Harsh Rangwani*, Lavish Bansal*, Kartik Sharma, Tejan Karmali, Varun Jampani, R. Venkatesh Babu
Vision and AI Lab, IISc Bangalore
CVPR 2023
Update: Thanks to @prajwalsingh for providing NoisyTwins implementation on base StyleGAN2: (https://github.com/prajwalsingh/StyleGAN2-NosiyTwins)
TLDR: NoisyTwins is a self-supervised regularization scheme for StyleGANs, which helps in alleviating mode collapse and leads to consistent conditional fine-grained image generation.
CIFAR-10 dataset will be downloaded automatically in ./data
folder in the project directory.
ImageNet-LT is a subset of ImageNet dataset which can be downloaded from this link.
The long-tailed version of the CIFAR-10 and ImageNet-LT datasets will be created automatically by code.
For iNaturalist2019 dataset please download the files from the following link: here. After download untar the images, following which use the script available here to create image file names.
NOTE: For iNaturalist19 and ImageNet-LT, the -l
and -hdf5
flags are passed to use the HDF5 files. As these datasets are very large in size, hdf5 files are required for faster processing which provide the data in chunks. Once the HDF5 is created for a specific dataset with a given image resolution in the first run, make sure the DATA_DIR
in run_script.sh
is set to the HDF5 file's directory, so that it doesn't reads the complete dataset everytime and simply use the file.
For Imagenet-Carnivores dataset, we provide the zip files in data
folder. Unzip the files there, which will correspond to the path in DATA_DIR
. The dataset was originally provided by the authors of Shahbazi et. al. Link which they used for their experiements.
For AnimalFace, the dataset can be downloaded from here. After downloading the zip file of dataset, unzip the file and remove the Natural
folder from it, as it contains 5-6 sample images.
For installing all the requirements use the following command:
conda env create -f environment.yml -n studiogan
To run the experiements, the scripts have been provided in run_script.sh
file.
Depending on the experiment setting that you want to keep for a run(experiment), changes can be done in the run_script.sh
file.
The scripts for 3 tasks and all 5 datasets used in the paper have been provided.
Tasks include Training the model, Evaluating a trained model, Visualizing the results from a trained model. Change the TASK
parameter to perform the required task. Similarly change the DATASET
name for the required dataset.
Change your Personal API KEY WANBD_API_KEY
in the WANDB_API
parameter to connect to experiement to your WANDB account.
For a specific dataset, changing the CFG
parameter will read the specified config file in src/configs/{DATASET}
. These configs specify the exhuastive set of hyperparameters for the experiments.
NOTE: Full path of the config file need to be specified.
For running the NoisyTwins experiment, keep the config as follows. (suppose for iNat19)
CFG="src/configs/iNat19/StyleGAN2-SPD-ADA-LC-NoisyTwins.yaml"
Similarly for a baseline
CFG="src/configs/iNat19/StyleGAN2-SPD-ADA-LC.yaml"
The DATA_DIR
parameter can also be changed depending on the location of the dataset or HDF5 files of the dataset.
You can change the config files for tuning hyper-parameters or create your own custom config files and datasets and run experiments accordingly.
For hyper-parameters not specified in config file, their default values will be taken from the config.py
file.
Most of the code related to the noise augmentation has been added in src/models/stylegan2.py
. The code of Barlow Twins based NoisyTwins Regulariser has been present in src/utils/barlow.py
and src/utils/sample.py
for creating the twins and loss calculation. In src/worker.py
, the NoisyTwins loss in integrated with other losses in the main training loop.
We provide the results below for ImageNet-LT (left) and iNaturalist 2019 (right). For more details please check our paper.
Method | FID ImgNet | FID_CLIP | iFID_CLIP | Precision | Recall | FID iNat | FID_CLIP | iFID_CLIP | Precision | Recall |
---|---|---|---|---|---|---|---|---|---|---|
SG2 | 41.25 | 11.64 | 46.93 | 0.50 | 0.48 | 19.34 | 3.33 | 38.24 | 0.74 | 0.17 |
SG2+ADA | 37.20 | 11.04 | 47.41 | 0.54 | 0.38 | 14.92 | 2.30 | 35.19 | 0.75 | 0.57 |
SG2+ADA+gSR | 24.78 | 8.21 | 44.42 | 0.63 | 0.35 | 15.17 | 2.06 | 36.22 | 0.74 | 0.46 |
SG2+ADA+Noise (Ours) | 22.17 | 7.11 | 41.20 | 0.72 | 0.33 | 12.87 | 1.37 | 31.43 | 0.81 | 0.63 |
+ NoisyTwins (Ours) | 21.29 | 6.41 | 39.74 | 0.67 | 0.49 | 11.46 | 1.14 | 31.50 | 0.79 | 0.67 |
We also provide evaluation metrics of FID and iFID using the CLIP backbone. We find these metrics to be very reliable in comparison to Inception based metrics. For evaluating metrics with CLIP backbone specify --eval_backbone CLIP
while running the evaluation. Please look at run_script.sh
for commands required for evaluation.
If you find our code or work useful in any way, please consider citing us:
@inproceedings{rangwani2023noisytwins,
author = {Rangwani, Harsh and Bansal, Lavish and Sharma, Kartik and Karmali, Tejan and Jampani, Varun and Babu, R. Venkatesh},
title = {NoisyTwins: Class-Consistent and Diverse Image Generation through StyleGANs},
booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2023},
}
Please contact [email protected]
in case you have any comments or suggestions.
Our code is based on StudioGAN and supports all the functionality offered by StudioGAN.
PyTorch-StudioGAN: https://github.com/POSTECH-CVLab/PyTorch-StudioGAN