Note: This code is heavily based off of the code release for Dilated Residual Networks located here.
Please make sure you have downloaded the Oil Change dataset. See the main README for download instructions.
You can download our pretrained segmentation models using the following command:
./download_pretrained.shOnce they are downloaded, the eval.ipynb and visualize.ipynb notebooks are already configured to evaluate and visualize the pretrained models on our Oil Change dataset.
The training scripts use Munch config files for starting and resuming training runs. You can reference the config directory for sample config files.
For example, you can start a new training job using one of the config files like so:
python train.py config/drn_d_22_OilChange.ymlEach new training job gets its own log directory and checkpoint directory, time-stamped with the job's start time. By default, log directories are created inside logs, and checkpoint directories are created inside checkpoints. The training script will write an updated config file to the log directory every time a checkpoint is written. The updated config file will contain all parameters required to resume the run from the last checkpoint.
To resume a training job, simply pass in the corresponding config file:
python train.py logs/log_dir/config.ymlYou can modify the resume parameter in the config file to resume from a specific checkpoint. Note that no config file will be written to the run's log directory until after the first model checkpoint has been written.
To monitor training, start TensorBoard and point it at the root log directory:
tensorboard --logdir logs/You can then view the TensorBoard interface in your browser at localhost:6006.
Please see the eval.ipynb notebook for code to evaluate a trained model.
Please see the visualize.ipynb notebook to look at example outputs from a trained model.