CommonConvnetKeras

This repository intends to implement common convolutional neural networks for image classification.

Dependency

1. Tensorflow
2. Keras
3. Opencv

Structure

1. ConvnetTrainBase is the base class for all convnet implementation.
2. Each folder contains the implementation of convnet and their training configuration

Usage

1. training:

train_data_dir = "/path/to/training/dataset"
validation_data_dir = "/path/to/validation/dataset"
test_data_dir = "/path/to/test/dataset"
nb_train_samples = 2000
nb_validation_samples = 600

dir_path = os.path.dirname(os.path.realpath(__file__))


def experiment_1():

    resnet = Resnet34Train(
        train_name = "exp_1",
        train_data_dir = train_data_dir,
        validation_data_dir = validation_data_dir,
        test_data_dir = test_data_dir,
        nb_train_samples = nb_train_samples,
        nb_validation_samples = nb_validation_samples,
        nb_epoch = 100
    )

    # binary
    resnet.set_complete_model(1)

    optimizer = "rmsprop"
    resnet.train_from_scratch_binary(
        optimizer=optimizer
    )

    return

2. Prediction

def experiment_1():

    densenet_fast = DensenetFastTrain(
        train_name = "exp_1",
        train_data_dir = train_data_dir,
        validation_data_dir = validation_data_dir,
        test_data_dir = test_data_dir,
        nb_train_samples = nb_train_samples,
        nb_validation_samples = nb_validation_samples,
        nb_epoch = 100,
        model_weight_folder = dir_path,
    )

    # predict
    final_model = os.path.join(dir_path, "densenet_fast_exp_1.h5")
    densenet_fast.predict_with_final_model(final_model)

TODO

1. training flow