Incorrect number of feature dimensions with VGG + RNN

Hi,

I just ran into the following problem when trying to train with the VGG convnet (encoder type ‘vggblstmp’).

For an input of size B x T x F, VGG always returns a tensor of size Batch x T/4 x 128. One would assume that the number of feature dimensions for the subsequent RNN should also be 128.

But to determine the number of feature dimensions for the RNN, we call

def get_vgg2l_odim(idim, in_channel=3, out_channel=128):
    idim = idim / in_channel
    idim = np.ceil(np.array(idim, dtype=np.float32) / 2)  # 1st max pooling
    idim = np.ceil(np.array(idim, dtype=np.float32) / 2)  # 2nd max pooling
    return int(idim) * out_channel  # numer of channels

which yields F / 4 / in_channel * 128, causing a dimension mismatch. Shouldn’t that method just return the number of filters of the final convolutional layer (128)?

If this is the case, I can create a pull request.

[sw005320]

Sure, I’ll fix it.

[sw005320]

Done. Thanks!