Optimizing the Architecture of a CNN Using Keras in Python3

Question

I am trying to increase my validation accuracy of my CNN from 76% (currently) to over 90%. I am going to show all of the information about my CNN's performance and configuration below.

In essence, I want my CNN to distinguish between two classes of mel-spectrograms:

Class # 1 Class # 2 Here is the graph of accuracy vs epoch:

Here is the graph of loss vs. epoch

Finally, here is the model architecture configuration

model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),activation='relu',input_shape=(3, 640, 480)))
model.add(Conv2D(64, (3, 3), activation='relu', dim_ordering="th"))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(2, activation='softmax'))

Here are my calls to model.compile() and model.fit()

model.compile(loss=keras.losses.categorical_crossentropy,
          optimizer=keras.optimizers.SGD(lr=0.001),
          metrics=['accuracy'])
print("Compiled model")

history = model.fit(X_train, Y_train,
      batch_size=8,
      epochs=50,
      verbose=1,
      validation_data=(X_test, Y_test))

How can I change my CNN configuration to increase the validation accuracy score?

Things I have tried:

1. Decrease the learning rate to prevent sporadic fluctuations in the accuracy.
2. Decrease the batch_size from 64 down to 8.
3. Increase the number of epochs to 50(However not sure if this is enough).

Any help would be greatly appreciated!

UPDATE #1 I increase the number of epochs to 200, and after letting the program run overnight I got a validated accuracy score of around 76.31%

I am posting a picture of accuracy vs. epoch and loss vs. epoch below

What else specifically about my model architecture can I change to get better accuracy?

Answer 1

First of all you have to get music_tagger_cnn.py and put it in the project path. After that you can build your model:

from music_tagger_cnn import *
input_tensor = Input(shape=(1, 18, 119))
model =MusicTaggerCNN(input_tensor=input_tensor, include_top=False, weights='msd')

You can change the input tensor by the dimension you want... I usually use Theano dim ordering but Tensorflow as backend, so that's why:

from keras import backend as K
K.set_image_dim_ordering('th')

Using Theano dim ordering you hav to take into account that the order of the sample's dimensions have to be changed

X_train = X_train.transpose(0, 3, 2, 1)
X_val = X_val.transpose(0, 3, 2, 1)

After that you have to freeze these layers that you don't want to be updated

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

Now you can set your own output, for example:

last_layer = model.get_layer('pool3').output
out = Flatten()(last_layer)
out = Dense(128, activation='relu', name='fc2')(out)
out = Dropout(0.5)(out)
out = Dense(n_classes, activation='softmax', name='fc3')(out)
model = Model(input=model.input, output=out)

After that you have to be able to train it just doing:

sgd = SGD(lr=0.01, momentum=0, decay=0.002, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
history = model.fit(X_train, labels_train,
                          validation_data=(X_val, labels_val), nb_epoch=100, batch_size=5)

Note that labels should be in one-hot encoding

I hope it will help!!

Update: Posting code so I can get help debugging these lines and prevent a crash.

input_tensor = Input(shape=(3, 640, 480))
model = MusicTaggerCNN(input_tensor=input_tensor, include_top=False, weights='msd')

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


last_layer = model.get_layer('pool3').output
out = Flatten()(last_layer)
out = Dense(128, activation='relu', name='fc2')(out)
out = Dropout(0.5)(out)
out = Dense(n_classes, activation='softmax', name='fc3')(out)
model = Model(input=model.input, output=out)

sgd = SGD(lr=0.01, momentum=0, decay=0.002, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
history = model.fit(X_train, labels_train,
                          validation_data=(X_test, Y_test), nb_epoch=100, batch_size=5)

EDIT # 2

    # -*- coding: utf-8 -*-
'''MusicTaggerCNN model for Keras.

# Reference:

- [Automatic tagging using deep convolutional neural networks](https://arxiv.org/abs/1606.00298)
- [Music-auto_tagging-keras](https://github.com/keunwoochoi/music-auto_tagging-keras)

