# VideoMAE: masked video modeling This page is the Transformers contract of the VideoMAE MIC and also a usage example. [doc-builder](https://github.com/huggingface/doc-builder/blob/main/docs/runnable-code-blocks.md) renders the `runnable` blocks below as documentation, and its pytest plugin runs them as written against `MCG-NJU/videomae-base`. The MIC runner runs the same blocks against each slim fixture and each checkpoint in [`integration.json`](../../integration.json). Lines marked `# doc-builder: hide` only matter there: they swap in the pinned checkpoint or variant and print the result record. ## Reconstruct masked video patches ```py runnable:videomae_pretraining import torch from transformers import VideoMAEImageProcessor, VideoMAEForPreTraining, set_seed import json # doc-builder: hide import os # doc-builder: hide import numpy as np model_id = "MCG-NJU/videomae-base" model_id = os.environ.get("MIC_MODEL", model_id) # doc-builder: hide variant = json.loads(os.environ.get("MIC_MODEL_KWARGS", "{}")) # doc-builder: hide device = "cuda" if torch.cuda.is_available() else "cpu" set_seed(0) # seeds torch, NumPy, and Python: loading, dropout, and data augmentation draw random numbers processor = VideoMAEImageProcessor.from_pretrained(model_id) model = VideoMAEForPreTraining.from_pretrained(model_id).to(device) if variant: # doc-builder: hide del model # doc-builder: hide model = VideoMAEForPreTraining.from_pretrained(model_id, **variant).to(device) # doc-builder: hide # Two synthetic clips with moving structure, not i.i.d. noise: a masked-patch model # cannot predict noise, and a constant patch has a zero normalized-pixel target. num_frames = 16 height = width = 224 videos = [] for offset in (0, 64): frames = [] for t in range(num_frames): y, x = np.mgrid[:height, :width] frame = np.empty((height, width, 3), dtype=np.uint8) frame[..., 0] = (x + 2 * y + 16 * t + offset) % 256 frame[..., 1] = (2 * x + y + 8 * t + offset) % 256 frame[..., 2] = (x + y + 32 * t + offset) % 256 frames.append(frame) videos.append(frames) inputs = processor(images=videos, return_tensors="pt") inputs = inputs.to(device, dtype=model.dtype) num_patches_per_frame = (model.config.image_size // model.config.patch_size) ** 2 seq_length = (num_frames // model.config.tubelet_size) * num_patches_per_frame bool_masked_pos = torch.randint(0, 2, (len(videos), seq_length), device=device).bool() with torch.no_grad(): outputs = model(pixel_values=inputs["pixel_values"], bool_masked_pos=bool_masked_pos) # Probe a few reconstructed patches for every batch row. print(outputs.logits[:, -1, :8].float().tolist()) ``` ## One training step ```py runnable:videomae_pretraining:2 model.train() optimizer = torch.optim.SGD(model.parameters(), lr=0.01) loss = model(pixel_values=inputs["pixel_values"], bool_masked_pos=bool_masked_pos).loss loss.backward() before = {n: p.detach().cpu().clone() for n, p in model.named_parameters() if p.requires_grad} # doc-builder: hide optimizer.step() print(f"loss: {loss.item():.4f}") if os.environ.get("MIC_RUNNER"): # doc-builder: hide named = dict(model.named_parameters()) checked = ( "videomae.encoder.layer.0.attention.attention.query.weight", "videomae.encoder.layer.0.intermediate.dense.weight", "decoder.decoder_layers.0.attention.attention.query.weight", ) result = { "runtime": { "device": model.device.type, "dtype": str(model.dtype).removeprefix("torch."), "attn_implementation": model.config._attn_implementation, }, "logits": { "probe": outputs.logits[:, -1, :8].float().tolist(), }, "training": { "loss": loss.item(), "gradients": {n: bool(named[n].grad is not None and torch.count_nonzero(named[n].grad) > 0 and torch.isfinite(named[n].grad).all()) for n in checked}, "updated": any(not torch.equal(p.detach().cpu(), before[n]) for n, p in model.named_parameters() if n in before), }, } print(json.dumps(result)) ```