train_texture_prop.md

Train Texture Propagation Stage

Pre-requisites

1. It is assumed that the texture synthesis stage is already trained, or a pre-trained model is configured.
2. Infer textures for photo observed surfaces using the VGG textureness score by running vgg_crop_selector.py. The texture propagation stage is trained using these inferences.

   # We simulate surface unobservations using the train_graph_generator and the val_graph_generator defined in texture_prop_conf. Therefore, drop_fraction is set to 0.0 in the following scripts.
   python code/scripts/plan2scene/preprocessing/fill_room_embeddings.py ./data/processed/texture_gen/train/drop_0.0 train --drop 0.0
   python code/scripts/plan2scene/preprocessing/fill_room_embeddings.py ./data/processed/texture_gen/val/drop_0.0 val --drop 0.0
   
   python code/scripts/plan2scene/crop_select/vgg_crop_selector.py ./data/processed/vgg_crop_select/train/drop_0.0 ./data/processed/texture_gen/train/drop_0.0 train --drop 0.0
   python code/scripts/plan2scene/crop_select/vgg_crop_selector.py ./data/processed/vgg_crop_select/val/drop_0.0 ./data/processed/texture_gen/val/drop_0.0 val --drop 0.0

Training Instructions

Run the following command to start training.

python code/scripts/plan2scene/texture_prop/train.py ./trained_models/texture_prop/default

You may use tensorboard to preview training curves.

tensorboard --logdir ./trained_models/texture_prop/default/tensorboard

Checkpoints are saved at ./trained_models/texture_prop/default/checkpoints.

Preview results of every 50th epoch using this script as shown below.

# Preview every 50th epoch
CUDA_VISIBLE_DEVICES=0 python ./code/scripts/plan2scene/texture_prop/preview_nth_epoch_all_prop.py ./trained_models/texture_prop/default/preview_results 50 ./trained_models/texture_prop/default
# Open ./trained_models/texture_prop/default/preview_results/preview.html.

We usually preview every 50th epoch and choose the best epoch. Early epochs tend to look washed out with pink color shifts. You may also use epoch_val_color, epoch_val_subs and epoch_val_freq plots in tensorboard to help your decision.

For the figures reported in the paper, we used the epoch 250.