simple_classifier.py

from keras import backend as K
from keras.applications.vgg16 import VGG16
from keras.layers import GlobalAveragePooling2D, Flatten, Dense, Input, Dropout
from keras.models import Model
from keras.optimizers import Adam
from keras.preprocessing.image import ImageDataGenerator

#setup
path = 'data/dogscats_redux/'
#path = 'data/dogscats_redux/sample/'
target_size = (224, 224)
batch_size = 64

#define batches
def get_batches(directory, target_size=target_size, batch_size=batch_size):
  datagen = ImageDataGenerator()
  return datagen.flow_from_directory(directory=directory,
                                    target_size=target_size,
                                    batch_size=batch_size,
                                    class_mode='categorical')

batches = get_batches(path + 'train')
valid_batches = get_batches(path + 'valid', batch_size=batch_size*2)

# fine-tuning
# include_top=False to remove the last layer
initial_model = VGG16(weights="imagenet", include_top=False, input_tensor = Input(shape=(3,224,224)))
x = Flatten()(initial_model.output)

# we add a Dense layer
x = Dense(batches.nb_class, activation='softmax')(x)

# new model with "frozen" layers (except the last one)
model = Model(initial_model.input, x)
for layer in initial_model.layers: layer.trainable=False

model.compile(optimizer=Adam(lr=0.001),
              loss='categorical_crossentropy', 
              metrics=['accuracy'])

model.fit_generator(batches, samples_per_epoch=batches.nb_sample, nb_epoch=1,
                    validation_data=valid_batches, nb_val_samples=valid_batches.nb_sample)

# predictions on new data
test_batches = get_batches(path+'test', batch_size=batch_size*2)
preds = model.predict_generator(test_batches, test_batches.nb_sample)