USAGE.md

Domain Flow for Mixture Style Generation on Latent Space Exploration and Control

流域應用於隱空間探索與控制完成混合風格生成

---

Dependency

pytorch, yaml, tensorboard (from https://github.com/dmlc/tensorboard), and tensorboardX (from https://github.com/lanpa/tensorboard-pytorch).

The code base was developed using Anaconda with the following packages.

pip install tensorboard tensorboardX;

Example Usage

Description

• master_train.py : 使用這個檔案來執行訓練；如果只想進行 Disentangle 的話，可把Flowing 部分 mark 掉執行。
• master_trainer.py : 整個模型架構的組合，以及loss function的設定。主要是用 MASTER_Trainer。
• master_networks.py : 網路的component。
• master_test.py : 在 inference 時使用。
• utils .py : 工具function
• Note.md : 實驗記錄的部分 & 訓練command

Training

1. Prepare dataset

   • 將 dataset 分成 trainA / trainB 和 testA/testB，放到 datasets

2. Setup the yaml file.

   • Check out configs/style_shoes_label_floder_OO.yaml for folder-based dataset organization.

3. 開啟 Visdom python -m visdom.server

4. Start training

   python master_train.py --config configs/style_shoes_label_folder_OO.yaml --trainer MASTER
   

5. Intermediate image outputs and model binary files are stored in outputs/style_shoes_label_floder_OO

Testing

First, download the pretrained models and put them in models folder.