Train CNNs for the fruits360 data set in NTOU CS「Machine Vision」class.
Build a CNN on a pretrained model, ResNet50.
Finetune the pretrained model when training my CNN.
def fruit_model_on_pretrained(height,width,channel):
model = Sequential(name="fruit_pretrained")
pretrained = tf.keras.applications.resnet.ResNet50(include_top=False,input_shape=(100,100,3))
model.add(pretrained)
model.add(tf.keras.layers.GlobalAveragePooling2D())
model.add(Dense(16, activation='relu'))
model.add(Dense(16, activation='relu'))
model.add(Dense(2,activation='softmax'))
pretrained.trainable = False
model.compile(loss=tf.keras.losses.SparseCategoricalCrossentropy(),optimizer='adam', metrics=['accuracy'])
return model
model = fruit_model_on_pretrained(100,100,3)
model.summary()Build a CNN whose neural architecture includes ResBlock.
images = keras.layers.Input(x_train.shape[1:])
x = keras.layers.Conv2D(filters=16, kernel_size=[1,1], padding='same')(images)
block = keras.layers.Conv2D(filters=16, kernel_size=[3,3], padding="same")(x)
block = keras.layers.BatchNormalization()(block)
block = keras.layers.Activation("relu")(block)
block = keras.layers.Conv2D(filters=16, kernel_size=[3,3],padding="same")(block)
net = keras.layers.add([x,block])
net = keras.layers.BatchNormalization()(net)
net = keras.layers.Activation("relu")(net)
net = keras.layers.MaxPooling2D(pool_size=(2,2),name="block_1")(net)
x = keras.layers.Conv2D(filters=32, kernel_size=[1,1], padding='same')(net)
block = keras.layers.Conv2D(filters=32, kernel_size=[3,3], padding="same")(x)
block = keras.layers.BatchNormalization()(block)
block = keras.layers.Activation("relu")(block)
block = keras.layers.Conv2D(filters=32, kernel_size=[3,3],padding="same")(block)
net = keras.layers.add([x,block])net=keras.layers.BatchNormalization()(net)
net = keras.layers.Activation("relu")(net)
net = keras.layers.MaxPooling2D(pool_size=(2,2),name="block_2")(net)
x = keras.layers.Conv2D(filters=64, kernel_size=[1,1], padding='same')(net)
block = keras.layers.Conv2D(filters=64, kernel_size=[3,3], padding="same")(x)
block = keras.layers.BatchNormalization()(block)
block = keras.layers.Activation("relu")(block)
block = keras.layers.Conv2D(filters=64, kernel_size=[3,3],padding="same")(block)
net = keras.layers.add([x,block])
net = keras.layers.Activation("relu", name="block_3")(net)
net = keras.layers.BatchNormalization()(net)
net = keras.layers.Dropout(0.25)(net)
net = keras.layers.GlobalAveragePooling2D()(net)
net = keras.layers.Dense(units=nclasses,activation="softmax")(net)
model = keras.models.Model(inputs=images,outputs=net)
model.summary()This package is MIT licensed.