Paper is available on https://arxiv.org/abs/2306.14708 Accepted by ICONIP2024
- python 3.8.0
- Pytorch 1.8.0
- Pandas 1.2.2
- tqdm 4.62.3
- torchvision 0.9.0
- Pillow 7.2.0
- matplotlib 3.3.4
- At least 1x6GB NVIDIA GPU
- Download the preprocessed metadata for birds coco and extract them to
data/
- Download the birds image data. Extract them to
data/birds/
- Download coco2014 dataset and extract the images to
data/coco/images/
- [DF-GAN for bird] It is in '/gen_weights', There are three pth file in it.
- [Text encoder for bird and coco] It is in '../text_encoder_weights/text_encoder200.pth'
cd src/
python train_segan.py
cd src/
python eval_example.py
python metrics_evaluation.py
##Tips
- We can slightly increase the learning rate and get the better result.
- Generator's LR ~ (0.0001,0.0004)
- Discriminator's LR ~ (0.0003,0.0012)
- Do not use sgd, adam is better.
Random images in training process
300<=Epoch<=500, Image is better.