from transformers import AutoImageProcessor, TimesformerForVideoClassification
import numpy as np
import torch

video = list(np.random.randn(16, 3, 448, 448))

processor = AutoImageProcessor.from_pretrained("facebook/timesformer-hr-finetuned-ssv2")
model = TimesformerForVideoClassification.from_pretrained("facebook/timesformer-hr-finetuned-ssv2")

inputs = feature_extractor(images=video, return_tensors="pt")

with torch.no_grad():
  outputs = model(**inputs)
  logits = outputs.logits

predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])

For more code examples, we refer to the documentation.

BibTeX entry and citation info

bibtex

@inproceedings{bertasius2021space,
  title={Is Space-Time Attention All You Need for Video Understanding?},
  author={Bertasius, Gedas and Wang, Heng and Torresani, Lorenzo},
  booktitle={International Conference on Machine Learning},
  pages={813--824},
  year={2021},
  organization={PMLR}
}