Skip to content

Latest commit

 

History

History
21 lines (18 loc) · 1.84 KB

train_texture_synth.md

File metadata and controls

21 lines (18 loc) · 1.84 KB
Raw

Training Texture Synthesis Stage

This stage uses a modified version of the neural texture project by Henzler et. al.

IMPORTANT: We use a separate python environment with torch==1.4.0 and torchvision==0.5.0 to train the texture_gen stage. We have noticed that later versions of PyTorch causes poor training of neural texture synthesis networks. This is likely due to a compatibility issue of the noise_kernel which requires further investigation. Also, note that authors of neural texture project specifies PyTorch 1.4. You can use this requirements file to setup the python environment for training.

For inference, we use the same python environment as the main project. The issue only affects training.

Training the network

  1. Copy the textures dataset to ./data/input/stationary-textures-dataset/train and ./data/input/stationary-textures-dataset/val.
  2. Start training by running the following. 
    python code/scripts/plan2scene/texture_gen/train.py ./trained_models/texture_gen/default ./conf/plan2scene/texture_synth_conf/v2.yml
    You may use tensorboard to preview training curves & previews, and then choose the best checkpoint based on epoch_val_texture_loss. 
    python -m tensorboard --logdir ./trained_models/texture_gen/default/tensorboard
    The best checkpoints are saved at ./trained_models/texture_gen/default/best_models. Periodic checkpoints are saved at ./trained_models/texture_gen/default/checkpoints.
  3. Update 'checkpoint_path' of ./conf/plan2scene/texture_gen.json to the path of the selected best checkpoint from the previous step. Update 'texture_synth_conf' of the same file to ./trained_models/texture_gen/default/config.yml.