Spaces:
Sleeping
Sleeping
Elle McFarlane
commited on
Commit
·
e142172
1
Parent(s):
91a6b16
add readme
Browse files- text2motion/README.md +110 -0
text2motion/README.md
ADDED
@@ -0,0 +1,110 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Text-driven Motion Generation
|
2 |
+
|
3 |
+
<!-- TOC -->
|
4 |
+
|
5 |
+
- [Installation](#installation)
|
6 |
+
- [Training](#prepare-environment)
|
7 |
+
- [Acknowledgement](#acknowledgement)
|
8 |
+
|
9 |
+
<!-- TOC -->
|
10 |
+
|
11 |
+
## Installation
|
12 |
+
|
13 |
+
Please refer to [install.md](install.md) for detailed installation.
|
14 |
+
|
15 |
+
## Training
|
16 |
+
|
17 |
+
Due to the requirement of a large batchsize, we highly recommend you to use DDP training. A slurm-based script is as below:
|
18 |
+
|
19 |
+
```shell
|
20 |
+
PYTHONPATH="$(dirname $0)/..":$PYTHONPATH \
|
21 |
+
srun -p ${PARTITION} -n8 --gres=gpu:8 -u \
|
22 |
+
python -u tools/train.py \
|
23 |
+
--name kit_baseline_ddp_8gpu_8layers_1000 \
|
24 |
+
--batch_size 128 \
|
25 |
+
--times 200 \
|
26 |
+
--num_epochs 50 \
|
27 |
+
--dataset_name kit \
|
28 |
+
--distributed
|
29 |
+
```
|
30 |
+
|
31 |
+
Besides, you can train the model on multi-GPUs with DataParallel:
|
32 |
+
|
33 |
+
```shell
|
34 |
+
PYTHONPATH="$(dirname $0)/..":$PYTHONPATH \
|
35 |
+
python -u tools/train.py \
|
36 |
+
--name kit_baseline_dp_2gpu_8layers_1000 \
|
37 |
+
--batch_size 128 \
|
38 |
+
--times 50 \
|
39 |
+
--num_epochs 50 \
|
40 |
+
--dataset_name kit \
|
41 |
+
--num_layers 8 \
|
42 |
+
--diffusion_steps 1000 \
|
43 |
+
--data_parallel \
|
44 |
+
--gpu_id 0 1
|
45 |
+
```
|
46 |
+
|
47 |
+
Otherwise, you can run the training code on a single GPU like:
|
48 |
+
|
49 |
+
```shell
|
50 |
+
PYTHONPATH="$(dirname $0)/..":$PYTHONPATH \
|
51 |
+
python -u tools/train.py \
|
52 |
+
--name kit_baseline_1gpu_8layers_1000 \
|
53 |
+
--batch_size 128 \
|
54 |
+
--times 25 \
|
55 |
+
--num_epochs 50 \
|
56 |
+
--dataset_name kit
|
57 |
+
```
|
58 |
+
|
59 |
+
Here, `times` means the duplication times of the original dataset. To retain the number of iterations, you can set `times` to 25 for 1 GPU, 50 for 2 GPUs, 100 for 4 GPUs, and 200 for 8 GPUs.
|
60 |
+
|
61 |
+
## Evaluation
|
62 |
+
|
63 |
+
```shell
|
64 |
+
# GPU_ID indicates which gpu you want to use
|
65 |
+
python -u tools/evaluation.py checkpoints/kit/kit_motiondiffuse/opt.txt GPU_ID
|
66 |
+
# Or you can omit this option and use cpu for evaluation
|
67 |
+
python -u tools/evaluation.py checkpoints/kit/kit_motiondiffuse/opt.txt
|
68 |
+
```
|
69 |
+
|
70 |
+
## Visualization
|
71 |
+
|
72 |
+
You can visualize human motion with the given language description and the expected motion length. We also provide a [Colab Demo](https://colab.research.google.com/drive/1Dp6VsZp2ozKuu9ccMmsDjyij_vXfCYb3?usp=sharing) and a [Hugging Face Demo](https://huggingface.co/spaces/mingyuan/MotionDiffuse) for your convenience.
|
73 |
+
|
74 |
+
```shell
|
75 |
+
# Currently we only support visualization of models trained on the HumanML3D dataset.
|
76 |
+
# Motion length can not be larger than 196, which is the maximum length during training
|
77 |
+
# You can omit `gpu_id` to run visualization on your CPU
|
78 |
+
# Optionally, you can store the xyz coordinates of each joint to `npy_path`. The shape of motion data is (T, 22, 3), where T denotes the motion length, 22 is the number of joints.
|
79 |
+
|
80 |
+
python -u tools/visualization.py \
|
81 |
+
--opt_path checkpoints/t2m/t2m_motiondiffuse/opt.txt \
|
82 |
+
--text "a person is jumping" \
|
83 |
+
--motion_length 60 \
|
84 |
+
--result_path "test_sample.gif" \
|
85 |
+
--npy_path "test_sample.npy" \
|
86 |
+
--gpu_id 0
|
87 |
+
```
|
88 |
+
|
89 |
+
Here are some visualization examples. The motion lengths are shown in the title of animations.
|
90 |
+
|
91 |
+
<table>
|
92 |
+
<tr>
|
93 |
+
<td><img src="../figures/gallery_t2m/gen_00.gif" width="100%"/></td>
|
94 |
+
<td><img src="../figures/gallery_t2m/gen_01.gif" width="100%"/></td>
|
95 |
+
<td><img src="../figures/gallery_t2m/gen_02.gif" width="100%"/></td>
|
96 |
+
<td><img src="../figures/gallery_t2m/gen_03.gif" width="100%"/></td>
|
97 |
+
</tr>
|
98 |
+
<tr>
|
99 |
+
<td><img src="../figures/gallery_t2m/gen_04.gif" width="100%"/></td>
|
100 |
+
<td><img src="../figures/gallery_t2m/gen_05.gif" width="100%"/></td>
|
101 |
+
<td><img src="../figures/gallery_t2m/gen_06.gif" width="100%"/></td>
|
102 |
+
<td><img src="../figures/gallery_t2m/gen_07.gif" width="100%"/></td>
|
103 |
+
</tr>
|
104 |
+
</table>
|
105 |
+
|
106 |
+
**Note:** You may install `matplotlib==3.3.1` to support visualization here.
|
107 |
+
|
108 |
+
## Acknowledgement
|
109 |
+
|
110 |
+
This code is developed on top of [Generating Diverse and Natural 3D Human Motions from Text](https://github.com/EricGuo5513/text-to-motion)
|