TensorBob

TensorFlow tools

0. dependency

• slim: tensorflow/models/research/slim.
  • add /path/to/models/research/slim in PYTHONPATH, so that we can use from nets import vgg to build models.

1. tools

1.1. dataset

• target:
  • create tf.data.Dataset objects with Data Argument.
  • create tf.data.Dataset object for Open Data Source, such as ImageNet, VOC, etc.
  • config by python dicts.
• Module Architecture:
  • base_dataset.py: tf.data.Dataset wrapper, including BaseDataset and MergedDataset.
  • dataset_utils: utils to create tf.data.Dataset objects configured by python dicts.
  • segmentation_dataset_utils.py: create BaseDataset and MergedDataset object for segmentation task.
  • Open Database:
    • imagenet.py: ImageNet(classification).
    • voc2012.py: VOC2012(classification & segmentation).
    • ade2016.py: Scene Parsing Challenge 2016(segmentation).
    • camvid.py: Cambridge-driving Labeled Video Database(segmentation).
• For more information about dataset, please check here.

1.2. training

• target:
  • train models by SingularMonitoredSession & hooks.
  • build basic training procedure.
• Module Architecture:
  • training_utils.py: produce hooks, creating train_op function and training function.
  • trainer_utils.py: learning rate utils, slim model utils and scaffold utils.
  • trainer.py:
    • produce basic training procedure by Trainer.
    • trainer for classification task BaseClassificationTrainer.
    • trainer for segmentation task BaseSegmentationTrainer.

1.3. evaluating

• evaluator: evaluate given metrics with existing models on val set.

1.4. Models

• creating reusable models for different tasks.
• semantic segmentation:
  • fcn(fcn_8s_vgg16, fcn_8s_resnet_v2_50)
  • segnet(segnet_vgg16)
  • fc_densenet(fc_densenet)

1.5. Utils

• initializers.py
• regularizers.py
• metrics_utils.py: utils to compute iou by confusion matrix.
• variables.py: very important and useful.
• preprocessing.py: utils for image preprocess.

2. Examples

• target: use tensorbob tools to train/test models.
• modules:
  • imagenet：classification models training/testing.
  • voc2012: classification & image segmentation training/testing.
  • kaggle:
    • whale: more information about this solution, please click here.
  • ade: segmentation training and predicting.
  • camvid: segmentation training and evaluating.