289a03c523", size = 2453811, upload-time = "2026-02-11T04:21:05.116Z" },
-    { url = "https://files.pythonhosted.org/packages/d5/11/6db24d4bd7685583caeae54b7009584e38da3c3d4488ed4cd25b439de486/pillow-12.1.1-cp313-cp313-ios_13_0_arm64_iphoneos.whl", hash = "sha256:d242e8ac078781f1de88bf823d70c1a9b3c7950a44cdf4b7c012e22ccbcd8e4e", size = 4062689, upload-time = "2026-02-11T04:21:06.804Z" },
-    { url = "https://files.pythonhosted.org/packages/33/c0/ce6d3b1fe190f0021203e0d9b5b99e57843e345f15f9ef22fcd43842fd21/pillow-12.1.1-cp313-cp313-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:02f84dfad02693676692746df05b89cf25597560db2857363a208e393429f5e9", size = 4138535, upload-time = "2026-02-11T04:21:08.452Z" },
-    { url = "https://files.pythonhosted.org/packages/a0/c6/d5eb6a4fb32a3f9c21a8c7613ec706534ea1cf9f4b3663e99f0d83f6fca8/pillow-12.1.1-cp313-cp313-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:e65498daf4b583091ccbb2556c7000abf0f3349fcd57ef7adc9a84a394ed29f6", size = 3601364, upload-time = "2026-02-11T04:21:10.194Z" },
-    { url = "https://files.pythonhosted.org/packages/14/a1/16c4b823838ba4c9c52c0e6bbda903a3fe5a1bdbf1b8eb4fff7156f3e318/pillow-12.1.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:6c6db3b84c87d48d0088943bf33440e0c42370b99b1c2a7989216f7b42eede60", size = 5262561, upload-time = "2026-02-11T04:21:11.742Z" },
-    { url = "https://files.pythonhosted.org/packages/bb/ad/ad9dc98ff24f485008aa5cdedaf1a219876f6f6c42a4626c08bc4e80b120/pillow-12.1.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:8b7e5304e34942bf62e15184219a7b5ad4ff7f3bb5cca4d984f37df1a0e1aee2", size = 4657460, upload-time = "2026-02-11T04:21:13.786Z" },
-    { url = "https://files.pythonhosted.org/packages/9e/1b/f1a4ea9a895b5732152789326202a82464d5254759fbacae4deea3069334/pillow-12.1.1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:18e5bddd742a44b7e6b1e773ab5db102bd7a94c32555ba656e76d319d19c3850", size = 6232698, upload-time = "2026-02-11T04:21:15.949Z" },
-    { url = "https://files.pythonhosted.org/packages/95/f4/86f51b8745070daf21fd2e5b1fe0eb35d4db9ca26e6d58366562fb56a743/pillow-12.1.1-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:fc44ef1f3de4f45b50ccf9136999d71abb99dca7706bc75d222ed350b9fd2289", size = 8041706, upload-time = "2026-02-11T04:21:17.723Z" },
-    { url = "https://files.pythonhosted.org/packages/29/9b/d6ecd956bb1266dd1045e995cce9b8d77759e740953a1c9aad9502a0461e/pillow-12.1.1-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:5a8eb7ed8d4198bccbd07058416eeec51686b498e784eda166395a23eb99138e", size = 6346621, upload-time = "2026-02-11T04:21:19.547Z" },
-    { url = "https://files.pythonhosted.org/packages/71/24/538bff45bde96535d7d998c6fed1a751c75ac7c53c37c90dc2601b243893/pillow-12.1.1-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:47b94983da0c642de92ced1702c5b6c292a84bd3a8e1d1702ff923f183594717", size = 7038069, upload-time = "2026-02-11T04:21:21.378Z" },
-    { url = "https://files.pythonhosted.org/packages/94/0e/58cb1a6bc48f746bc4cb3adb8cabff73e2742c92b3bf7a220b7cf69b9177/pillow-12.1.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:518a48c2aab7ce596d3bf79d0e275661b846e86e4d0e7dec34712c30fe07f02a", size = 6460040, upload-time = "2026-02-11T04:21:23.148Z" },
