Keras: Accuracy Drops While Finetuning Inception

Question

I am having trouble fine tuning an Inception model with Keras.

I have managed to use tutorials and documentation to generate a model of fully connected top layers that classifies my dataset into their proper categories with an accuracy over 99% using bottleneck features from Inception.

import numpy as np
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential
from keras.layers import Dropout, Flatten, Dense
from keras import applications


# dimensions of our images.
img_width, img_height = 150, 150

#paths for saving weights and finding datasets
top_model_weights_path = 'Inception_fc_model_v0.h5'
train_data_dir = '../data/train2'
validation_data_dir = '../data/train2' 

#training related parameters?
inclusive_images = 1424
nb_train_samples = 1424
nb_validation_samples = 1424
epochs = 50
batch_size = 16


def save_bottlebeck_features():
    datagen = ImageDataGenerator(rescale=1. / 255)

    # build bottleneck features
    model = applications.inception_v3.InceptionV3(include_top=False, weights='imagenet', input_shape=(img_width,img_height,3))

    generator = datagen.flow_from_directory(
        train_data_dir,
        target_size=(img_width, img_height),
        batch_size=batch_size,
        class_mode='categorical',
        shuffle=False)

    bottleneck_features_train = model.predict_generator(
        generator, nb_train_samples // batch_size)

    np.save('bottleneck_features_train', bottleneck_features_train)

    generator = datagen.flow_from_directory(
        validation_data_dir,
        target_size=(img_width, img_height),
        batch_size=batch_size,
        class_mode='categorical',
        shuffle=False)

    bottleneck_features_validation = model.predict_generator(
        generator, nb_validation_samples // batch_size)

    np.save('bottleneck_features_validation', bottleneck_features_validation)

def train_top_model():
    train_data = np.load('bottleneck_features_train.npy')
    train_labels = np.array(range(inclusive_images))

    validation_data = np.load('bottleneck_features_validation.npy')
    validation_labels = np.array(range(inclusive_images))

    print('base size ', train_data.shape[1:])

    model = Sequential()
    model.add(Flatten(input_shape=train_data.shape[1:]))
    model.add(Dense(1000, activation='relu'))
    model.add(Dense(inclusive_images, activation='softmax'))
    model.compile(loss='sparse_categorical_crossentropy',
             optimizer='Adam',
             metrics=['accuracy'])

    proceed = True

    #model.load_weights(top_model_weights_path)

    while proceed:
        history = model.fit(train_data, train_labels,
              epochs=epochs,
              batch_size=batch_size)#,
              #validation_data=(validation_data, validation_labels), verbose=1)
        if history.history['acc'][-1] > .99:
            proceed = False

    model.save_weights(top_model_weights_path)


save_bottlebeck_features()
train_top_model()

Epoch 50/50 1424/1424 [==============================] - 17s 12ms/step - loss: 0.0398 - acc: 0.9909

I have also been able to stack this model on top of inception to create my full model and use that full model to successfully classify my training set.

from keras import Model
from keras import optimizers
from keras.callbacks import EarlyStopping

img_width, img_height = 150, 150

top_model_weights_path = 'Inception_fc_model_v0.h5'
train_data_dir = '../data/train2'
validation_data_dir = '../data/train2' 

#how many inclusive examples do we have?
inclusive_images = 1424
nb_train_samples = 1424
nb_validation_samples = 1424
epochs = 50
batch_size = 16

# build the complete network for evaluation
base_model = applications.inception_v3.InceptionV3(weights='imagenet', include_top=False, input_shape=(img_width,img_height,3))

top_model = Sequential()
top_model.add(Flatten(input_shape=base_model.output_shape[1:]))
top_model.add(Dense(1000, activation='relu'))
top_model.add(Dense(inclusive_images, activation='softmax'))

top_model.load_weights(top_model_weights_path)

#combine base and top model
fullModel = Model(input= base_model.input, output= top_model(base_model.output))

#predict with the full training dataset
results = fullModel.predict_generator(ImageDataGenerator(rescale=1. / 255).flow_from_directory(
        train_data_dir,
        target_size=(img_width, img_height),
        batch_size=batch_size,
        class_mode='categorical',
        shuffle=False))

inspection of the results from processing on this full model match the accuracy of the bottleneck generated fully connected model.

import matplotlib.pyplot as plt
import operator

#retrieve what the softmax based class assignments would be from results
resultMaxClassIDs = [ max(enumerate(result), key=operator.itemgetter(1))[0] for result in results]

#resultMaxClassIDs should be equal to range(inclusive_images) so we subtract the two and plot the log of the absolute value 
#looking for spikes that indicate the values aren't equal 
plt.plot([np.log(np.abs(x)+10) for x in (np.array(resultMaxClassIDs) - np.array(range(inclusive_images)))])

Here is the problem: When I take this full model and attempt to train it, Accuracy drops to 0 even though validation remains above 99%.

