ECCV 2024
Oral Talk in ICML 2024 Workshop on Long Context Foundation Models (LCFM)
This repository represents the official implementation of the paper titled "ZigMa: A DiT-style Zigzag Mamba Diffusion Model (ECCV 2024)".
Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes Fischer, Björn Ommer
We present ZigMa, a scanning scheme that follows a zigzag pattern, considering both spatial continuity and parameter efficiency. We further adapt this scheme to video, separating the reasoning between spatial and temporal dimensions, thus achieving efficient parameter utilization. Our design allows for greater incorporation of inductive bias for non-1D data and improves parameter efficiency in diffusion models.
Please cite our paper:
@InProceedings{hu2024zigma,
title={ZigMa: A DiT-style Zigzag Mamba Diffusion Model},
author={Vincent Tao Hu and Stefan Andreas Baumann and Ming Gui and Olga Grebenkova and Pingchuan Ma and Johannes Fischer and Björn Ommer},
booktitle = {ECCV},
year={2024}
}May. 24th, 2024: 🚀🚀🚀 New checkpoints for FacesHQ1024, landscape1024, Churches256 datasets.April. 6th, 2024: Support for FP16 training, and checkpoint function, and torch.compile to achieve better memory utilization and speed boosting.April. 2th, 2024: Main code released.
from model_zigma import ZigMa
img_dim = 32
in_channels = 3
model = ZigMa(
in_channels=in_channels,
embed_dim=640,
depth=18,
img_dim=img_dim,
patch_size=1,
has_text=True,
d_context=768,
n_context_token=77,
device="cuda",
scan_type="zigzagN8",
use_pe=2,
)
x = torch.rand(10, in_channels, img_dim, img_dim).to("cuda")
t = torch.rand(10).to("cuda")
_context = torch.rand(10, 77, 768).to("cuda")
o = model(x, t, y=_context)
print(o.shape)In comparison to the original implementation, we implement a selection of training speed acceleration and memory saving features including gradient checkpointing
| torch.compile | gradient checkpointing | training speed | memory |
|---|---|---|---|
| ❌ | ❌ | 1.05 iters/sec | 18G |
| ❌ | ✔ | 0.93 steps/sec | 9G |
| ✔ | ❌ | 1.8 iters/sec | 18G |
torch.compiles is for indexing operation: here and here
Sweep-2, 1GPU
accelerate launch --num_processes 1 --num_machines 1 --mixed_precision fp16 train_acc.py model=sweep2_b1 use_latent=1 data=celebamm256_uncond ckpt_every=10_000 data.sample_fid_n=5_000 data.sample_fid_bs=4 data.sample_fid_every=10_000 data.batch_size=8 note=_ Zigzag-8, 1GPU
CUDA_VISIBLE_DEVICES=4 accelerate launch --num_processes 1 --num_machines 1 --mixed_precision fp16 --main_process_ip 127.0.0.1 --main_process_port 8868 train_acc.py model=zigzag8_b1 use_latent=1 data=celebamm256_uncond ckpt_every=10_000 data.sample_fid_n=5_000 data.sample_fid_bs=4 data.sample_fid_every=10_000 data.batch_size=4 note=_ Baseline, multi-GPU
CUDA_VISIBLE_DEVICES="0,1,2,3" accelerate launch --num_processes 4 --num_machines 1 --multi_gpu --mixed_precision fp16 --main_process_ip 127.0.0.1 --main_process_port 8868 train_acc.py model=3d_sweep2_b2 use_latent=1 data=ucf101 ckpt_every=10_000 data.sample_fid_n=20_0 data.sample_fid_bs=4 data.sample_fid_every=10_000 data.batch_size=4 note=_ Factorized 3D Zigzag: sst, multi-GPU
CUDA_VISIBLE_DEVICES="0,1,2,3" accelerate launch --num_processes 4 --num_machines 1 --multi_gpu --mixed_precision fp16 --main_process_ip 127.0.0.1 --main_process_port 8868 train_acc.py model=3d_zigzag8sst_b2 use_latent=1 data=ucf101 ckpt_every=10_000 data.sample_fid_n=20_0 data.sample_fid_bs=4 data.sample_fid_every=10_000 data.batch_size=4 note=_ You can directly download the model in this repository. You also can download the model in python script:
from huggingface_hub import hf_hub_download
hf_hub_download(
repo_id="taohu/zigma",
filename="faceshq1024_0090000.pt",
local_dir="./checkpoints",
)| Dataset | Checkingpoint | Model | data |
|---|---|---|---|
| faceshq1024.pt | faceshq1024_0090000.pt | model=s1024_zigzag8_b2_old | data=facehq_1024 |
| landscape1024 | landscape1024_0210000.pt | model=s1024_zigzag8_b2_old | data=landscapehq_1024 |
| Churches256 | churches256_0280000.pt | model=zigzag8_b1_pe2 | data=churches256 |
| Coco256 | zigzagN8_b1_pe2_coco14_bs48_0400000.pt | mode=zigzag8_b1_pe2 | data=coco14 (31.0) |
1GPU sampling
CUDA_VISIBLE_DEVICES="2" accelerate launch --num_processes 1 --num_machines 1 sample_acc.py model=s1024_zigzag8_b2_old use_latent=1 data=facehq_1024 ckpt_every=10_000 data.sample_fid_n=5_000 data.sample_fid_bs=4 data.sample_fid_every=10_000 data.batch_size=8 sample_mode=ODE likelihood=0 num_fid_samples=5_000 sample_debug=0 ckpt=checkpoints/faceshq1024_0060000.pt The sampled images will be saved both on wandb (disable with use_wandb=False) and directory samples/
cuda==11.8,python==3.11, torch==2.2.0, gcc==11.3(for SSM enviroment)
python=3.11 # support the torch.compile for the time being. pytorch/pytorch#120233 (comment)
conda create -n zigma python=3.11
conda activate zigma
conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit
conda install pytorch torchvision pytorch-cuda=11.8 -c pytorch -c nvidia
pip install torchdiffeq matplotlib h5py timm diffusers accelerate loguru blobfile ml_collections wandb
pip install hydra-core opencv-python torch-fidelity webdataset einops pytorch_lightning
pip install torchmetrics --upgrade
pip install opencv-python causal-conv1d
cd dis_causal_conv1d && pip install -e . && cd ..
cd dis_mamba && pip install -e . && cd ..
pip install moviepy imageio #wandb.Video() need it
pip install scikit-learn --upgrade
pip install transformers==4.36.2
pip install numpy-hilbert-curve # (optional) for generating the hilbert path
pip install av # (optional) to use the ucf101 frame extracting
pip install -v -U git+https://github.com/facebookresearch/xformers.git@main#egg=xformers #for FDD metricsInstalling Mamba may cost a lot of effort. If you encounter problems, this issues in Mamba may be very helpful.
Create a file under the directory ./config/wandb/default.yaml:
key: YOUR_WANDB_KEY
entity: YOUR_ENTITY
project: YOUR_PROJECT_NAME- If you meeet some issues for installing ssm, maybe you can find solution here: https://github.com/state-spaces/mamba/issues
Due to privacy issue, we cannot share the dataset here, basically, we use MM-CelebA-HQ-Dataset from https://github.com/IIGROUP/MM-CelebA-HQ-Dataset, we organize into the format of webdataset to enable the scalable training in multi-gpu.
Webdataset Format:
- image: image.jpg # ranging from [-1,1], shape should be [3,256,256]
- latent: img_feature256.npy # latent feature for latent generation, shape should be [4,32,32]
The dataset we use include:
- MM-CelebA-HQ for 256 and 512 resolution training
- FacesHQ1024 for 1024 resolution
- UCF101 for 16x256x256 resolution
This work is licensed under the Apache License, Version 2.0 (as defined in the LICENSE).
By downloading and using the code and model you agree to the terms in the LICENSE.