-    { url = "https://files.pythonhosted.org/packages/6c/57/9045cb3ff11eeb6c1adce3b2d60d7d299d7b273a2e6c8381a524abfdc474/pillow-12.1.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:a550ae29b95c6dc13cf69e2c9dc5747f814c54eeb2e32d683e5e93af56caa029", size = 7164523, upload-time = "2026-02-11T04:21:25.01Z" },
-    { url = "https://files.pythonhosted.org/packages/73/f2/9be9cb99f2175f0d4dbadd6616ce1bf068ee54a28277ea1bf1fbf729c250/pillow-12.1.1-cp313-cp313-win32.whl", hash = "sha256:a003d7422449f6d1e3a34e3dd4110c22148336918ddbfc6a32581cd54b2e0b2b", size = 6332552, upload-time = "2026-02-11T04:21:27.238Z" },
-    { url = "https://files.pythonhosted.org/packages/3f/eb/b0834ad8b583d7d9d42b80becff092082a1c3c156bb582590fcc973f1c7c/pillow-12.1.1-cp313-cp313-win_amd64.whl", hash = "sha256:344cf1e3dab3be4b1fa08e449323d98a2a3f819ad20f4b22e77a0ede31f0faa1", size = 7040108, upload-time = "2026-02-11T04:21:29.462Z" },
-    { url = "https://files.pythonhosted.org/packages/d5/7d/fc09634e2aabdd0feabaff4a32f4a7d97789223e7c2042fd805ea4b4d2c2/pillow-12.1.1-cp313-cp313-win_arm64.whl", hash = "sha256:5c0dd1636633e7e6a0afe7bf6a51a14992b7f8e60de5789018ebbdfae55b040a", size = 2453712, upload-time = "2026-02-11T04:21:31.072Z" },
-    { url = "https://files.pythonhosted.org/packages/19/2a/b9d62794fc8a0dd14c1943df68347badbd5511103e0d04c035ffe5cf2255/pillow-12.1.1-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:0330d233c1a0ead844fc097a7d16c0abff4c12e856c0b325f231820fee1f39da", size = 5264880, upload-time = "2026-02-11T04:21:32.865Z" },
-    { url = "https://files.pythonhosted.org/packages/26/9d/e03d857d1347fa5ed9247e123fcd2a97b6220e15e9cb73ca0a8d91702c6e/pillow-12.1.1-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:5dae5f21afb91322f2ff791895ddd8889e5e947ff59f71b46041c8ce6db790bc", size = 4660616, upload-time = "2026-02-11T04:21:34.97Z" },
-    { url = "https://files.pythonhosted.org/packages/f7/ec/8a6d22afd02570d30954e043f09c32772bfe143ba9285e2fdb11284952cd/pillow-12.1.1-cp313-cp313t-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:2e0c664be47252947d870ac0d327fea7e63985a08794758aa8af5b6cb6ec0c9c", size = 6269008, upload-time = "2026-02-11T04:21:36.623Z" },
-    { url = "https://files.pythonhosted.org/packages/3d/1d/6d875422c9f28a4a361f495a5f68d9de4a66941dc2c619103ca335fa6446/pillow-12.1.1-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:691ab2ac363b8217f7d31b3497108fb1f50faab2f75dfb03284ec2f217e87bf8", size = 8073226, upload-time = "2026-02-11T04:21:38.585Z" },
-    { url = "https://files.pythonhosted.org/packages/a1/cd/134b0b6ee5eda6dc09e25e24b40fdafe11a520bc725c1d0bbaa5e00bf95b/pillow-12.1.1-cp313-cp313t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:e9e8064fb1cc019296958595f6db671fba95209e3ceb0c4734c9baf97de04b20", size = 6380136, upload-time = "2026-02-11T04:21:40.562Z" },
-    { url = "https://files.pythonhosted.org/packages/7a/a9/7628f013f18f001c1b98d8fffe3452f306a70dc6aba7d931019e0492f45e/pillow-12.1.1-cp313-cp313t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:472a8d7ded663e6162dafdf20015c486a7009483ca671cece7a9279b512fcb13", size = 7067129, upload-time = "2026-02-11T04:21:42.521Z" },
