MingyuJ666/SEAttnGAN

[ICONIP'24]Mingyu.Jin's final year project

AccAttnGAN

The whole structure

Paper is available on https://arxiv.org/abs/2306.14708 Accepted by ICONIP2024

Requirements

• 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

Preparation

Datasets

1. Download the preprocessed metadata for birds coco and extract them to data/
2. Download the birds image data. Extract them to data/birds/
3. Download coco2014 dataset and extract the images to data/coco/images/

Pretrained Model

• [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'

Training

cd src/

Train the model

• python train_segan.py

Evaluation

cd src/

Input the sentence in the model

• python eval_example.py

compute IS and FID

• 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.

Image in Epoch 330

Random images in training process

300<=Epoch<=500, Image is better.

Some perfect images