train3.py

import model_q2
from keras.optimizers import Adam
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import ModelCheckpoint
import os

img_width, img_height = 224, 224
train_data_dir = '/data/datasets/rbonatti/data_processed/2/train'
validation_data_dir = '/data/datasets/rbonatti/data_processed/2/train20'
save_models_dir= '/data/datasets/rbonatti/ml_weights3/'
# nb_train_samples = 7998*2
nb_train_samples = 1000*64
nb_validation_samples = 1000*4
epochs = 200
batch_size = 32

if __name__ == "__main__":

    # Test pretrained model
    network = model_q2.VGG_16('/data/datasets/rbonatti/vgg16_weights_with_name.h5')
    adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
    network.compile(optimizer=adam, loss='categorical_crossentropy', metrics=['accuracy'])

    # prepare data augmentation configuration
    datagen = ImageDataGenerator(rescale=1. / 255,zoom_range=0.7)

    train_generator = datagen.flow_from_directory(
        train_data_dir,
        target_size=(img_height, img_width),
        batch_size=batch_size,
        class_mode='categorical',
        color_mode='grayscale'
        )

    validation_generator = datagen.flow_from_directory(
        validation_data_dir,
        target_size=(img_height, img_width),
        batch_size=batch_size,
        class_mode='categorical',
        color_mode='grayscale')

    # define callbacks
    # cleanup_callback = LambdaCallback(on_train_end=lambda logs: [p.terminate() for p in processes if p.is_alive()])
    # plot_loss_callback = LambdaCallback(on_epoch_end=lambda epoch, logs: plt.plot(np.arange(epoch),logs['loss']))
    # batch_print_callback = LambdaCallback(on_batch_begin=lambda batch,logs: print(batch))
    # saves the model weights after each epoch if the validation loss decreased
    checkpointer = ModelCheckpoint(filepath="/data/datasets/rbonatti/ml_weights3/weights.{epoch:02d}-{val_loss:.2f}.hdf5", verbose=1)

    # folder to save the weights
    if not os.path.exists(save_models_dir):
        os.makedirs(save_models_dir)

    # fine-tune the model
    network.fit_generator(
        generator=train_generator,
        samples_per_epoch=nb_train_samples,
        nb_epoch=epochs,
        validation_data=validation_generator,
        nb_val_samples=nb_validation_samples,
        callbacks=[checkpointer]
        )