Aprendizaje de transferencia: el modelo está dando resultados de pérdida sin cambios. ¿No es entrenamiento? [cerrado]

Question

Estoy intentando entrenar un modelo de regresión en Inception V3. Las entradas son imágenes de tamaño (96,320,3). Hay un total de 16k + imágenes de las cuales 12k + son para entrenamiento y el resto para validación. He congelado todas las capas en Inception, pero descongelarlas tampoco ayuda (ya lo intenté). Reemplacé la parte superior del modelo previamente entrenado con algunas capas como se indica en el código a continuación.

X_train = preprocess_input(X_train)
inception = InceptionV3(weights='imagenet', include_top=False, input_shape=(299,299,3))
inception.trainable = False
print(inception.summary())

driving_input = Input(shape=(96,320,3))
resized_input = Lambda(lambda image: tf.image.resize(image,(299,299)))(driving_input)
inp = inception(resized_input)

x = GlobalAveragePooling2D()(inp)

x = Dense(512, activation = 'relu')(x)
x = Dense(256, activation = 'relu')(x)
x = Dropout(0.25)(x)
x = Dense(128, activation = 'relu')(x)
x = Dense(64, activation = 'relu')(x)
x = Dropout(0.25)(x)
result = Dense(1, activation = 'relu')(x)

lr_schedule = ExponentialDecay(initial_learning_rate=0.1, decay_steps=100000, decay_rate=0.95)
optimizer = Adam(learning_rate=lr_schedule)
loss = Huber(delta=0.5, reduction="auto", name="huber_loss")
model = Model(inputs = driving_input, outputs = result)
model.compile(optimizer=optimizer, loss=loss)

checkpoint = ModelCheckpoint(filepath="./ckpts/model.h5", monitor='val_loss', save_best_only=True)
stopper = EarlyStopping(monitor='val_loss', min_delta=0.0003, patience = 10)

batch_size = 32
epochs = 100

model.fit(x=X_train, y=y_train, shuffle=True, validation_split=0.2, epochs=epochs, 
          batch_size=batch_size, verbose=1, callbacks=[checkpoint, stopper])

Esto resulta en esto:

¿Por qué mi modelo no se está entrenando y qué puedo hacer para solucionarlo?

Answer 1

Como su problema es un problema de regresión, la activación de la última capa debería ser en linearlugar de relu. Además, la tasa de aprendizaje es demasiado alta, debería considerar reducirla de acuerdo con su configuración general. Aquí estoy mostrando una muestra de código con MNIST.

# data 
(xtrain, train_target), (xtest, test_target) = tf.keras.datasets.mnist.load_data()
# train_x, MNIST is gray scale, so in order to use it in pretrained weights , extending it to 3 axix
x_train = np.expand_dims(xtrain, axis=-1)
x_train = np.repeat(x_train, 3, axis=-1)
x_train = x_train.astype('float32') / 255
# prepare the label for regression model 
ytrain4 = tf.square(tf.cast(train_target, tf.float32))

# base model 
inception = InceptionV3(weights='imagenet', include_top=False, input_shape=(75,75,3))
inception.trainable = False

# inputs layer
driving_input = tf.keras.layers.Input(shape=(28,28,3))
resized_input = tf.keras.layers.Lambda(lambda image: tf.image.resize(image,(75,75)))(driving_input)
inp = inception(resized_input)

# top model 
x = GlobalAveragePooling2D()(inp)
x = Dense(512, activation = 'relu')(x)
x = Dense(256, activation = 'relu')(x)
x = Dropout(0.25)(x)
x = Dense(128, activation = 'relu')(x)
x = Dense(64, activation = 'relu')(x)
x = Dropout(0.25)(x)
result = Dense(1, activation = 'linear')(x)

# hyper-param
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(initial_learning_rate=0.0001, 
                                                             decay_steps=100000, decay_rate=0.95)
optimizer = tf.keras.optimizers.Adam(learning_rate=lr_schedule)
loss = tf.keras.losses.Huber(delta=0.5, reduction="auto", name="huber_loss")

# build models
model = tf.keras.Model(inputs = driving_input, outputs = result)
model.compile(optimizer=optimizer, loss=loss)

# callbacks
checkpoint = tf.keras.callbacks.ModelCheckpoint(filepath="./ckpts/model.h5", monitor='val_loss', save_best_only=True)
stopper = tf.keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=0.0003, patience = 10)

batch_size = 32
epochs = 10

# fit 
model.fit(x=x_train, y=ytrain4, shuffle=True, validation_split=0.2, epochs=epochs, 
          batch_size=batch_size, verbose=1, callbacks=[checkpoint, stopper])

Salida

1500/1500 [==============================] - 27s 18ms/step - loss: 5.2239 - val_loss: 3.6060
Epoch 2/10
1500/1500 [==============================] - 26s 17ms/step - loss: 3.5634 - val_loss: 2.9022
Epoch 3/10
1500/1500 [==============================] - 26s 17ms/step - loss: 3.0629 - val_loss: 2.5063
Epoch 4/10
1500/1500 [==============================] - 26s 17ms/step - loss: 2.7615 - val_loss: 2.3764
Epoch 5/10
1500/1500 [==============================] - 26s 17ms/step - loss: 2.5371 - val_loss: 2.1303
Epoch 6/10
1500/1500 [==============================] - 26s 17ms/step - loss: 2.3848 - val_loss: 2.1373
Epoch 7/10
1500/1500 [==============================] - 26s 17ms/step - loss: 2.2653 - val_loss: 1.9039
Epoch 8/10
1500/1500 [==============================] - 26s 17ms/step - loss: 2.1581 - val_loss: 1.9087
Epoch 9/10
1500/1500 [==============================] - 26s 17ms/step - loss: 2.0518 - val_loss: 1.7193
Epoch 10/10
1500/1500 [==============================] - 26s 17ms/step - loss: 1.9699 - val_loss: 1.8837