/rcg

PyTorch implementation of RCG https://arxiv.org/abs/2312.03701

Primary LanguagePythonMIT LicenseMIT

RCG PyTorch Implementation

This is a PyTorch/GPU implementation of the paper Return of Unconditional Generation: A Self-supervised Representation Generation Method:

@Article{RCG2023,
  author  = {Tianhong Li and Dina Katabi and Kaiming He},
  journal = {arXiv:2312.03701},
  title   = {Return of Unconditional Generation: A Self-supervised Representation Generation Method},
  year    = {2023},
}

RCG is a self-conditioned image generation framework that achieves SOTA unconditional image generation performance on ImageNet 256x256, bridging the long-standing performance gap between unconditional and class-conditional image generation.

Updates

March 2024

  • Update FID evaluation and results to follow ADM suite by hard-coding ADM statistics in torch-fidelity. The modified torch-fidelity could be installed via
pip install -e git+https://github.com/LTH14/torch-fidelity.git@master#egg=torch-fidelity
  • Update an ADM checkpoint trained for 400 epochs (same as the original paper).
  • Include training scripts and pre-trained checkpoint (400 epochs) for DiT-XL with RCG.
  • Update Arxiv.

Preparation

Dataset

Download ImageNet dataset, and place it in your IMAGENET_DIR. Prepare the ImageNet validation set for FID evaluation against validation set:

python prepare_imgnet_val.py --data_path ${IMAGENET_DIR} --output_dir imagenet-val

To evaluate FID against the validation set, do pip install torch-fidelity which install the original torch-fidelity package.

Installation

Download the code

git clone https://github.com/LTH14/rcg.git
cd rcg

A suitable conda environment named rcg can be created and activated with:

conda env create -f environment.yaml
conda activate rcg

Use this link to download the pre-trained VQGAN tokenzier as vqgan_jax_strongaug.ckpt.

Use this link to download the pre-trained moco v3 ViT-B encoder and name it as pretrained_enc_ckpts/mocov3/vitb.pth.tar.

Use this link to download the pre-trained moco v3 ViT-L encoder and name it as pretrained_enc_ckpts/mocov3/vitl.pth.tar.

Usage

RDM

To train the Moco v3 ViT-B representation diffusion model using 4 V100 GPUs:

python -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 --node_rank=0 \
main_rdm.py \
--config config/rdm/mocov3vitb_simplemlp_l12_w1536.yaml \
--batch_size 128 --input_size 256 \
--epochs 200 \
--blr 1e-6 --weight_decay 0.01 \
--output_dir ${OUTPUT_DIR} \
--data_path ${IMAGENET_DIR} \
--dist_url tcp://${MASTER_SERVER_ADDRESS}:2214

To continue a previously interrupted training session, set --resume to the OUTPUT_DIR where checkpoint-last.pth is stored.

The following table provides the pre-trained weights of the Moco v3 ViT-B/ViT-L RDM used in the paper:

Moco v3 ViT-B Moco v3 ViT-L
Class-unconditional RDM Google Drive / config Google Drive / config
Class-conditional RDM Google Drive / config Google Drive / config

Pixel Generator: MAGE

To train a MAGE-B conditioned on Moco v3 ViT-B representations, using 64 V100 GPUs for 200 epochs:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=8 --node_rank=0 \
main_mage.py \
--pretrained_enc_arch mocov3_vit_base \
--pretrained_enc_path pretrained_enc_ckpts/mocov3/vitb.pth.tar --rep_drop_prob 0.1 \
--use_rep --rep_dim 256 --pretrained_enc_withproj --pretrained_enc_proj_dim 256 \
--pretrained_rdm_cfg ${RDM_CFG_PATH} --pretrained_rdm_ckpt ${RDM_CKPT_PATH} \
--rdm_steps 250 --eta 1.0 --temp 6.0 --num_iter 20 --num_images 50000 --cfg 0.0 \
--batch_size 64 --input_size 256 \
--model mage_vit_base_patch16 \
--mask_ratio_min 0.5 --mask_ratio_max 1.0 --mask_ratio_mu 0.75 --mask_ratio_std 0.25 \
--epochs 200 \
--warmup_epochs 10 \
--blr 1.5e-4 --weight_decay 0.05 \
--output_dir ${OUTPUT_DIR} \
--data_path ${IMAGENET_DIR} \
--dist_url tcp://${MASTER_SERVER_ADDRESS}:2214

To train a MAGE-L conditioned on Moco v3 ViT-L representations, change RDM_CFG_PATH and RDM_CKPT_PATH for Moco v3 ViT-L RDM, as well as the following arguments:

--pretrained_enc_arch mocov3_vit_large --pretrained_enc_path pretrained_enc_ckpts/mocov3/vitl.pth.tar --temp 11.0 --model mage_vit_large_patch16

Resume: set --resume to the OUTPUT_DIR where checkpoint-last.pth is stored.

Evaluation: set --resume to the pre-trained MAGE checkpoint and include the --evaluate flag in the above script.

