keras自定义生成器categorical_crossentropy修复输出形状问题

Question

I have an issue with batch labels. How I can pass labels correctly to fit_generator?

I assume that it should be something like this: If #0 = jazz, #1 = rock, #2 = electronic.

So if batch is 4 for example...

batch_features batch_labels

array_with_rock 1

array_with_rock 1

array_with_electro 2

array_with_jazz 0

where array_with_tag is mel-spectrogram in memory.

My task is: train NN_Music tagger https://github.com/Vital95/NN_Music/blob/master/music_tagger_crnn.py

On this dataset: https://github.com/mdeff/fma

progress for now: I was able to train binary_crossentropy model for rock only.

Issue: Error when checking target: expected output to have shape (None, 3) but got array with shape (12, 1)

And some main code:

Main

if __name__ == '__main__':
    batch_features, batch_labels = getFeturesAndLabelsFromTarget(target)
    model = getModel()
    my_generator = MelGenerator(features = batch_features, labels = batch_labels, batch_size =  12)
    model.compile(loss='categorical_crossentropy',
                optimizer='rmsprop',
                metrics=['accuracy'])
    model.fit_generator(my_generator, samples_per_epoch = 10, nb_epoch = 10, verbose=2,  callbacks=None, validation_data=None, class_weight=None, nb_worker=1)
    model.save_weights('test012.h5')
    print("done")

Generator

def MelGenerator(features, labels, batch_size):
 # Create empty arrays to contain batch of features and labels#
    batch_features = np.zeros((batch_size, 96, 1366, 1))
    batch_labels = np.zeros((batch_size,1))
    while True:
        for i in range(batch_size):
        # choose random index in features
            #to do review this line
            #index = random.choice(np.size(batch_features,1))
            index = randint(0, np.size(batch_features,0))
            batch_features[i] = features[index]
            batch_labels[i] = labels[index]
        yield batch_features, batch_labels

Load Mel Spectro from .wav files to memory (this is else case)

def getFeturesAndLabelsFromTarget(targetFolder):
    wavFiles = WTI.GetListOfFilesByExt(targetFolder, extention = '.wav')
    upperFolder = WTI.getUpperFolders(wavFiles)
    n = len(wavFiles)
    if len(set(upperFolder)) == 1:
        batch_features = np.zeros((n, 96, 1366, 1))
        batch_labels = np.zeros((n,1),dtype=np.int8)
        for i in range(0,n):
            batch_features[i] = preprocess_input(wavFiles[i])
            #correct
            # print(str(np.size(batch_features,0)))
            batch_labels[i] = 0
        return batch_features, batch_labels
    else:
        s = set(upperFolder)
        theList = list(s)
        batch_features = np.zeros((n, 96, 1366, 1))
        batch_labels = np.zeros((n,1),dtype=np.int8)    
        for i in range(0,n):
            batch_features[i] = preprocess_input(wavFiles[i])
            if(theList[0] in wavFiles[i]):
                #0 = jazz
                batch_labels[i] = 0
            if(theList[1] in wavFiles[i]):
                #1 = rock
                batch_labels[i] = 1
            if(theList[2] in wavFiles[i]):
                #2 = electronic
                batch_labels[i] = 2
        return batch_features, batch_labels

model

def getModel(input_tensor=None):
    if K.image_dim_ordering() == 'th':
        input_shape = (1, 96, 1366)
    else:
        input_shape = (96, 1366, 1)

    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 = ZeroPadding2D(padding=(0, 37))(melgram_input)
    x = BatchNormalization(axis=time_axis, name='bn_0_freq')(x)

    # 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, 2), strides=(2, 2), 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=(3, 3), strides=(3, 3), 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=(4, 4), strides=(4, 4), name='pool3')(x)

    # Conv block 4
    x = Convolution2D(128, 3, 3, border_mode='same', name='conv4')(x)
    x = BatchNormalization(axis=channel_axis, mode=0, name='bn4')(x)
    x = ELU()(x)
    x = MaxPooling2D(pool_size=(4, 4), strides=(4, 4), name='pool4')(x)

    # reshaping
    if K.image_dim_ordering() == 'th':
        x = Permute((3, 1, 2))(x)
    x = Reshape((15, 128))(x)

    # GRU block 1, 2, output
    x = GRU(32, return_sequences=True, name='gru1')(x)
    x = GRU(32, return_sequences=False, name='gru2')(x)

    x = Dense(3, activation='sigmoid', name='output')(x)

    # Create model
    model = Model(melgram_input, x)
    return model

Answer 1

This happens because you are not using one-hot encoded labels, which are required by the categorical_crossentropy loss.

You can transform your labels using keras.utils.np_utils.to_categorical and it should work.