# Butterfly GAN

## Model description

Based on paper: Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis

which states: "Notably, the model converges from scratch with just a few hours of training on a single RTX-2080 GPU, and has a consistent performance, even with less than 100 training samples"

also dubbed the Light-GAN model. This model was trained using the script here which is adapted from the lucidrains repo.

Differently from the script above, I used the transforms from the official repo. Because our training images were already cropped and aligned. official paper implementation repo

transform_list = [
transforms.Resize((int(im_size),int(im_size))),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
]


## Intended uses & limitations

Intended for fun & learning~

#### How to use


import torch
from huggan.pytorch.lightweight_gan.lightweight_gan import LightweightGAN # install the community-events repo above

gan = LightweightGAN.from_pretrained("ceyda/butterfly_cropped_uniq1K_512")
gan.eval()
batch_size = 1
ims = gan.G(torch.randn(batch_size, gan.latent_dim)).clamp_(0., 1.)*255
ims = ims.permute(0,2,3,1).detach().cpu().numpy().astype(np.uint8)
# ims is [BxWxHxC] call Image.fromarray(ims[0])


#### Limitations and bias

• During training I filtered the dataset to have only 1 butterfly from each species available. Otherwise the model generated less varied butterflies (a few species with more images would dominate).
• The dataset was also filtered using CLIP scores for ['pretty butterfly','one butterfly','butterfly with open wings','colorful butterfly']. While this was done to eliminate images that contained no butterflies(just scientific tags, cluttered images) from the full dataset. It is easy to imagine where this type of approach would be problematic in certain scenarios; who is to say which butterfly is "pretty" and should be in the dataset.ie; CLIP failing to identify a butterfly might exclude it from the dataset causing bias.

## Training data

1000 images are used, while it was possible to increase this number, we didn't have time to manually curate the dataset.

& also wanted to see if it was possible to do low data training as mention in the paper.

More details are on the data card

## Training procedure

Trained on 2xA4000s for ~1day. Can see good results within 7-12h. Importans params: "--batch_size 64 --gradient_accumulate_every 4 --image_size 512 --mixed_precision fp16" Training logs can be seen here

## Eval results

calculated FID score on 100 images. results for different checkpoints are here

but can't say it is too meaningful (due to the shortcomings of FID score)

## Generated Images

Play with the demo