VGG16 Keras fine tuning: low accuracy

Question

I've already asked similar question here, but now I have slightly different problem, therefore asking new question.

I decided to use slightly different approach instead of proposed among answers in the referenced question to train, and then fine-tune model.

Update: I've replaced old question provided here with more suitable version

Here is my sequence of actions:

1. Build VGG16 model and drop top layer (call it no-top model)
2. Generate bottleneck features using no-top model
3. Train a separate fully-connected model using bottleneck features
4. Build new VGG16 model, drop top layers, and attach pretrained top-model
5. Train concatenated model on dogs/cats data

And here is a code I use to implement aforementioned sequence of actions:

import warnings
warnings.simplefilter('ignore', UserWarning)
warnings.simplefilter('ignore', DeprecationWarning)

from __future__ import print_function
from itertools import izip_longest as zip_longest
from pprint import pformat as pf
from pprint import pprint as pp
import os

from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img
from keras.layers import Conv2D, MaxPooling2D, MaxPooling2D, ZeroPadding2D
from keras.layers import Dropout, Flatten, Dense, InputLayer, Lambda
from keras.models import Sequential, Model, load_model
from keras.utils.data_utils import get_file
from keras.optimizers import SGD
import keras.backend as K

import numpy as np


RANDOM_STATE = 1
IMAGE_WIDTH = 224
IMAGE_HEIGHT = 224
BATCH_SIZE = 4
VGG_MEAN = np.array([123.68, 116.779, 103.939]).reshape((3, 1, 1))
VGG16_WEIGHTS_PATH = 'http://www.platform.ai/models/vgg16.h5'
DATA_ROOT = os.path.join(os.path.expanduser('~'), 'data', 'dogscats')
TRAIN_DIR = os.path.join(DATA_ROOT, 'train')
VALID_DIR = os.path.join(DATA_ROOT, 'valid')
SAMPLES_DIR = os.path.expanduser('~/dogscats_samples')


np.random.seed(RANDOM_STATE)
K.set_image_dim_ordering('th')


def get_batches(dirname, gen=ImageDataGenerator(), shuffle=True, 
                batch_size=BATCH_SIZE, class_mode='categorical'):
    return gen.flow_from_directory(
        os.path.join(SAMPLES_DIR, dirname),
        target_size=(IMAGE_WIDTH, IMAGE_HEIGHT),
        class_mode=class_mode,
        shuffle=shuffle,
        batch_size=batch_size)

def vgg_preprocess(x):
    x = x - VGG_MEAN
    return x[:, ::-1]

def conv_block(model, n_layers, n_filters, name='block'):
    for i in range(n_layers):
        model.add(ZeroPadding2D((1, 1), name='%s_padding_%s' % (name, i)))
        model.add(Conv2D(n_filters, (3, 3), activation='relu', name='%s_conv2d_%s' % (name, i)))
    model.add(MaxPooling2D((2, 2), strides=(2, 2), name='%s_maxpool' % name))

def fc_block(model, name='block'):
    model.add(Dense(4096, activation='relu', name=name + '_dense'))
    model.add(Dropout(0.5))

def build_vgg_16():
    model = Sequential()
    input_shape = (3, IMAGE_WIDTH, IMAGE_HEIGHT) 
    model.add(InputLayer(input_shape=input_shape))
    model.add(Lambda(vgg_preprocess))
    conv_block(model, n_layers=2, n_filters=64, name='block1')
    conv_block(model, n_layers=2, n_filters=128, name='block2')
    conv_block(model, n_layers=3, n_filters=256, name='block3')
    conv_block(model, n_layers=3, n_filters=512, name='block4')
    conv_block(model, n_layers=3, n_filters=512, name='block5')
    model.add(Flatten())
    fc_block(model)
    fc_block(model)
    model.add(Dense(1000, activation='softmax'))
    return model