-    { url = "https://files.pythonhosted.org/packages/1e/f8/66ab30a2193b277785601e82ee2d49f68ea575d9637e5e234faaa98efa4c/pillow-12.1.1-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:89b54027a766529136a06cfebeecb3a04900397a3590fd252160b888479517bf", size = 6491807, upload-time = "2026-02-11T04:21:44.22Z" },
-    { url = "https://files.pythonhosted.org/packages/da/0b/a877a6627dc8318fdb84e357c5e1a758c0941ab1ddffdafd231983788579/pillow-12.1.1-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:86172b0831b82ce4f7877f280055892b31179e1576aa00d0df3bb1bbf8c3e524", size = 7190954, upload-time = "2026-02-11T04:21:46.114Z" },
-    { url = "https://files.pythonhosted.org/packages/83/43/6f732ff85743cf746b1361b91665d9f5155e1483817f693f8d57ea93147f/pillow-12.1.1-cp313-cp313t-win32.whl", hash = "sha256:44ce27545b6efcf0fdbdceb31c9a5bdea9333e664cda58a7e674bb74608b3986", size = 6336441, upload-time = "2026-02-11T04:21:48.22Z" },
-    { url = "https://files.pythonhosted.org/packages/3b/44/e865ef3986611bb75bfabdf94a590016ea327833f434558801122979cd0e/pillow-12.1.1-cp313-cp313t-win_amd64.whl", hash = "sha256:a285e3eb7a5a45a2ff504e31f4a8d1b12ef62e84e5411c6804a42197c1cf586c", size = 7045383, upload-time = "2026-02-11T04:21:50.015Z" },
-    { url = "https://files.pythonhosted.org/packages/a8/c6/f4fb24268d0c6908b9f04143697ea18b0379490cb74ba9e8d41b898bd005/pillow-12.1.1-cp313-cp313t-win_arm64.whl", hash = "sha256:cc7d296b5ea4d29e6570dabeaed58d31c3fea35a633a69679fb03d7664f43fb3", size = 2456104, upload-time = "2026-02-11T04:21:51.633Z" },
-    { url = "https://files.pythonhosted.org/packages/03/d0/bebb3ffbf31c5a8e97241476c4cf8b9828954693ce6744b4a2326af3e16b/pillow-12.1.1-cp314-cp314-ios_13_0_arm64_iphoneos.whl", hash = "sha256:417423db963cb4be8bac3fc1204fe61610f6abeed1580a7a2cbb2fbda20f12af", size = 4062652, upload-time = "2026-02-11T04:21:53.19Z" },
-    { url = "https://files.pythonhosted.org/packages/2d/c0/0e16fb0addda4851445c28f8350d8c512f09de27bbb0d6d0bbf8b6709605/pillow-12.1.1-cp314-cp314-ios_13_0_arm64_iphonesimulator.whl", hash = "sha256:b957b71c6b2387610f556a7eb0828afbe40b4a98036fc0d2acfa5a44a0c2036f", size = 4138823, upload-time = "2026-02-11T04:22:03.088Z" },
-    { url = "https://files.pythonhosted.org/packages/6b/fb/6170ec655d6f6bb6630a013dd7cf7bc218423d7b5fa9071bf63dc32175ae/pillow-12.1.1-cp314-cp314-ios_13_0_x86_64_iphonesimulator.whl", hash = "sha256:097690ba1f2efdeb165a20469d59d8bb03c55fb6621eb2041a060ae8ea3e9642", size = 3601143, upload-time = "2026-02-11T04:22:04.909Z" },
-    { url = "https://files.pythonhosted.org/packages/59/04/dc5c3f297510ba9a6837cbb318b87dd2b8f73eb41a43cc63767f65cb599c/pillow-12.1.1-cp314-cp314-macosx_10_15_x86_64.whl", hash = "sha256:2815a87ab27848db0321fb78c7f0b2c8649dee134b7f2b80c6a45c6831d75ccd", size = 5266254, upload-time = "2026-02-11T04:22:07.656Z" },
-    { url = "https://files.pythonhosted.org/packages/05/30/5db1236b0d6313f03ebf97f5e17cda9ca060f524b2fcc875149a8360b21c/pillow-12.1.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:f7ed2c6543bad5a7d5530eb9e78c53132f93dfa44a28492db88b41cdab885202", size = 4657499, upload-time = "2026-02-11T04:22:09.613Z" },