'''
from __future__ import print_function
from __future__ import absolute_import

from keras import backend as K
from keras.layers import Input, Dense
from keras.models import Model
from keras.layers import Dense, Dropout, Flatten
from keras.layers.convolutional import Convolution2D
from keras.layers.convolutional import MaxPooling2D, ZeroPadding2D
from keras.layers.normalization import BatchNormalization
from keras.layers.advanced_activations import ELU
from keras.utils.data_utils import get_file
from keras.layers import Input, Dense

TH_WEIGHTS_PATH = 'https://github.com/keunwoochoi/music-auto_tagging-keras/blob/master/data/music_tagger_cnn_weights_theano.h5'
TF_WEIGHTS_PATH = 'https://github.com/keunwoochoi/music-auto_tagging-keras/blob/master/data/music_tagger_cnn_weights_tensorflow.h5'


def MusicTaggerCNN(weights='msd', input_tensor=None,
                   include_top=True):
    '''Instantiate the MusicTaggerCNN architecture,
    optionally loading weights pre-trained
    on Million Song Dataset. Note that when using TensorFlow,
    for best performance you should set
    `image_dim_ordering="tf"` in your Keras config
    at ~/.keras/keras.json.

    The model and the weights are compatible with both
    TensorFlow and Theano. The dimension ordering
    convention used by the model is the one
    specified in your Keras config file.

    For preparing mel-spectrogram input, see
    `audio_conv_utils.py` in [applications](https://github.com/fchollet/keras/tree/master/keras/applications).
    You will need to install [Librosa](http://librosa.github.io/librosa/)
    to use it.

    # Arguments
        weights: one of `None` (random initialization)
            or "msd" (pre-training on ImageNet).
        input_tensor: optional Keras tensor (i.e. output of `layers.Input()`)
            to use as image input for the model.
        include_top: whether to include the 1 fully-connected
            layer (output layer) at the top of the network.
            If False, the network outputs 256-dim features.


    # Returns
        A Keras model instance.
    '''
    if weights not in {'msd', None}:
        raise ValueError('The `weights` argument should be either '
                         '`None` (random initialization) or `msd` '
                         '(pre-training on Million Song Dataset).')

    # Determine proper input shape
    if K.image_dim_ordering() == 'th':
        input_shape = (3, 640, 480)
    else:
        input_shape = (3, 640, 480)

    if input_tensor is None:
        melgram_input = Input(shape=input_shape)
    else:
        if not K.is_keras_tensor(input_tensor):
            melgram_input = Input(tensor=input_tensor, shape=input_shape)
        else:
            melgram_input = input_tensor

    # Determine input axis
    if K.image_dim_ordering() == 'th':
        channel_axis = 1
        freq_axis = 2
        time_axis = 3
    else:
        channel_axis = 3
        freq_axis = 1
        time_axis = 2

    # Input block
    x = BatchNormalization(axis=freq_axis, name='bn_0_freq')(melgram_input)

    # Conv block 1
    x = Convolution2D(64, 3, 3, border_mode='same', name='conv1')(x)
    x = BatchNormalization(axis=channel_axis, mode=0, name='bn1')(x)
    x = ELU()(x)
    x = MaxPooling2D(pool_size=(2, 4), name='pool1')(x)

    # Conv block 2
    x = Convolution2D(128, 3, 3, border_mode='same', name='conv2')(x)
    x = BatchNormalization(axis=channel_axis, mode=0, name='bn2')(x)
    x = ELU()(x)
    x = MaxPooling2D(pool_size=(2, 4), name='pool2')(x)

    # Conv block 3
    x = Convolution2D(128, 3, 3, border_mode='same', name='conv3')(x)
    x = BatchNormalization(axis=channel_axis, mode=0, name='bn3')(x)
    x = ELU()(x)
    x = MaxPooling2D(pool_size=(2, 4), name='pool3')(x)



    # Output
    x = Flatten()(x)
    if include_top:
        x = Dense(50, activation='sigmoid', name='output')(x)

    # Create model
    model = Model(melgram_input, x)
    if weights is None:
        return model
    else:
        # Load input
        if K.image_dim_ordering() == 'tf':
            raise RuntimeError("Please set image_dim_ordering == 'th'."
                               "You can set it at ~/.keras/keras.json")
        model.load_weights('data/music_tagger_cnn_weights_%s.h5' % K._BACKEND,
                           by_name=True)
        return model

EDIT #3

I tried the keras example for using the MusicTaggerCRNN as a feature extractor of the melgrams. Then i trained a simple NN with 2 Dense layers and a binary output. The samples taken in my example don't apply in your case but it's also a binary classifier I used keras==1.2.2 and tensorflow-gpu==1.0.0 and works for me.

