Open-Sora

mirror of https://github.com/hpcaitech/Open-Sora.git synced 2026-04-11 05:13:31 +02:00

History

Shen-Chenhui 1c3e910fc6 allow loading of discriminator in inference		2024-04-19 14:19:29 +08:00
..
__init__.py
discriminator_3d.py
lpips.py
magvit2.py
model_utils.py
README.md	inference config	2024-04-19 10:47:09 +08:00
vae.py
vae_3d.py
vae_3d_v2.py	allow loading of discriminator in inference	2024-04-19 14:19:29 +08:00

README.md

Commands

1. VAE 3D

1.1 Train

# train on pexel dataset
WANDB_API_KEY=<wandb_api_key> CUDA_VISIBLE_DEVICES=<n> torchrun --master_port=<port_num> --nnodes=1 --nproc_per_node=1 scripts/train-vae.py configs/vae_3d/train/16x256x256.py --data-path /home/shenchenhui/data/pexels/train.csv --wandb True

1.2 Inference

CUDA_VISIBLE_DEVICES=6 torchrun --standalone --nnodes=1 --nproc_per_node=1 scripts/inference-vae.py configs/vae_3d/inference/16x256x256.py --ckpt-path /home/shenchenhui/Open-Sora-dev/outputs/train_pexel_028/epoch3-global_step20000/ --data-path /home/shenchenhui/data/pexels/debug.csv --save-dir outputs/pexel


# resume training debug
CUDA_VISIBLE_DEVICES=5 torchrun --master_port=29530 --nnodes=1 --nproc_per_node=1 scripts/train-vae.py configs/vae_3d/train/16x256x256.py --data-path /home/shenchenhui/data/pexels/debug.csv  --load /home/shenchenhui/Open-Sora-dev/outputs/006-F16S3-VAE_3D_B/epoch49-global_step50

version 2 pipeline

# NOTE: first VAE is pretrained 2D, 16x128x128 --> 16x16x16
# then we train our own temporal VAE, 16x16x16 --> 4x16x16
# we use a 3 layer discriminator on the intermediate of 16x16x16
WANDB_API_KEY=<wandb_api_key> CUDA_VISIBLE_DEVICES=7 torchrun --master_port=29580 --nnodes=1 --nproc_per_node=1 scripts/train-vae-v2.py configs/vae_magvit_v2/train/pipeline_16x128x128.py --data-path /home/shenchenhui/data/trial_data/train_short.csv --wandb True

2. MAGVIT-v2

2.1 dependencies

'accelerate>=0.24.0',
'beartype',
'einops>=0.7.0',
'ema-pytorch>=0.2.4',
'pytorch-warmup',
'gateloop-transformer>=0.2.2',
'kornia',
'opencv-python',
'pillow',
'pytorch-custom-utils>=0.0.9',
'numpy',
'vector-quantize-pytorch>=1.11.8',
'taylor-series-linear-attention>=0.1.5',
'torch',
'torchvision',
'x-transformers'

Note: uses hotfix/zero branch of https://github.com/ver217/ColossalAI.git. clone the repo, go to the branch, then do pip install .

2.2 Train

CUDA_VISIBLE_DEVICES7 torchrun --master_port=29510 --nnodes=1 --nproc_per_node=1 scripts/train-vae-v2.py configs/vae_magvit_v2/train/17x128x128.py --data-path /home/shenchenhui/data/pexels/train.csv