def train_finetuned_model():
    file_path = get_file('vgg16.h5', VGG16_WEIGHTS_PATH, cache_subdir='models')
    print('Building VGG16 (no-top) model to generate bottleneck features')
    vgg16_notop = build_vgg_16()
    vgg16_notop.load_weights(file_path)
    for _ in range(6):
        vgg16_notop.pop()
    vgg16_notop.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])

    train_batches = get_batches('train', shuffle=False, class_mode=None)
    train_labels = np.array([0]*1000 + [1]*1000)
    bottleneck_train = vgg16_notop.predict_generator(train_batches, steps=2000 // BATCH_SIZE)

    valid_batches = get_batches('valid', shuffle=False, class_mode=None)
    valid_labels = np.array([0]*400 + [1]*400)
    bottleneck_valid = vgg16_notop.predict_generator(valid_batches, steps=800 // BATCH_SIZE)

    print('Training top model on bottleneck features')
    top_model = Sequential()
    top_model.add(Flatten(input_shape=bottleneck_train.shape[1:]))
    top_model.add(Dense(256, activation='relu'))
    top_model.add(Dropout(0.5))
    top_model.add(Dense(1, activation='sigmoid'))
    top_model.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'])
    top_model.fit(bottleneck_train, train_labels, 
                  batch_size=32, epochs=50, 
                  validation_data=(bottleneck_valid, valid_labels))

    print('Concatenate new VGG16 (without top layer) with pretrained top model')
    vgg16_fine = build_vgg_16()    
    vgg16_fine.load_weights(file_path)
    for _ in range(6):
        vgg16_fine.pop()
    vgg16_fine.add(Flatten(name='top_flatten'))    
    vgg16_fine.add(Dense(256, activation='relu', name='top_dense'))
    vgg16_fine.add(Dropout(0.5, name='top_dropout'))
    vgg16_fine.add(Dense(1, activation='sigmoid', name='top_sigmoid'))
    for i, layer in enumerate(reversed(top_model.layers), 1):
        pretrained_weights = layer.get_weights()
        vgg16_fine.layers[-i].set_weights(pretrained_weights)
    for layer in vgg16_fine.layers[:26]:
        layer.trainable = False
    vgg16_fine.compile(optimizer=SGD(lr=1e-4, momentum=0.9),
                       loss='binary_crossentropy',
                       metrics=['accuracy'])

    print('Train concatenated model on dogs/cats dataset sample')
    train_datagen = ImageDataGenerator(rescale=1./255,
                                       shear_range=0.2,
                                       zoom_range=0.2,
                                       horizontal_flip=True)
    test_datagen = ImageDataGenerator(rescale=1./255)
    train_batches = get_batches('train', gen=train_datagen, class_mode='binary')
    valid_batches = get_batches('valid', gen=test_datagen, class_mode='binary')
    vgg16_fine.fit_generator(train_batches,
                             steps_per_epoch=2000 // BATCH_SIZE,
                             epochs=50,
                             validation_data=valid_batches,
                             validation_steps=800 // BATCH_SIZE)
    return vgg16_fine


final_model = train_finetuned_model()

But the problem is that model's accuracy drastically dropped. After 50 epochs, its accuracy is around 50%. Therefore, probably I've done something wrong.

Maybe something wrong with parameters, i.e. learning rate, batch size, etc.?

Answer 1

Your fully connected layers look totally different from the original VGG architecture.

# yours
Flatten()
Dense(256, activation='relu')
Dense(1, activation='sigmoid')

# original
Flatten()
Dense(4096, activation='relu')
Dense(4096, activation='relu')
Dense(2, activation='softmax')

Two points.

1. The last layer should be 2-class-softmax instead of sigmoid. The accuracy is not computed as you expect if you use sigmoid, I guess.