-    { url = "https://files.pythonhosted.org/packages/6f/18/008d2ca0eb612e81968e8be0bbae5051efba24d52debf930126d7eaacbba/pillow-12.1.1-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:652a2c9ccfb556235b2b501a3a7cf3742148cd22e04b5625c5fe057ea3e3191f", size = 6232137, upload-time = "2026-02-11T04:22:11.434Z" },
-    { url = "https://files.pythonhosted.org/packages/70/f1/f14d5b8eeb4b2cd62b9f9f847eb6605f103df89ef619ac68f92f748614ea/pillow-12.1.1-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:d6e4571eedf43af33d0fc233a382a76e849badbccdf1ac438841308652a08e1f", size = 8042721, upload-time = "2026-02-11T04:22:13.321Z" },
-    { url = "https://files.pythonhosted.org/packages/5a/d6/17824509146e4babbdabf04d8171491fa9d776f7061ff6e727522df9bd03/pillow-12.1.1-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:b574c51cf7d5d62e9be37ba446224b59a2da26dc4c1bb2ecbe936a4fb1a7cb7f", size = 6347798, upload-time = "2026-02-11T04:22:15.449Z" },
-    { url = "https://files.pythonhosted.org/packages/d1/ee/c85a38a9ab92037a75615aba572c85ea51e605265036e00c5b67dfafbfe2/pillow-12.1.1-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a37691702ed687799de29a518d63d4682d9016932db66d4e90c345831b02fb4e", size = 7039315, upload-time = "2026-02-11T04:22:17.24Z" },
-    { url = "https://files.pythonhosted.org/packages/ec/f3/bc8ccc6e08a148290d7523bde4d9a0d6c981db34631390dc6e6ec34cacf6/pillow-12.1.1-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:f95c00d5d6700b2b890479664a06e754974848afaae5e21beb4d83c106923fd0", size = 6462360, upload-time = "2026-02-11T04:22:19.111Z" },
-    { url = "https://files.pythonhosted.org/packages/f6/ab/69a42656adb1d0665ab051eec58a41f169ad295cf81ad45406963105408f/pillow-12.1.1-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:559b38da23606e68681337ad74622c4dbba02254fc9cb4488a305dd5975c7eeb", size = 7165438, upload-time = "2026-02-11T04:22:21.041Z" },
-    { url = "https://files.pythonhosted.org/packages/02/46/81f7aa8941873f0f01d4b55cc543b0a3d03ec2ee30d617a0448bf6bd6dec/pillow-12.1.1-cp314-cp314-win32.whl", hash = "sha256:03edcc34d688572014ff223c125a3f77fb08091e4607e7745002fc214070b35f", size = 6431503, upload-time = "2026-02-11T04:22:22.833Z" },
-    { url = "https://files.pythonhosted.org/packages/40/72/4c245f7d1044b67affc7f134a09ea619d4895333d35322b775b928180044/pillow-12.1.1-cp314-cp314-win_amd64.whl", hash = "sha256:50480dcd74fa63b8e78235957d302d98d98d82ccbfac4c7e12108ba9ecbdba15", size = 7176748, upload-time = "2026-02-11T04:22:24.64Z" },
-    { url = "https://files.pythonhosted.org/packages/e4/ad/8a87bdbe038c5c698736e3348af5c2194ffb872ea52f11894c95f9305435/pillow-12.1.1-cp314-cp314-win_arm64.whl", hash = "sha256:5cb1785d97b0c3d1d1a16bc1d710c4a0049daefc4935f3a8f31f827f4d3d2e7f", size = 2544314, upload-time = "2026-02-11T04:22:26.685Z" },
-    { url = "https://files.pythonhosted.org/packages/6c/9d/efd18493f9de13b87ede7c47e69184b9e859e4427225ea962e32e56a49bc/pillow-12.1.1-cp314-cp314t-macosx_10_15_x86_64.whl", hash = "sha256:1f90cff8aa76835cba5769f0b3121a22bd4eb9e6884cfe338216e557a9a548b8", size = 5268612, upload-time = "2026-02-11T04:22:29.884Z" },
