from userconfig import *
import util
from layers_custom import *
###################################################################################################################
# create Theano variables for input minibatch
input_var = T.tensor4('X')
# note that in general, the main data tensors will have these axes:
# - minibatchsize
# - numchannels (always 1 for us, since spectrograms)
# - numfilts (or specbinnum for input)
# - numtimebins
if example_is_audio:
# load our example audio file as a specgram
examplegram = util.standard_specgram((util.load_soundfile(examplewavpath, 0)))
print("examplegram is of shape %s" % str(np.shape(examplegram)))
###################################################################################################################
# here we define our "semi-convolutional" autoencoder
# NOTE: lasagne assumes pooling is on the TRAILING axis of the tensor, so we always use time as the trailing axis
def make_custom_convlayer(network, in_num_chans, out_num_chans):
"Applies our special padding and reshaping to do 1D convolution on 2D data"
network = lasagne.layers.PadLayer(network, width=(featframe_len-1)/2, batch_ndim=3) # NOTE: the "batch_ndim" is used to stop batch dims being padded, but here ALSO to skip first data dim
print("shape after pad layer: %s" % str(network.output_shape))
network = lasagne.layers.Conv2DLayer(network, out_num_chans, (in_num_chans, featframe_len), stride=(1,1), pad=0, nonlinearity=very_leaky_rectify, W=lasagne.init.Orthogonal()) # we pad "manually" in order to do it in one dimension only
filters = network.W
network = lasagne.layers.ReshapeLayer(network, ([0], [2], [1], [3])) # reinterpret channels as rows
print("shape after conv layer: %s" % str(network.output_shape))
return network, filters
import util
from layers_custom import *
###################################################################################################################
# create Theano variables for input minibatch
input_var = T.tensor4('X')
# note that in general, the main data tensors will have these axes:
# - minibatchsize
# - numchannels (always 1 for us, since spectrograms)
# - numfilts (or specbinnum for input)
# - numtimebins
if example_is_audio:
# load our example audio file as a specgram
foregroundgram = util.standard_specgram(
(util.load_soundfile(foregroundwavpath, 0)))
backgroundgram = util.standard_specgram(
(util.load_soundfile(backgroundwavpath, 0)))
print("foregroundgram is of shape %s" % str(np.shape(foregroundgram)))
print("backgroundgram is of shape %s" % str(np.shape(backgroundgram)))
###################################################################################################################
# here we define our "semi-convolutional" autoencoder
# NOTE: lasagne assumes pooling is on the TRAILING axis of the tensor, so we always use time as the trailing axis
def make_custom_convlayer(network, in_num_chans, out_num_chans):
"Applies our special padding and reshaping to do 1D convolution on 2D data"
network = lasagne.layers.PadLayer(
network, width=(featframe_len - 1) / 2, batch_ndim=3
) # NOTE: the "batch_ndim" is used to stop batch dims being padded, but here ALSO to skip first data dim
from userconfig import *
import util
from layers_custom import *
###################################################################################################################
# create Theano variables for input minibatch
input_var = T.tensor4('X')
# note that in general, the main data tensors will have these axes:
# - minibatchsize
# - numchannels (always 1 for us, since spectrograms)
# - numfilts (or specbinnum for input)
# - numtimebins
if example_is_audio:
# load our example audio file as a specgram
foregroundgram = util.standard_specgram((util.load_soundfile(foregroundwavpath, 0)))
backgroundgram = util.standard_specgram((util.load_soundfile(backgroundwavpath, 0)))
print("foregroundgram is of shape %s" % str(np.shape(foregroundgram)))
print("backgroundgram is of shape %s" % str(np.shape(backgroundgram)))
###################################################################################################################
# here we define our "semi-convolutional" autoencoder
# NOTE: lasagne assumes pooling is on the TRAILING axis of the tensor, so we always use time as the trailing axis
def make_custom_convlayer(network, in_num_chans, out_num_chans):
"Applies our special padding and reshaping to do 1D convolution on 2D data"
network = lasagne.layers.PadLayer(network, width=(featframe_len-1)/2, batch_ndim=3) # NOTE: the "batch_ndim" is used to stop batch dims being padded, but here ALSO to skip first data dim
print("shape after pad layer: %s" % str(network.output_shape))
network = lasagne.layers.Conv2DLayer(network, out_num_chans, (in_num_chans, featframe_len), stride=(1,1), pad=0, nonlinearity=very_leaky_rectify, W=lasagne.init.Orthogonal()) # we pad "manually" in order to do it in one dimension only
filters = network.W