model2 = fullModel

for layer in model2.layers[:-2]:
    layer.trainable = False

# compile the model with a SGD/momentum optimizer
# and a very slow learning rate.
#model.compile(loss='binary_crossentropy', optimizer=optimizers.SGD(lr=1e-4, momentum=0.9),  metrics=['accuracy'])

model2.compile(loss='categorical_crossentropy',
             optimizer=optimizers.SGD(lr=1e-4, momentum=0.9), 
             metrics=['accuracy'])

train_datagen = ImageDataGenerator(rescale=1. / 255)

test_datagen = ImageDataGenerator(rescale=1. / 255)

train_generator = train_datagen.flow_from_directory(
    train_data_dir,
    target_size=(img_height, img_width),
    batch_size=batch_size,
    class_mode='categorical')

validation_generator = test_datagen.flow_from_directory(
    validation_data_dir,
    target_size=(img_height, img_width),
    batch_size=batch_size,
    class_mode='categorical')

callback = [EarlyStopping(monitor='acc', min_delta=0, patience=3, verbose=0, mode='auto', baseline=None)]
# fine-tune the model
model2.fit_generator(
    #train_generator,
    validation_generator,
    steps_per_epoch=nb_train_samples//batch_size,
    validation_steps = nb_validation_samples//batch_size,
    epochs=epochs,
    validation_data=validation_generator)

Epoch 1/50 89/89 [==============================] - 388s 4s/step - loss: 13.5787 - acc: 0.0000e+00 - val_loss: 0.0353 - val_acc: 0.9937

and it gets worse as things progress

Epoch 21/50 89/89 [==============================] - 372s 4s/step - loss: 7.3850 - acc: 0.0035 - val_loss: 0.5813 - val_acc: 0.8272

The only thing I could think of is that somehow the training labels are getting improperly assigned on this last train, but I've successfully done this with similar code using VGG16 before.

I have searched over the code trying to find a discrepancy to explain why a model making accurate predictions over 99% of the time drops its training accuracy while maintaining validation accuracy during fine tuning, but I can't figure it out. Any help would be appreciated.

Information about the code and environment:

Things that are going to stand out as weird, but are meant to be that way:

• There is only 1 image per class. This NN is intended to classify objects whose environmental and orientation conditions are controlled. Their is only one acceptable image for each class corresponding to the correct environmental and rotational situation.
• The test and validation set are the same. This NN is only ever designed to be used on the classes it is being trained on. The images it will process will be carbon copies of the class examples. It is my intent to overfit the model to these classes

I am using:

• Windows 10
• Python 3.5.6 under Anaconda client 1.6.14
• Keras 2.2.2
• Tensorflow 1.10.0 as the backend
• CUDA 9.0
• CuDNN 8.0

I have checked out:

1. Keras accuracy discrepancy in fine-tuned model
2. VGG16 Keras fine tuning: low accuracy
3. Keras: model accuracy drops after reaching 99 percent accuracy and loss 0.01
4. Keras inception v3 retraining and finetuning error
5. How to find which version of TensorFlow is installed in my system?

but they appear unrelated.

Answer 1

Note: Since your problem is a bit strange and difficult to debug without having your trained model and dataset, this answer is just a (best) guess after considering many things that may have could go wrong. Please provide your feedback and I will delete this answer if it does not work.

Since the inception_V3 contains BatchNormalization layers, maybe the problem is due to (somehow ambiguous or unexpected) behavior of this layer when you set trainable parameter to False (1, 2, 3, 4).

Now, let's see if this is the root of the problem: as suggested by @fchollet, set the learning phase when defining the model for fine-tuning:

from keras import backend as K

K.set_learning_phase(0)

base_model = applications.inception_v3.InceptionV3(weights='imagenet', include_top=False, input_shape=(img_width,img_height,3))

for layer in base_model.layers:
    layer.trainable = False

K.set_learning_phase(1)

top_model = Sequential()
top_model.add(Flatten(input_shape=base_model.output_shape[1:]))
top_model.add(Dense(1000, activation='relu'))
top_model.add(Dense(inclusive_images, activation='softmax'))

top_model.load_weights(top_model_weights_path)

#combine base and top model
fullModel = Model(input= base_model.input, output= top_model(base_model.output))

fullModel.compile(loss='categorical_crossentropy',
             optimizer=optimizers.SGD(lr=1e-4, momentum=0.9), 
             metrics=['accuracy'])


#####################################################################
# Here, define the generators and then fit the model same as before #
#####################################################################

---

Side Note: This is not causing any problem in your case, but keep in mind that when you use top_model(base_model.output) the whole Sequential model (i.e. top_model) is stored as one layer of fullModel. You can verify this by either using fullModel.summary() or print(fullModel.layers[-1]). Hence when you used:

for layer in model2.layers[:-2]:
    layer.trainable = False 

you are actually not freezing the last layer of base_model as well. However, since it is a Concatenate layer, and therefore does not have trainable parameters, no problem occurs and it would behave as you intended.

Answer 2

Like the previous reply, I'll try to share some thoughts to see whether it helps.

There are a couple of things that called my attention (and maybe are worth reviewing). Note: some of them should have given you issues with the separate models as well.

• Correct if I'm wrong, but it seems you used sparse_categorical_crossentropy for the first training while you used categorical_crossentropy for the second one. Is it correct? Because I believe they assume labels differently (sparse assumes integers and the other assumes one-hot).
• Have you tried to set the layers you added in the end as trainable = True? I know that you have already set the others to trainable = False, but maybe that's something worth checking too.
• It seems the data generator is not making use of the default preprocessing function used in Inception v3, which uses a per-mean channel.
• Have you tried any experiment using Functional instead of Sequential API?

I hope that helps.