# opt = keras.optimizers.Adadelta(lr=1.0, rho=0.95, epsilon=None, decay=0.0)
opt = keras.optimizers.Adamax(lr=0.002,
beta_1=0.9,
beta_2=0.999,
epsilon=None,
decay=0.0)
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
log_path = '/tmp/tflearn_logs/NASNetMobile_LCZ42_Adadelta'
callback = TensorBoard(log_path)
callback.set_model(model)
model.fit(
x_train,
y_train,
batch_size=1024, # 128,1024
epochs=10,
shuffle="batch",
validation_data=(x_test, y_test))
modelpath = 'NASNetMobile_Adadelta_epochs_10.h5'
model.save(modelpath)
print('Saved trained model at %s ' % modelpath)
# Score trained model.
scores = model.evaluate(x_test, y_test, verbose=1)
print('Test loss:', scores[0])
print('Test accuracy:', scores[1])
# This callback will log model stats to Tensorboard.
tb_callback = TensorBoard()
# This callback will checkpoint the best model at every epoch.
mc_callback = ModelCheckpoint(filepath='current_best.hdf5',
verbose=1,
save_best_only=True)
# This is the train DataSequence.
train_sequence = DataSequence(train_pd, "./images", batch_size=batch_size)
train_steps = len(train_pd) // batch_size
# This is the validation DataSequence.
validation_sequence = DataSequence(test_pd, "./images", batch_size=batch_size)
validation_steps = len(test_pd) // batch_size
# These are the callbacks.
callbacks = [lr_callback, tb_callback, mc_callback]
# This line will train the model.
model.fit_generator(train_sequence,
validation_data=validation_sequence,
epochs=20,
use_multiprocessing=True,
workers=80,
steps_per_epoch=train_steps,
validation_steps=validation_steps,
callbacks=callbacks)
# Finally, we save the model.
model.save(MODEL_NAME)