Transfer learning bad accuracy

Question

I have a task to classify seeds depending on the defect. I have around 14k images in 7 classes (they are not equal size, some classes have more photos, some have less). I tried to train Inception V3 from scratch and I've got around 90% accuracy. Then I tried transfer learning using pre-trained model with ImageNet weights. I imported inception_v3 from applications without top fc layers, then added my own like in documentation. I ended with the following code:

# Setting dimensions
img_width = 454
img_height = 227

###########################
# PART 1 - Creating Model #
###########################

# Creating InceptionV3 model without Fully-Connected layers
base_model = InceptionV3(weights='imagenet', include_top=False, input_shape = (img_height, img_width, 3))

# Adding layers which will be fine-tunned
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(7, activation='softmax')(x)

# Creating final model
model = Model(inputs=base_model.input, outputs=predictions)

# Plotting model
plot_model(model, to_file='inceptionV3.png')

# Freezing Convolutional layers
for layer in base_model.layers:
    layer.trainable = False

# Summarizing layers
print(model.summary())

# Compiling the CNN
model.compile(optimizer = 'adam', loss = 'categorical_crossentropy', metrics = ['accuracy'])

##############################################
# PART 2 - Images Preproccessing and Fitting #
##############################################

# Fitting the CNN to the images

train_datagen = ImageDataGenerator(rescale = 1./255,
                                   rotation_range=30,
                                   width_shift_range=0.2,
                                   height_shift_range=0.2,
                                   shear_range = 0.2,
                                   zoom_range = 0.2,
                                   horizontal_flip = True,
                                   preprocessing_function=preprocess_input,)

valid_datagen = ImageDataGenerator(rescale = 1./255,
                                   preprocessing_function=preprocess_input,)

train_generator = train_datagen.flow_from_directory("dataset/training_set",
                                                    target_size=(img_height, img_width),
                                                    batch_size = 4,
                                                    class_mode = "categorical",
                                                    shuffle = True,
                                                    seed = 42)

valid_generator = valid_datagen.flow_from_directory("dataset/validation_set",
                                                    target_size=(img_height, img_width),
                                                    batch_size = 4,
                                                    class_mode = "categorical",
                                                    shuffle = True,
                                                    seed = 42)

STEP_SIZE_TRAIN = train_generator.n//train_generator.batch_size
STEP_SIZE_VALID = valid_generator.n//valid_generator.batch_size

# Save the model according to the conditions  
checkpoint = ModelCheckpoint("inception_v3_1.h5", monitor='val_acc', verbose=1, save_best_only=True, save_weights_only=False, mode='auto', period=1)
early = EarlyStopping(monitor='val_acc', min_delta=0, patience=10, verbose=1, mode='auto')

#Training the model
history = model.fit_generator(generator=train_generator,
                         steps_per_epoch=STEP_SIZE_TRAIN,
                         validation_data=valid_generator,
                         validation_steps=STEP_SIZE_VALID,
                         epochs=25,
                         callbacks = [checkpoint, early])

But I've got terrible results: 45% accuracy. I thought it should be better. I have some hypothesis what could go wrong:

• I trained from scratch on scaled images (299x299) and on non-scaled while transfer-learning (227x454) and it failed something (or maybe I failed dimensions order).
• While transfer-learning I used preprocessing_function=preprocess_input (found article on the web that it is extremely important, so I decided to add that).
• Added rotation_range=30, width_shift_range=0.2, height_shift_range=0.2, and horizontal_flip = True while transfer learning to augment data even more.
• Maybe Adam optimizer is a bad idea? Should I try RMSprop for example?
• Should I fine-tune some conv layers with SGD with small learning rate too?

Or did I failed something else?

EDIT: I post a plot of training history. Maybe it contains valuable information:

EDIT2: With changing parameters of InceptionV3:

VGG16 for comparison:

Answer 1

@today, I found a problem. It is because of some changes in Batch Normalisation layers and its behavior while freezing them. Mr. Chollet gave a workaround, but I used a Keras fork made by datumbox, which solved my problem. The main problem is described here:

https://github.com/keras-team/keras/pull/9965

Now I get ~85% accuracy and am trying to raise it.

Answer 2

If you want to preprocess the input using the preprocess_input method from Keras, then remove the rescale=1./255 argument. Otherwise, keep the rescale argument and remove the preprocessing_function argument. Plus, try a lower learning rate like 1e-4 or 3e-5 or 1e-5 (the default learning rate of Adam optimizer is 1e-3) if loss does not decrease:

from keras.optimizers import Adam

model.compile(optimizer = Adam(lr=learning_rate), ...)

Edit: After adding the training plot, you can see that is it overfitting on the training set. You can:

• add some kind of regularization like a Dropout layer,
• or decrease the network size by lowering the number of units in the Dense layer which is before the last layer.