Here's the code:

from keras.applications.music_tagger_crnn import MusicTaggerCRNN
from keras.applications.music_tagger_crnn import preprocess_input, decode_predictions
import numpy as np
from keras.layers import Input, Dense
from keras.models import Model
from keras.layers import Dense, Dropout, Flatten
from keras.optimizers import SGD


model = MusicTaggerCRNN(weights='msd', include_top=False)
#Samples simulation
audio_paths_train = ['data/genres/blues/blues.00000.au','data/genres/classical/classical.00000.au','data/genres/classical/classical.00002.au', 'data/genres/blues/blues.00003.au']
audio_paths_test = ['data/genres/blues/blues.00001.au', 'data/genres/classical/classical.00001.au', 'data/genres/blues/blues.00002.au', 'data/genres/classical/classical.00003.au']
labels_train = [0,1,1,0]
labels_test = [0, 1, 0, 1]
melgrams_train = [preprocess_input(audio_path) for audio_path in audio_paths_train]
melgrams_test = [preprocess_input(audio_path) for audio_path in audio_paths_test]
feats_train = [model.predict(np.expand_dims(melgram, axis=0)) for melgram in melgrams_train]
feats_test = [model.predict(np.expand_dims(melgram, axis=0)) for melgram in melgrams_test]
feats_train = np.array(feats_train)
feats_test = np.array(feats_test)

_input = Input(shape=(1,32))
x = Flatten(name='flatten')(_input)
x = Dense(128, activation='relu', name='fc6')(x)
x = Dense(64, activation='relu', name='fc7')(x)
x = Dense(1, activation='softmax', name='fc8')(x)
class_model = Model(_input, x)

sgd = SGD(lr=0.01, momentum=0, decay=0.02, nesterov=True)
class_model.compile(loss='binary_crossentropy', optimizer=sgd, metrics=['accuracy'])
history = class_model.fit(feats_train, labels_train, validation_data=(feats_test, labels_test), nb_epoch=100, batch_size=5, class_weight='auto')
print(history.history['acc'])

# Final evaluation of the model
scores = class_model.evaluate(feats_test, labels_test, verbose=0)
print("Accuracy: %.2f%%" % (scores[1] * 100))

Answer 2

1. Dropout: It turns out that simple Dropout is not effective with CNNs. Try changing the first Dropout layer to SpatialDropout2D. The second Dropout is just for a standard Dense layer, so that is the right kind of Dropout. CNN networks create a series of overlapping pyramids. Standard Dropout just turns some of those pyramids into stumps. SpatialDropout complete zeroes out some of the pyramids, which is a better mode of "forgetting".

2. Repeated structure: Use a few different copies of your Conv2D->Conv2D->SpatialDropout chain, making downsized images with more filters. All of the successful image-processing designs use multiple repeated blocks.

3. Your data: Spectrograms have two different measurements as the different dimensions of the "image". The normal design for image processing is to consider all of the neighbor pixels as equally contributing to one output feature. It may be that you want many pixels in one row to contribute to a feature map, but only one or two rows. So, maybe each pyramid needs a long&narrow base.

Answer 3

you can do several preliminary modifications to your code and check result:

1. increase layers count
2. increase layers size
3. change learning_rate
4. change optimizer (usually Adam works better than SGD ,...)
5. shuffle your data(maybe your test data include some samples which are far from training samples
6. try adding batch normalization layers
7. also sometimes increasing MFCC features number helps(instead of 640, 480 try extracting more features