Transfer learning with tf.keras and Inception-v3: No training is happening

Question

I am attempting to training a model based on a frozen Inception_v3 model with 3 classes as an output. When I run the training, training accuracy goes up but not validation accuracy which is more or less exactly at 33.33% i.e. showing completely random prediction. I can't figure where is the bug in my code and/or approach

I tried various form of output after the Inception v3 core with no differences at all.

# Model definition
# InceptionV3 frozen, flatten, dense 1024, dropout 50%, dense 1024, dense 3, lr 0.001 --> does not train
# InceptionV3 frozen, flatten, dense 1024, dense 3, lr 0.001 --> does not train
# InceptionV3 frozen, flatten, dense 1024, dense 3, lr 0.005 --> does not train
# InceptionV3 frozen, GlobalAvgPooling, dense 1024, dense 1024, dense 512, dense 3, lr 0.001 --> does not train
# InceptionV3 frozen, GlobalAvgPooling dropout 0.4 dense 3, lr 0.001, custom pre-process --> does not train
# InceptionV3 frozen, GlobalAvgPooling dropout 0.4 dense 3, lr 0.001, custom pre-process, batch=32 --> does not train
# InceptionV3 frozen, GlobalAvgPooling dropout 0.4 dense 3, lr 0.001, custom pre-process, batch=32, rebalance train/val sets --> does not train

IMAGE_SIZE = 150
BATCH_SIZE = 32

def build_model(image_size):
  input_tensor = tf.keras.layers.Input(shape=(image_size, image_size, 3))

  inception_base = InceptionV3(include_top=False, weights='imagenet', input_tensor=input_tensor)
  for layer in inception_base.layers:
    layer.trainable = False

  x = inception_base.output
  x = tf.keras.layers.GlobalAveragePooling2D()(x)
  x = tf.keras.layers.Dropout(0.2)(x)
  output_tensor = tf.keras.layers.Dense(3, activation="softmax")(x)

  model = tf.keras.Model(inputs=input_tensor, outputs=output_tensor)

  return model

model = build_model(IMAGE_SIZE)
model.compile(optimizer=RMSprop(lr=0.002), loss='categorical_crossentropy', metrics=['acc'])

# Data generators with Image augmentations
train_datagen = ImageDataGenerator(
      rescale=1./255,
      preprocessing_function=tf.keras.applications.inception_v3.preprocess_input,
      rotation_range=40,
      width_shift_range=0.2,
      height_shift_range=0.2,
      shear_range=0.2,
      zoom_range=0.2,
      horizontal_flip=True,
      fill_mode='nearest')

# Do not augment validation!
validation_datagen = ImageDataGenerator(
    rescale=1./255,
    preprocessing_function=tf.keras.applications.inception_v3.preprocess_input)

train_generator = train_datagen.flow_from_directory(
      train_dir,
      target_size=(IMAGE_SIZE, IMAGE_SIZE),
      batch_size=BATCH_SIZE,
      class_mode='categorical')

validation_generator = validation_datagen.flow_from_directory(
      valid_dir,
      target_size=(IMAGE_SIZE, IMAGE_SIZE),
      batch_size=BATCH_SIZE,
      class_mode='categorical')

Output of this cell is:

Found 1697 images belonging to 3 classes. Found 712 images belonging to 3 classes.

Output of last two epochs of training:

Epoch 19/20
23/23 [==============================] - 6s 257ms/step - loss: 1.1930 - acc: 0.3174
54/54 [==============================] - 20s 363ms/step - loss: 0.7870 - acc: 0.6912 - val_loss: 1.1930 - val_acc: 0.3174
Epoch 20/20
23/23 [==============================] - 6s 255ms/step - loss: 1.1985 - acc: 0.3160
54/54 [==============================] - 20s 362ms/step - loss: 0.7819 - acc: 0.7018 - val_loss: 1.1985 - val_acc: 0.3160

Answer 1

The only big thing that jumps out at me is to ditch the rescale=1./255 ImageDataGenerators, because this is also being handled by tf.keras.applications.inception_v3.preprocess_input, which scales the from -1 to 1; the network's expected input.