---
datasets:
- HuggingFaceM4/vatex
language:
- en
metrics:
- bleu
- meteor
- rouge
pipeline_tag: video-text-to-text
inference: true
tags:
- video-captioning
model-index:
- name: Caelen
  results:
  - task:
      type: video-captioning
    dataset:
      type: video-captioning
      name: VATEX
    metrics:
    - name: CIDEr
      type: image-captioning
      value: 67.3
      verified: false
---
<h1 align='center'> SpaceTimeGPT - Video Captioning Model </h1>

<div align="center">
  <a href="https://github.com/Neleac/SpaceTimeGPT">
    <img src="https://img.shields.io/badge/GitHub-Neleac/SpaceTimeGPT-purple.svg">
  </a>
  <img src="https://raw.githubusercontent.com/Neleac/SpaceTimeGPT/main/model.JPG" width="75%" height="75%">
  <p> (partial diagrams from <a href="https://arxiv.org/abs/2103.15691">1</a>, <a href="https://arxiv.org/abs/2102.05095">2</a>, <a href="https://arxiv.org/abs/1706.03762">3</a>) </p>
</div>

SpaceTimeGPT is a video description generation model capable of spatial and temporal reasoning. Given a video, eight frames are sampled and analyzed by the model. The output is a sentence description of the events that occured in the video, generated using autoregression.

## Architecture and Training
Vision Encoder: [timesformer-base-finetuned-k600](https://huggingface.co/facebook/timesformer-base-finetuned-k600) \
Text Decoder: [gpt2](https://huggingface.co/gpt2)

The encoder and decoder are initialized using pretrained weights for video classification and sentence completion, respectively. Encoder-decoder cross attention is used to unify the visual and linguistic domains. The model is fine-tuned end-to-end on the video captioning task. See [GitHub repository](https://github.com/Neleac/SpaceTimeGPT) for details.

#### Example Inference Code:
```python
import av
import numpy as np
import torch
from transformers import AutoImageProcessor, AutoTokenizer, VisionEncoderDecoderModel

device = "cuda" if torch.cuda.is_available() else "cpu"

# load pretrained processor, tokenizer, and model
image_processor = AutoImageProcessor.from_pretrained("MCG-NJU/videomae-base")
tokenizer = AutoTokenizer.from_pretrained("gpt2")
model = VisionEncoderDecoderModel.from_pretrained("Neleac/timesformer-gpt2-video-captioning").to(device)

# load video
video_path = "never_gonna_give_you_up.mp4"
container = av.open(video_path)

# extract evenly spaced frames from video
seg_len = container.streams.video[0].frames
clip_len = model.config.encoder.num_frames
indices = set(np.linspace(0, seg_len, num=clip_len, endpoint=False).astype(np.int64))
frames = []
container.seek(0)
for i, frame in enumerate(container.decode(video=0)):
    if i in indices:
        frames.append(frame.to_ndarray(format="rgb24"))

# generate caption
gen_kwargs = {
    "min_length": 10, 
    "max_length": 20, 
    "num_beams": 8,
}
pixel_values = image_processor(frames, return_tensors="pt").pixel_values.to(device)
tokens = model.generate(pixel_values, **gen_kwargs)
caption = tokenizer.batch_decode(tokens, skip_special_tokens=True)[0]
print(caption) # A man and a woman are dancing on a stage in front of a mirror.
```

#### Author Information:
👾 [Discord](https://discordapp.com/users/297770280863137802) \
🐙 [GitHub](https://github.com/Neleac) \
🤝 [LinkedIn](https://www.linkedin.com/in/caelenw/)