Code adapted from Aladdin Persson.

This is my project on Deep Learning. It generates traditional thai house images using DCGANs. Total training images = 1,000 images (10,000 after augmented). The result images are still not good. More training, data, and tuning needed.

├───dataset
│   └───[your_dataset]
      └───images
├───gen_img
│   └───[name_your_folder]
├───saved_models
│   └───[name_your_folder]
└───tensorboard_logs
    └───[name_your_folder]

See section 1, 2 in the notebook

• Put training images in dataset/[your_dataset]/images/
• Set project path in the notebook

  DATASET_PATH = "./dataset/[your_dataset]"
  PROJECT_PATH = "[name_your_folder]"
  

• Run Thaihouse-gen_DCGANs.ipynb
• Weights will be saved every epochs in saved_models/
• Tensorboard logs will be in tensorboard_logs/

Training with 50epochs

Tensorboard

See section 3, 4 in the notebook. Generate one random tensor and put into generator.

See section 5 in the notebook. But as you can see, generator loss is still not converging.