def trainable_model(non_trainable_model):
'''
Parameters
----------
non_trainable_model: This is the non_trianable base network whose output features are used to trian this
network.
Returns
-------
trainable_model: The fine_tuning model object that we will train.
'''
input_tensor = Input(shape=non_trainable_model.output_shape[1:])
# input_tensor = Input(shape = non_trainable_model.get_layer(BOTTLENECK_TENSOR_NAME).output_shape[1:])
# weights = non_trainable_model.get_weights()
trainable_model = transfer_model(input_tensor, LABEL_LENGTH)
# trainable_model.summary()
return (trainable_model)
Returns
------
predictions: The probability outputs of the network for the given image
'''
feature_extracted = non_trainable_model.predict(img)
print(feature_extracted.shape)
predictions = trainable_model.predict(feature_extracted)
class_name = label_map[np.argmax(predictions[0])]
if want_probability:
return (class_name, predictions[0])
else:
return (class_name, predictions[0][np.argmax(predictions[0])])
# base_model = applications.ResNet50(weights = 'imagenet', include_top = False, input_shape = (256,256,3))
input_tensor = Input(
shape=base_model.get_layer(BOTTLENECK_TENSOR_NAME).output_shape[1:])
trainable_model = transfer_model(input_tensor, LABEL_LENGTH)
trainable_model = load_model(args.weight_file)
non_trainable_model = create_non_trainable_model(base_model,
BOTTLENECK_TENSOR_NAME)
print(trainable_model.input_shape)
for img_addr in natsorted(glob(args.img_dir + "/*")):
img = cv2.imread(img_addr)
img = cv2.resize(img, (256, 256))
img = np.expand_dims(img, axis=0)
(class_name, confidence) = predict_complete_output(non_trainable_model,
trainable_model, img,
inv_label_map)
print((class_name, confidence))
# input_tensor = Input(shape = non_trainable_model.get_layer(BOTTLENECK_TENSOR_NAME).output_shape[1:])
# weights = non_trainable_model.get_weights()
trainable_model = transfer_model(input_tensor, LABEL_LENGTH)
# trainable_model.summary()
return (trainable_model)
# for i in (base_model.layers):
# print (i.name)
# exit(0)
non_trainable_model = create_non_trainable_model(base_model,
BOTTLENECK_TENSOR_NAME)
trainable_model = trainable_model(non_trainable_model)
non_bottleneck_net = Model(inputs=non_trainable_model.input,
output=transfer_model(non_trainable_model.output,
LABEL_LENGTH,
bottleneck_used=False))
non_bottleneck_net.compile(loss='categorical_crossentropy',
optimizer=optimizers.SGD(lr=1e-2, momentum=0.9),
metrics=['accuracy'])
# Train data generator
train_datagen = ImageDataGenerator(preprocessing_function=preprocess_input)
# Callback list
checkpoint = ModelCheckpoint(args.weight_file)
tb_callback = keras.callbacks.TensorBoard(log_dir=args.logs, write_graph=True)
# early_stopping = EarlyStopping(monitor = 'val_loss')
callback_list = [checkpoint, tb_callback] #, early_stopping]
h1 = train_datagen.flow_from_directory(args.train, batch_size=BATCH_SIZE)