This is the repo used for human motion prediction with non-autoregressive transformers published with our paper

• Pytorch>=1.7.
• Numpy.
• Tensorboard for pytorch.

We have performed experiments with 2 different datasets

1. H36M
2. NTURGB+D (60 actions)

Follow the instructions to download each dataset and place it in data.

Note. You can download the H36M dataset using wget http://www.cs.stanford.edu/people/ashesh/h3.6m.zip. However, the code expects files to be npy files instead of txt. You can use the script in data/h36_convert_txt_to_numpy.py to convert to npy files.

To run training with H3.6M dataset and save experiment results in POTR_OUT folder run the following:

python training/transformer_model_fn.py \
  --model_prefix=${POTR_OUT} \
  --batch_size=16 \
  --data_path=${H36M} \
  --learning_rate=0.0001 \
  --max_epochs=500 \
  --steps_per_epoch=200 \
  --loss_fn=l1 \
  --model_dim=128 \
  --num_encoder_layers=4 \
  --num_decoder_layers=4 \
  --num_heads=4 \
  --dim_ffn=2048 \
  --dropout=0.3 \
  --lr_step_size=400 \
  --learning_rate_fn=step \
  --warmup_epochs=100 \
  --pose_format=rotmat \
  --pose_embedding_type=gcn_enc \
  --dataset=h36m_v2 \
  --pre_normalization \
  --pad_decoder_inputs \
  --non_autoregressive \
  --pos_enc_alpha=10 \
  --pos_enc_beta=500 \
  --predict_activity \
  --action=all

Where pose_embedding_type controls the type of architectures of networks to be used for encoding and decoding skeletons (\phi and \psi in our paper). See models/PoseEncoderDecoder.py for the types of architectures. Tensorboard curves and pytorch models will be saved in ${POTR_OUT}.

If you happen to use the code for your research, please cite the following paper

@inproceedings{Martinez_ICCV_2021,
author = "Mart\'inez-Gonz\'alez, A. and Villamizar, M. and Odobez, J.M.",
title = {Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive Transformers},
booktitle = {IEEE/CVF International Conference on Computer Vision - Workshops (ICCV)},
year = {2021}
}