name='GRUEncoder',
mask_input=l_mask_enc)
print lasagne.layers.get_output(l_enc,
inputs={
l_in: x_sym,
l_mask_enc: xmask_sym
}).eval({
x_sym: inputs,
xmask_sym: input_masks
}).shape
##### START OF DECODER######
l_dec = LSTMAttentionDecodeFeedbackLayer(l_enc,
num_units=NUM_UNITS_DEC,
aln_num_units=20,
n_decodesteps=MAX_DIGITS + 1,
name='LSTMDecoder')
# We need to do some reshape voodo to connect a softmax layer to the decoder.
# See http://lasagne.readthedocs.org/en/latest/modules/layers/recurrent.html#examples
# In short this line changes the shape from
# (batch_size, decode_len, num_dec_units) -> (batch_size*decodelen,num_dec_units).
# We need to do this since the softmax is applied to the last dimension and we want to
# softmax the output at each position individually
l_reshape = lasagne.layers.ReshapeLayer(l_dec, (-1, [2]))
l_softmax = lasagne.layers.DenseLayer(
l_reshape,
num_units=NUM_OUTPUTS,
nonlinearity=lasagne.nonlinearities.softmax,
l_enc = lasagne.layers.GRULayer(l_in,
num_units=NUM_UNITS_ENC,
mask_input=l_in_mask)
print "GRU Encoder: {}".format(
lasagne.layers.get_output(l_enc,
inputs={
l_in: x_sym,
l_in_mask: x_mask_sym
}).eval({
x_sym: v_train_x,
x_mask_sym: v_train_x_mask
}).shape)
l_dec = LSTMAttentionDecodeFeedbackLayer(l_enc,
num_units=NUM_UNITS_DEC,
aln_num_units=52,
n_decodesteps=MAX_LENGHT_OUPUT)
print "LSTM Decoder Layer: {}".format(
lasagne.layers.get_output(l_dec,
inputs={
l_in: x_sym,
l_in_mask: x_mask_sym
}).eval({
x_sym: v_train_x,
x_mask_sym: v_train_x_mask
}).shape)
# classic reshape voodoo
# we have an input that is batch_size, time_steps, num_units
# we want batch_size * time_steps, num_units