Pre-trained Models:

Rep. Cond. MAGE-B Rep. Cond. MAGE-L
Checkpoint Google Drive Google Drive
Class-unconditional Generation (w/o CFG) FID=3.98, IS=177.8 FID=3.44, IS=186.9
Class-unconditional Generation (w/ CFG) FID=3.19, IS=214.9 (cfg=1.0) FID=2.15, IS=253.4 (cfg=6.0)
Class-conditional Generation (w/o CFG) FID=3.50, IS=194.9 FID=2.99, IS=215.5
Class-conditional Generation (w/ CFG) FID=3.18, IS=242.6 (cfg=1.0) FID=2.25, IS=300.7 (cfg=6.0)

Visualization: use viz_rcg.ipynb to visualize generation results.

Class-unconditional generation examples:

Class-conditional generation examples:

Pixel Generator: DiT

To train a DiT-L conditioned on Moco v3 ViT-B representations, using 128 V100 GPUs for 400 epochs:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=16 --node_rank=0 \
main_dit.py \
--rep_cond --rep_dim 256 \
--pretrained_enc_arch mocov3_vit_base \
--pretrained_enc_path pretrained_enc_ckpts/mocov3/vitb.pth.tar \
--pretrained_rdm_cfg ${RDM_CFG_PATH} \
--pretrained_rdm_ckpt ${RDM_CKPT_PATH} \
--batch_size 16 --image_size 256 --dit_model DiT-L/2 --num-sampling-steps ddim25 \
--epochs 400  \
--lr 1e-4 --weight_decay 0.0 \
--output_dir ${OUTPUT_DIR} \
--data_path ${IMAGENET_DIR} \
--dist_url tcp://${MASTER_SERVER_ADDRESS}:2214

Note: sometimes batch_size=16 causes out of memory for DiT-XL. Changing it to 12 or 14 has a negligible effect on performance.

Resume: set --resume to the OUTPUT_DIR where checkpoint-last.pth is stored.

Evaluation: set --resume to the pre-trained DiT checkpoint and include the --evaluate flag in the above script. Set --num-sampling-steps 250 for better generation performance.

Pre-trained DiT-XL/2 (400 epochs) conditioned on Moco v3 ViT-B representations can be downloaded here (FID=4.89, IS=143.2).

Pixel Generator: ADM

To train an ADM conditioned on Moco v3 ViT-B representations, using 128 V100 GPUs for 100 epochs:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=16 --node_rank=0 \
main_adm.py \
--rep_cond --rep_dim 256 \
--pretrained_enc_arch mocov3_vit_base \
--pretrained_enc_path pretrained_enc_ckpts/mocov3/vitb.pth.tar \
--pretrained_rdm_cfg ${RDM_CFG_PATH} \
--pretrained_rdm_ckpt ${RDM_CKPT_PATH} \
--batch_size 2 --image_size 256 \
--epochs 100  \
--lr 1e-4 --weight_decay 0.0 \
--attention_resolutions 32,16,8 --diffusion_steps 1000 \
--learn_sigma --noise_schedule linear \
--num_channels 256 --num_head_channels 64 --num_res_blocks 2 --resblock_updown \
--use_scale_shift_norm \
--gen_timestep_respacing ddim25 --use_ddim \
--output_dir ${OUTPUT_DIR} \
--data_path ${IMAGENET_DIR} \
--dist_url tcp://${MASTER_SERVER_ADDRESS}:2214

Resume: set --resume to the OUTPUT_DIR where checkpoint-last.pth is stored.

Evaluation: set --resume to the pre-trained ADM checkpoint and include the --evaluate flag in the above script. Set --gen_timestep_respacing 250 and disable --use_ddim for better generation performance.

Pre-trained ADM (400 epochs) conditioned on Moco v3 ViT-B representations can be downloaded here (FID=6.24, IS=136.9).

Pixel Generator: LDM

Use this link to download the tokenizer and name it as vqgan-ckpts/ldm_vqgan_f8_16384/checkpoints/last.ckpt.

To train an LDM-8 conditioned on Moco v3 ViT-B representations, using 64 V100 GPUs for 40 epochs:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=8 --node_rank=0 \
main_ldm.py \
--config config/ldm/cin-ldm-vq-f8-repcond.yaml \
--batch_size 4 \
--epochs 40 \
--blr 2.5e-7 --weight_decay 0.01 \
--output_dir ${OUTPUT_DIR} \
--data_path ${IMAGENET_DIR} \
--dist_url tcp://${MASTER_SERVER_ADDRESS}:2214

Resume: set --resume to the OUTPUT_DIR where checkpoint-last.pth is stored.

Evaluation: set --resume to the pre-trained LDM checkpoint and include the --evaluate flag in the above script.

Pre-trained LDM (40 epochs) conditioned on Moco v3 ViT-B representations can be downloaded here (FID=11.30, IS=101.9).

Contact

If you have any questions, feel free to contact me through email (tianhong@mit.edu). Enjoy!