shallow complex model on Musicnet Dataset

[linglingxu]

Hi all,
I got this error when trying to run the command: python train.py shallow_complex_model --in-memory --model=complex_shallow_convnet --complex --local-data data/musicnet_11khz.npz

then i change

then i make a change that let "channels= 2", which is " channels=feature_dim[1]=1" before. Then i pass the previous question and get another error.

[OisinMoran]

I have the same issue. Did you ever solve this @linglingxu?

[austinmw]

I'm getting the same thing and found that fan_in is equaling zero due to an input of (512,0,32) on line 236 of complexnn/init.py : fan_in, fan_out = initializers._compute_fans(tuple(self.kernel_size) + (self.input_dim, self.nb_filters)), where self.input_dim equals 0. This causes division by zero on line 71 of complexnn/init.

I think it stems from musicnet/musicnet/dataset.py lines 167 to 190:

        channels = 2 if self.complex_ else 1
        features_out = numpy.zeros(
            [features.shape[0], self.window_size, channels])
        if self.fourier:
            if self.complex_:
                data = fft(features, axis=1)
                features_out[:, :, 0] = numpy.real(data[:, :, 0])
                features_out[:, :, 1] = numpy.imag(data[:, :, 0])
            else:
                data = numpy.abs(fft(features, axis=1))
                features_out = data
        elif self.stft:
            _, _, data = stft(features, nperseg=120, noverlap=60, axis=1)
            length = data.shape[1]
            n_feats = data.shape[3]
            if self.complex_:
                features_out = numpy.zeros(
                    [len(self.train_data), length, n_feats * 2])
                features_out[:, :, :n_feats] = numpy.real(data)
                features_out[:, :, n_feats:] = numpy.imag(data)
            else:
                features_out = numpy.abs(data[:, :, 0, :])
        else:
            features_out = features
        return features_out, output

When I try to run shallow complex on the resampled musicnet dataset, first it sets channels=2 and features_out.shape=(22500,4096,2), but then it overwrites this to set features_out to equal features which is of shape (22500,4096,1). Channels is 1, but then it is later divided and floored by 2 to get input_dim, which then equals 0.