2. Complexity (number of neurons and layers) seems to be too low.

Answer 2

Well, not sure if it is a right solution, but I was able to increase accuracy at least up to 70% with this code (probably the main reason is decreased learning rate and more epochs):

def train_finetuned_model(lr=1e-5, verbose=True):
    file_path = get_file('vgg16.h5', VGG16_WEIGHTS_PATH, cache_subdir='models')
    if verbose:
        print('Building VGG16 (no-top) model to generate bottleneck features.')

    vgg16_notop = build_vgg_16()
    vgg16_notop.load_weights(file_path)
    for _ in range(6):
        vgg16_notop.pop()
    vgg16_notop.compile(optimizer=RMSprop(lr=lr), loss='categorical_crossentropy', metrics=['accuracy'])    

    if verbose:
        print('Bottleneck features generation.')

    train_batches = get_batches('train', shuffle=False, class_mode=None, batch_size=BATCH_SIZE)
    train_labels = np.array([0]*1000 + [1]*1000)
    train_bottleneck = vgg16_notop.predict_generator(train_batches, steps=2000 // BATCH_SIZE)
    valid_batches = get_batches('valid', shuffle=False, class_mode=None, batch_size=BATCH_SIZE)
    valid_labels = np.array([0]*400 + [1]*400)
    valid_bottleneck = vgg16_notop.predict_generator(valid_batches, steps=800 // BATCH_SIZE)

    if verbose:
        print('Training top model on bottleneck features.')

    top_model = Sequential()
    top_model.add(Flatten(input_shape=train_bottleneck.shape[1:]))
    top_model.add(Dense(4096, activation='relu'))
    top_model.add(Dropout(0.5))
    top_model.add(Dense(4096, activation='relu'))
    top_model.add(Dropout(0.5))
    top_model.add(Dense(2, activation='softmax'))
    top_model.compile(optimizer=RMSprop(lr=lr), loss='categorical_crossentropy', metrics=['accuracy'])
    top_model.fit(train_bottleneck, to_categorical(train_labels),
                  batch_size=32, epochs=10,
                  validation_data=(valid_bottleneck, to_categorical(valid_labels)))

    if verbose:
        print('Concatenate new VGG16 (without top layer) with pretrained top model.')

    vgg16_fine = build_vgg_16()
    vgg16_fine.load_weights(file_path)
    for _ in range(6):
        vgg16_fine.pop()
    vgg16_fine.add(Flatten(name='top_flatten'))    
    vgg16_fine.add(Dense(4096, activation='relu'))
    vgg16_fine.add(Dropout(0.5))
    vgg16_fine.add(Dense(4096, activation='relu'))
    vgg16_fine.add(Dropout(0.5))
    vgg16_fine.add(Dense(2, activation='softmax'))
    vgg16_fine.compile(optimizer=RMSprop(lr=lr), loss='categorical_crossentropy', metrics=['accuracy'])

    if verbose:
        print('Loading pre-trained weights into concatenated model')

    for i, layer in enumerate(reversed(top_model.layers), 1):
        pretrained_weights = layer.get_weights()
        vgg16_fine.layers[-i].set_weights(pretrained_weights)

    for layer in vgg16_fine.layers[:26]:
        layer.trainable = False

    if verbose:
        print('Layers training status:')
        for layer in vgg16_fine.layers:
            print('[%6s] %s' % ('' if layer.trainable else 'FROZEN', layer.name))        

    vgg16_fine.compile(optimizer=RMSprop(lr=1e-6), loss='binary_crossentropy', metrics=['accuracy'])

    if verbose:
        print('Train concatenated model on dogs/cats dataset sample.')

    train_datagen = ImageDataGenerator(rescale=1./255,
                                       shear_range=0.2,
                                       zoom_range=0.2,
                                       horizontal_flip=True)
    test_datagen = ImageDataGenerator(rescale=1./255)
    train_batches = get_batches('train', gen=train_datagen, class_mode='categorical', batch_size=BATCH_SIZE)
    valid_batches = get_batches('valid', gen=test_datagen, class_mode='categorical', batch_size=BATCH_SIZE)
    vgg16_fine.fit_generator(train_batches, epochs=100,
                             steps_per_epoch=2000 // BATCH_SIZE,
                             validation_data=valid_batches,
                             validation_steps=800 // BATCH_SIZE)
    return vgg16_fine    

I guess there is a way to achieve much better results with fine-tuning (up to 98%), but I wasn't able to achieve it with provided code.