keras 模型上的迁移学习总是给出相同的预测

Question

I'm trying to train an image classifier using keras applications module. When I run predictions on validation set, all images are predicted as the same class. It is not always the same class, it varies during training. I'm using MobileNetV2 with weights from ImageNet but I also tried other models with same result.

I've tried using model from TensorFlow hub like described in this tutorial:https://www.tensorflow.org/beta/tutorials/images/hub_with_keras and it worked fine, so it is not a data set issue.

My code snippet:

image_size = 224
batch_size = 32
train_datagen = tf.keras.preprocessing.image.ImageDataGenerator(
    preprocessing_function=tf.keras.applications.mobilenet_v2.preprocess_input)
validation_datagen = tf.keras.preprocessing.image.ImageDataGenerator(
    preprocessing_function=tf.keras.applications.mobilenet_v2.preprocess_input)
train_generator = train_datagen.flow_from_directory(training_data_dir,
                                                    target_size=(image_size, image_size),
                                                    batch_size=batch_size)
validation_generator = train_datagen.flow_from_directory(validation_data_dir,
                                                         target_size=(image_size, image_size),
                                                         batch_size=batch_size)

IMG_SHAPE = (image_size, image_size, 3)
base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,
                                               include_top=False,
                                               weights="imagenet")
base_model.trainable = False

model = tf.keras.Sequential([
    base_model,
    tf.keras.layers.GlobalAveragePooling2D(),
    tf.keras.layers.Dense(train_generator.num_classes, activation='softmax')
])
model.compile(optimizer=tf.keras.optimizers.RMSprop(lr=0.001),
              loss="categorical_crossentropy",
              metrics=["accuracy"])
model.summary()

batch_stats = CollectBatchStats()
epoch_stats = CollectEpochStats(model, validation_generator)
checkpoint = tf.keras.callbacks.ModelCheckpoint(...)

epochs = 10
steps_per_epoch = train_generator.n // train_generator.batch_size
validation_steps = validation_generator.n // validation_generator.batch_size

history = model.fit_generator(train_generator,
                              epochs=epochs,
                              steps_per_epoch=steps_per_epoch,
                              callbacks=[batch_stats, epoch_stats, checkpoint],
                              workers=4,
                              validation_data=validation_generator,
                              validation_steps=validation_steps)

Answer 1

Issue resolved: in my code I had following lines after model compilation:

sess = keras_backend.get_session()
init = tf.compat.v1.global_variables_initializer()
sess.run(init)

After removing them everything works fine.