Keras Trainer

An abstraction to train Keras CNN models for image classification.

This package provides an easy framework to apply extra training steps such as image preprocessing, random cropping, balanced sampling, training with probabilistic labels and using multi-loss functions.

The models supported are the ones specified in the keras-model-specs package that correspond to the latest models available in keras-applications

These are:

VGG:

• vgg16, vgg19

ResNet

• resnet50, resnet101 resnet152, resnet50_v2, resnet101_v2, resnet152_v2

ResNeXt

• ResNeXt50, ResNeXt101

MobileNet

• mobilenet_v1, mobilenet_v2, mobilenet_v3_small, mobilenet_v3_large

Inception

• inception_resnet_v2, inception_v3

Xception

• xception

NasNet

• nasnet_large, nasnet_mobile

DenseNet

• densenet_169, densenet_121, densenet_201

EfficientNet

• efficientnetb0, efficientnetb1, efficientnetb2, efficientnetb3, efficientnetb4, efficientnetb5, efficientnetb6, efficientnetb7

And the default model configurations are specified here.

Keras Trainer definition

These are the default options:

OPTIONS = {
        'model_spec': {'type': str},
        'output_logs_dir': {'type': str},
        'output_model_dir': {'type': str},
        'activation': {'type': str, 'default': 'softmax'},
        'batch_size': {'type': int, 'default': 1},
        'callback_list': {'type': list, 'default': []},
        'checkpoint_path': {'type': str, 'default': None},
        'class_weights': {'type': None, 'default': None},
        'custom_crop': {'type': bool, 'default': False},
        'custom_model': {'type': None, 'default': None},
        'decay': {'type': float, 'default': 0.0005},
        'dropout_rate': {'type': float, 'default': 0.0},
        'epochs': {'type': int, 'default': 1},
        'freeze_layers_list': {'type': None, 'default': None},
        'loss_function': {'type': str, 'default': 'categorical_crossentropy'},
        'include_top': {'type': bool, 'default': False},
        'input_shape': {'type': None, 'default': None},
        'iterator_mode': {'type': str, 'default': None},
        'loss_weights': {'type': None, 'default': None},
        'max_queue_size': {'type': int, 'default': 16},
        'metrics': {'type': list, 'default': ['accuracy']},
        'model_kwargs': {'type': dict, 'default': {}},
        'momentum': {'type': float, 'default': 0.9},
        'num_classes': {'type': int, 'default': None},
        'num_gpus': {'type': int, 'default': 0},
        'optimizer': {'type': None, 'default': None},
        'pooling': {'type': str, 'default': 'avg'},
        'random_crop_size': {'type': float, 'default': None},
        'regularization_function': {'type': None, 'default': None},
        'regularization_layers': {'type': None, 'default': None},
        'regularize_bias': {'type': bool, 'default': False},
        'save_training_options': {'type': bool, 'default': True},
        'sgd_lr': {'type': float, 'default': 0.01},
        'top_layers': {'type': None, 'default': None},
        'track_sensitivity': {'type': bool, 'default': False},
        'train_data_generator': {'type': None, 'default': None},
        'train_dataset_dataframe_path': {'type': str, 'default': None},
        'train_dataset_dir': {'type': str, 'default': None},
        'train_generator': {'type': None, 'default': None},
        'val_data_generator': {'type': None, 'default': None},
        'val_dataset_dataframe_path': {'type': str, 'default': None},
        'val_dataset_dir': {'type': str, 'default': None},
        'val_generator': {'type': None, 'default': None},
        'verbose': {'type': bool, 'default': False},
        'weights': {'type': str, 'default': 'imagenet'},
        'workers': {'type': int, 'default': 1},
    }

You will find a guide of use here