VPTR: Efficient Transformers for Video Prediction

Video future frames prediction based on Transformers. Accepted by ICPR2022, http://arxiv.org/abs/2203.15836

The overall framework for video prediction.

Fully autoregressive (left) and non-autoregressive VPTR (right).

Pretrained-models

Download the checkpoints from here: https://polymtlca0-my.sharepoint.com/:f:/g/personal/xi_ye_polymtl_ca/EuxjSddJ7wNIsiSTOfB-u7AB7qQhP5H0iX2a5mbaowiSZw?e=IEj1bd

See Test_AutoEncoder.ipynb and Test_VPTR.ipynb for the detatiled test functions.

Training

Stage 1: train_AutoEncoder.py

Train the autoencoder firstly, save the ckpt, load it for stage 2

Stage 2: Train Transformer for the video prediction

train_FAR.py: Fully autoregressive model
train_FAR_mp.py: multiple gpu training (single machine)
train_NAR.py: Non-autoregressive model
train_NAR_mp.py: multiple gpu training (single machine)

Dataset folder structure

/MovingMNIST
     moving-mnist-train.npz
     moving-mnist-test.npz
     moving-mnist-val.npz

/KTH
     boxing/
         person01_boxing_d1/
            image_0001.png
            image_0002.png
            ...
         person01_boxing_d2/
            image_0001.png
            image_0002.png
            ...

     handclapping/
         ...
     handwaving/
         ...
     jogging_no_empty/
         ...
     running_no_empty/
         ...
     walking_no_empty/
         ...

/BAIR
     test/
         example_0/
            0000.png
            0001.png
            ...
         example_1/
            0000.png
            0001.png
            ...
         example_...
     train/
         example_0/
            0000.png
            0001.png
            ...
         example_...

Citing

Please cite the paper if you find our work is helpful.

@article{ye_2022,
       title = {VPTR: Efficient Transformers for Video Prediction},
       author = {Xi Ye and Guillaume-Alexandre Bilodeau},
       journal={arXiv preprint arXiv:2203.15836},
       year={2022}
} 

Correction about the paper

Recently, we found a mistake in our ICPR paper. For the BAIR experiments, the previous papers predict 28 future frames instead of 10. Specifically, the results in "TABLE II: Results on BAIR" are for 10 future frames instead of 28. The results for 28 predicted frames are updated here, see the following correct table.

We apologize for the mistake, the correction does not affect our conclusions.