Apostasi0225/susie-finetuned

1. Where is the code

All code can be found in https://huggingface.co/Apostasi0225/susie-finetuned.
It also contains the dataset and our3 model checkpoint.

2. Environment Setup

System requirements:

• Linux
• CUDA 11.8
• CUDNN 8.6

Build the environment using conda:

conda create -n susie python=3.10
conda activate susie

cd susie
pip install -r requirements.txt
pip install -e .

Additionally, install the following packages:

# Install Pytorch for model loading
pip install torch==2.6.0

# Install jaxlib 0.4.11
pip install https://storage.googleapis.com/jax-releases/cuda11/jaxlib-0.4.11+cuda11.cudnn86-cp310-cp310-manylinux2014_x86_64.whl

pip install scipy==1.12.0

# Downgrade numpy and orbax-checkpoint
pip install orbax-checkpoint==0.3.5
pip install numpy==1.24

# Install skimage
pip install scikit-image

pip install ipykernel

Modify diffusers library code:

1. Find the location of the package

pip show diffusers # e.g. Location: /home/username/miniconda3/envs/susie/lib/python3.10/site-packages/diffusers

2. Open the folder and find utils/dynamic_modules_utils.py

3. Remove cached_download in line 28: from huggingface_hub import ...

3. Getting Started

Step 1: Generate TFRecords data

Run 0-generate_dataset.ipynb to generate the TFRecords data.
This will automatically split the dataset into training and validation sets.

Step 2: Train the Model

Follow the command in 1-train.md to start training.
You can modify the training configuration parameters by editing susie/configs/base.py

Step 3: Evaluate the Model

We provide the checkpoint for ours3 model.
After training, evaluate the model using 3-eval.ipynb.
This will test the model on the validation set and output evaluation metrics such as SSIM and PSNR for each task.