(ICCV 2023) AttT2M
Code of ICCV 2023 paper: "AttT2M: Text-Driven Human Motion Generation with Multi-Perspective Attention Mechanism"
[Paper] [Bilibili Video]
The pre-train model and train/eval method are Updated. Please see below for more details.
If our paper or code is helpful to you, please cite our paper:
@InProceedings{Zhong_2023_ICCV,
author = {Zhong, Chongyang and Hu, Lei and Zhang, Zihao and Xia, Shihong},
title = {AttT2M: Text-Driven Human Motion Generation with Multi-Perspective Attention Mechanism},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2023},
pages = {509-519}
}
1. Results
1.1 Visual Results
Text-driven motion generation
Compare with SOTA
Generation diversity
Fine-grained generation
1.2 Quantitative Results
For more results, please refer to our [Demo])
2. Installation
2.1. Environment
conda env create -f environment.yml
conda activate Att-T2M
The code was tested on Python 3.8 and PyTorch 1.8.1.
2.2. Datasets and others
We use two dataset: HumanML3D and KIT-ML. For both datasets, the details about them can be found [here].
Motion & text feature extractors are also provided by t2m to evaluate our generated motions
3. Quick Start
1.First step: Download the pre-train models from Google Drive
pretrain_models/
├── HumanML3D/
├── Trans/
├──net_best_fid.pth
├──run.log
├── VQVAE/
├──net_last.pth
├── KIT/
├── Trans/
├──net_last_290000.pth
├──run.log
├── VQVAE/
├──net_last.pth
- Second step:Download other models from Google Drive
3.Third step:run the visualize script:
python vis.py
4. Train
Preparation: you need to download the necessary material from Google Drive:material1, material2
4.1. VQ-VAE
The VAVAE trian parameters are almost the same as T2M GPT
VQ training
```bash
python3 train_vq.py \
--batch-size 256 \
--lr 2e-4 \
--total-iter 300000 \
--lr-scheduler 200000 \
--nb-code 512 \
--down-t 2 \
--depth 3 \
--dilation-growth-rate 3 \
--out-dir output \
--dataname t2m \
--vq-act relu \
--quantizer ema_reset \
--loss-vel 0.5 \
--recons-loss l1_smooth \
--exp-name VQVAE
```
4.2. GPT
The results are saved in the folder output.
GPT training
```bash
python3 train_t2m_trans.py \
--num_layers_cross 2 \
--exp-name GPT \
--batch-size 128 \
--num-layers 9 \
--embed-dim-gpt 1024 \
--nb-code 512 \
--n-head-gpt 16 \
--block-size 51 \
--ff-rate 4 \
--drop-out-rate 0.1 \
--resume-pth output/VQVAE/net_last.pth \
--vq-name VQVAE \
--out-dir output \
--total-iter 300000 \
--lr-scheduler 150000 \
--lr 0.0001 \
--dataname t2m \
--down-t 2 \
--depth 3 \
--quantizer ema_reset \
--eval-iter 10000 \
--pkeep 0.5 \
--dilation-growth-rate 3 \
--vq-act relu
```
5. Evaluation
GPT eval
```bash
python3 GPT_eval_multi.py \
--exp-name TEST_GPT \
--batch-size 128 \
--num-layers 9 \
--num_layers_cross 2 \
--embed-dim-gpt 1024 \
--nb-code 512 \
--n-head-gpt 16 \
--block-size 51 \
--ff-rate 4 \
--drop-out-rate 0.1 \
--resume-pth output/VQVAE/net_last.pth \
--vq-name VQVAE \
--out-dir output \
--total-iter 300000 \
--lr-scheduler 150000 \
--lr 0.0001 \
--dataname t2m \
--down-t 2 \
--depth 3 \
--quantizer ema_reset \
--eval-iter 10000 \
--pkeep 0.5 \
--dilation-growth-rate 3 \
--vq-act relu \
--resume-trans output/GPT/net_best_fid.pth
```
Please repalce "--resume-pth" and "--resume-trans" with the VQVAE and Transformer models you want to evaluate.
The evaluation for multimodality will take a long time. So for a quicker evaluation without multimodality, you can comment out line 452 and line 453 in ./utils/eval_trans.py
6. Acknowledgement