Skip to content

train_substance_classifier.md

Latest commit

 

History

History
45 lines (37 loc) · 2.82 KB

train_substance_classifier.md

File metadata and controls

45 lines (37 loc) · 2.82 KB

Substance Classifier

The substance classifier is used by the SUBS metric. We train it a mix of crops obtained from the opensurfaces dataset and stationary textures dataset. More details are provided in the supplementary metrics section of the paper.

Preparation of Open Surfaces Crops

Download the opensurfaces dataset. The download script will extract the rectified surface masks. From those extracted rectified surface masks, we prepare a train_masks directory and a val_masks directory having masks with following file names.

Then, we use the following script to extract crops from these rectified surface masks.

python scripts/plan2scene/metric_impl/substance_classifier/prepare_opensurfaces_crops.py ./data/processed/open-surfaces-crops/train PATH/TO/train_masks 
python scripts/plan2scene/metric_impl/substance_classifier/prepare_opensurfaces_crops.py ./data/processed/open-surfaces-crops/val PATH/TO/val_masks

Preparation of Texture Dataset Crops

We use the following script to extract crops from the texture dataset.

python scripts/plan2scene/metric_impl/substance_classifier/prepare_texture_crops.py ./data/processed/stationary-textures-crops/train PATH/TO/TRAIN/TEXTURES 
python scripts/plan2scene/metric_impl/substance_classifier/prepare_texture_crops.py ./data/processed/stationary-textures-crops/val PATH/TO/VAL/TEXTURES

Training Substance Classifier

  1. Make sure the crops extracted from the texture dataset are in the [PROJECT_ROOT]/data/processed/stationary-textures-dataset-crops/train and /data/processed/stationary-textures-crops/val directories.

  2. Make sure the crops extracted from opensurfaces rectified surface masks are in the /data/processed/open-surfaces-crops/train and [PROJECT_ROOT]/data/processed/open-surfaces-crops/val directories.

  3. Run the following command to start training. This will continue training for 200 epochs.

    export PYTHONPATH=./code/src
python ./code/scripts/plan2scene/metric_impl/substance_classifier/train.py ./trained_models/substance_classifier/default ./conf/plan2scene/substance_classifier_conf/default.json --save-model-interval 1

    Checkpoints are saved at './trained_models/substance_classifier/default/checkpoints' directory.

  4. Preview learning curves using Tensorboard.

    tensorboard --logdir=./trained_models/substance_classifier/default/tensorboard

  5. Choose the best checkpoint based on substance classification accuracy. Update substance_classifier.checkpoint_path field of ./conf/plan2scene/metric.json to point to the best checkpoint. Update substance_classifier.conf_path field of the same file to ./trained_models/substance_classifier/default/conf/substance_classifier_conf.json