(args, model, device)
| 97 | |
| 98 | |
| 99 | def process_sequence(args, model, device): |
| 100 | print(f"Loading frames from {args.input}...") |
| 101 | |
| 102 | native_size = None |
| 103 | input_video_fps = None |
| 104 | input_image = [] |
| 105 | |
| 106 | for frame, orig_shape, extracted_fps in get_frames(args.input, args.fix_width): |
| 107 | if native_size is None: |
| 108 | native_size = orig_shape |
| 109 | |
| 110 | # Capture the FPS from the first yielded frame |
| 111 | if input_video_fps is None and extracted_fps is not None: |
| 112 | input_video_fps = extracted_fps |
| 113 | |
| 114 | img_t = torch.from_numpy(frame).permute(2, 0, 1).float() |
| 115 | input_image.append(img_t) |
| 116 | |
| 117 | if len(input_image) < 2: |
| 118 | print("Sequence too short (requires at least 2 frames).") |
| 119 | return |
| 120 | |
| 121 | input_scene = torch.stack(input_image, dim=0)[None] |
| 122 | B, T, C, H, W = input_scene.shape |
| 123 | |
| 124 | is_video_out = False |
| 125 | video_writer = None |
| 126 | out_dir = None |
| 127 | |
| 128 | if args.output.lower().endswith(('.mp4', '.avi', '.mkv', '.webm')): |
| 129 | is_video_out = True |
| 130 | os.makedirs(os.path.dirname(os.path.abspath(args.output)) or ".", exist_ok=True) |
| 131 | |
| 132 | # Prioritize native video FPS, fallback to args.fps for image folders |
| 133 | final_fps = input_video_fps if input_video_fps is not None else args.fps |
| 134 | print(f"Encoding output video at {final_fps:.2f} FPS") |
| 135 | |
| 136 | video_writer = imageio.get_writer(args.output, fps=final_fps, codec='libx264', macro_block_size=None) |
| 137 | else: |
| 138 | out_dir = args.output |
| 139 | os.makedirs(out_dir, exist_ok=True) |
| 140 | |
| 141 | print(f"Processing sequence of {T} frames via window size {args.window_size}...") |
| 142 | infer_window = args.window_size |
| 143 | |
| 144 | for start in tqdm(range(0, T - 1, infer_window - 1)): |
| 145 | end = min(start + infer_window, T) |
| 146 | chunk = input_scene[:, start:end].to(device) |
| 147 | |
| 148 | compute_dtype = torch.bfloat16 if device == "cuda" and torch.cuda.is_bf16_supported() else torch.float16 |
| 149 | with torch.autocast(device_type=device, dtype=compute_dtype, enabled=(device == "cuda")): |
| 150 | results_dict = model(chunk, num_reg_refine=args.iters) |
| 151 | |
| 152 | flow_pr = results_dict['flow_preds'][-1] |
| 153 | |
| 154 | if native_size is not None and getattr(args, 'restore_size', False): |
| 155 | scaled_flow = F.interpolate(flow_pr.view(-1, 2, H, W), size=native_size, mode='bilinear', align_corners=True) |
| 156 |
no test coverage detected