Person-Image-Generation

Readme

Pytorch implementation for reproducing PIG-MM

• the main backbone is based on HDGanPytorch implementation

Dependencies

• Python 3+
• Pytorch 0.3.+
• Anaconda 3.6+

Data

• please viist CUHK_PEDES and download images of CUHK-PEDES (You may need to contact with the author to acqurie the dataset).
• The 'reid_raw.json' is from https://github.com/ShuangLI59/Person-Search-with-Natural-Language-Description. This file can be used to split the train/val/test.
• we use BERT embedding for the text embedding, you can also use char-CNN-RNN to get text embedding.reedscot/icml2016

pretrain model

• BaiduYuncode: f4kf

Train

• go to train/train_gan:

• some pathes to dataset and model in the code should redirection to your own path

python train_worker.py

• To use multiple GPUs, simply set device='0,1,..' as a set of gpu ids.

Test

• go to test:

python test_worker.py

Tips

• we use the first 10k training set due to limited computing resources, but we also do the experiment on the whole training set.
• 3206 is class_num of 10k training set && 11003 is class_num of whole training set
• the training on the whole train set (30k) is hard, model collapse may happen sometimes, we suggest you to finetune the parameter, e.g, id_loss_rate.. or just re-train the model use the same parameter(I have tried, it is useful.)

Acknowledgements

HDGanPytorch implementation

StakGAN Pytorch implementation

AttanGanPytorch implementation