Keras beginner: CNN applied to spectrograms. Help building the network.

[Fran Carranza]

Hi, I'm very new to this topics. I'm trying to build a CNN in order to feed with spectrograms. A spectrogram is a matrix where the x axis is time and y axis, frequency.

For example my data is 128 mel components (frequency) and 130 timesteps.

To summarize the instructions for the network is:

The network consisted of two convolutional layers with 128 filters span-
ning 6 timesteps and the entire frequency range, yielding one-dimensional
convolutions. A max-pooling layer with a pooling window of 4 timesteps
followed the first convolutional layer, and one with a pooling window of 5
timesteps followed the second convolutional layer. A fully connected hidden
layer with 400 units was stacked on top of this, and finally the output layer
had a number of units equaling the number of latent factors to predict.

That is expanded in this reddit's thread.

I'm currently testing with a tiny portion of the dataset just to test if the network is properly settled.

My data:

• 80 Spectrograms with 128 mel frequencies and 130 time frames
  

I reshaped the data to match de 2d Convolution requirements. So my input_shape is (128, 130, 1)

K.set_image_data_format('channels_first')
input_shape = (128, 130, 1)
n_filters = 128
num_classes = 50

And the network is built like this:

model = Sequential()
model.add(Conv2D(n_filters, 4, padding='valid', strides=1,
                 input_shape=input_shape))
model.add(Activation('relu'))
model.add(Conv2D(n_filters, 5))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(4,1)))
model.add(Dropout(0.25))


model.add(Flatten())
model.add(Dense(400))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes))
model.add(Activation('softmax'))

Please tell me if the network parameters I've filled match with the above description.

Also, when I excecute the last piece of code I get this error:

ValueError: ('The specified size contains a dimension with value <= 0', (-23400, 400))

I also read this answer https://groups.google.com/forum/#!topic/keras-users/vWG8cjCip8Q, but I can't get it to work.

Thanks for the support!

[Fran Carranza]

model.add(Conv2D(n_filters, 4, padding='valid', strides=1,
                 input_shape=input_shape))

I've found that the second number (kernel_size) must be 2 dimensional. If I put (4,1), now it compiles!

But I think that something is still wrong. The model scores almost 0 of accuracy.

[baptist...@gmail.com]

Keras has a Conv1D layer that is more appropriate to your particular task; you're doing 2D convolutions on a (130,1) vector. Also, make sure that your backend is using the "channel_first" data format. Theano does, Tensorflow, as far as I know, does not. Also, your strides for a 2D layer should also be a tuple. Finally, the number of filters in each conv layer does not need to be 128. It can be anything you want, it just changes the complexity of the network. Good luck!