-    { url = "https://files.pythonhosted.org/packages/f8/f1/4f42eb2b388eb2ffc660dcb7f7b556c1015c53ebd5f7f754965ef997585b/pillow-12.1.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:1f1be78ce9466a7ee64bfda57bdba0f7cc499d9794d518b854816c41bf0aa4e9", size = 4660567, upload-time = "2026-02-11T04:22:31.799Z" },
-    { url = "https://files.pythonhosted.org/packages/01/54/df6ef130fa43e4b82e32624a7b821a2be1c5653a5fdad8469687a7db4e00/pillow-12.1.1-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:42fc1f4677106188ad9a55562bbade416f8b55456f522430fadab3cef7cd4e60", size = 6269951, upload-time = "2026-02-11T04:22:33.921Z" },
-    { url = "https://files.pythonhosted.org/packages/a9/48/618752d06cc44bb4aae8ce0cd4e6426871929ed7b46215638088270d9b34/pillow-12.1.1-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:98edb152429ab62a1818039744d8fbb3ccab98a7c29fc3d5fcef158f3f1f68b7", size = 8074769, upload-time = "2026-02-11T04:22:35.877Z" },
-    { url = "https://files.pythonhosted.org/packages/c3/bd/f1d71eb39a72fa088d938655afba3e00b38018d052752f435838961127d8/pillow-12.1.1-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:d470ab1178551dd17fdba0fef463359c41aaa613cdcd7ff8373f54be629f9f8f", size = 6381358, upload-time = "2026-02-11T04:22:37.698Z" },
-    { url = "https://files.pythonhosted.org/packages/64/ef/c784e20b96674ed36a5af839305f55616f8b4f8aa8eeccf8531a6e312243/pillow-12.1.1-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6408a7b064595afcab0a49393a413732a35788f2a5092fdc6266952ed67de586", size = 7068558, upload-time = "2026-02-11T04:22:39.597Z" },
-    { url = "https://files.pythonhosted.org/packages/73/cb/8059688b74422ae61278202c4e1ad992e8a2e7375227be0a21c6b87ca8d5/pillow-12.1.1-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:5d8c41325b382c07799a3682c1c258469ea2ff97103c53717b7893862d0c98ce", size = 6493028, upload-time = "2026-02-11T04:22:42.73Z" },
-    { url = "https://files.pythonhosted.org/packages/c6/da/e3c008ed7d2dd1f905b15949325934510b9d1931e5df999bb15972756818/pillow-12.1.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:c7697918b5be27424e9ce568193efd13d925c4481dd364e43f5dff72d33e10f8", size = 7191940, upload-time = "2026-02-11T04:22:44.543Z" },
-    { url = "https://files.pythonhosted.org/packages/01/4a/9202e8d11714c1fc5951f2e1ef362f2d7fbc595e1f6717971d5dd750e969/pillow-12.1.1-cp314-cp314t-win32.whl", hash = "sha256:d2912fd8114fc5545aa3a4b5576512f64c55a03f3ebcca4c10194d593d43ea36", size = 6438736, upload-time = "2026-02-11T04:22:46.347Z" },
-    { url = "https://files.pythonhosted.org/packages/f3/ca/cbce2327eb9885476b3957b2e82eb12c866a8b16ad77392864ad601022ce/pillow-12.1.1-cp314-cp314t-win_amd64.whl", hash = "sha256:4ceb838d4bd9dab43e06c363cab2eebf63846d6a4aeaea283bbdfd8f1a8ed58b", size = 7182894, upload-time = "2026-02-11T04:22:48.114Z" },
-    { url = "https://files.pythonhosted.org/packages/ec/d2/de599c95ba0a973b94410477f8bf0b6f0b5e67360eb89bcb1ad365258beb/pillow-12.1.1-cp314-cp314t-win_arm64.whl", hash = "sha256:7b03048319bfc6170e93bd60728a1af51d3dd7704935feb228c4d4faab35d334", size = 2546446, upload-time = "2026-02-11T04:22:50.342Z" },
-]
-
 [[package]]
 name = "pyyaml"
 version = "6.0.3"