Train FCN (Fully Convolutional Network)

This page shows how to train FCN with your own dataset.

FCN is a neural network model used for semantic segmentation.

Any size of image can be applied to this network as long as your GPU has enough memory.

Available Dataset Class

SemanticSegmentationDataset (imported from jsk_recognition_utils.datasets)

This class assumes the following directory structure for each split.

path_to_awesome_dataset/
|-- JPEGImages
|   |-- foo.jpg
|   |-- bar.jpg
|   `-- etc.
|-- SegmentationClass
|   |-- foo.npy
|   |-- bar.npy
|   `-- etc.
|-- class_names.txt
`-- etc.

Arguments

  • --train_dataset_dir (string, default: $(rospack find jsk_perception)/learning_datasets/kitchen_dataset/train)

  • --val_dataset_dir (string, default: $(rospack find jsk_perception)/learning_datasets/kitchen_dataset/test)

    Directory name which contains dataset for training and validation respectively.

  • --model_name (string, default: fcn32s)

    Model name. Currently, fcn32s, fcn16s, fcn8s and fcn8s_at_once are supported.

  • --gpu (int, default: 0)

    GPU id. -1 means CPU mode, but we recommend to use GPU for much faster computing.

  • --batch_size (int, default: 1)

    Number of images used simultaneously in each iteration.

    You should decrease this number when you face memory allocation error.

  • --max_epoch (int, default: 100)

    Stop trigger for training.

  • --lr (float, default: 1e-10)

    Learning rate.

    Perhaps you should decrease this number when you face NaN value as loss.

  • --weight_decay (float, default: 0.0001)

    Weight decay.

  • --out_dir (string, default: ${ROS_HOME}/learning_logs/<timestamp>)

    Output directory name.

  • --progressbar_update_interval (float, default: 10)

    Interval for updating progress bar shown while training.

    The unit is [iteration].

  • --print_interval (float, default: 100)

  • --print_interval_type (string, default: iteration)

    Interval for printing information like current epoch, elapsed time, loss, etc. on terminal.

    Note that XXX_interval_type can be chosen from {’epoch’, ‘iteration’}.

  • --log_interval (float, default: 10)

  • --log_interval_type (string, default: iteration)

    Interval for logging information to <out_dir>/log.json.

  • --plot_interval (float, default: 5)

  • --plot_interval_type (string, default: epoch)

    Interval for plotting loss to <out_dir>/loss_plot.png.

  • --eval_interval (float, default: 10)

  • --eval_interval_type (string, default: epoch)

    Interval for running evaluation.

  • --save_interval (float, default: 10)

  • --save_interval_type (string, default: epoch)

    Interval for saving snapshot of model to <out_dir>/model_snapshot.npz.

    Trainer listens to loss for validation dataset according to this interval, and saves snapshot when the loss value becomes minimum.

Output

All these files will be automatically generated under <out_dir>.

  • log.json

  • loss_plot.png

  • model_name.yaml

  • model_snapshot.npz

  • network_architecture.dot

  • params.yaml

  • target_names.yaml

Usage

rosrun jsk_perception train_fcn.py [ARGS]