padded_markers = numpy.array([mark_word_boundaries([map_ind_2_chr[ind] for ind in seq]) for seq in seq_batch])
padded_markers = padded_markers.flatten(order="C")
return padded_markers[mask_batch.flatten(order="C") == 1]
def mark_letter(seq_batch, mask_batch, letter):
padded_markers = 1*numpy.array([[map_ind_2_chr[char] == letter for char in seq] for seq in seq_batch])
padded_markers = padded_markers.flatten(order="C")
return padded_markers[mask_batch.flatten(order="C") == 1]
map_chr_2_ind = cPickle.load(open("char_to_ind.pkl"))
map_ind_2_chr = cPickle.load(open("ind_to_char.pkl"))
lstm_net = Network(NetworkType.LSTM, input_dim=len(map_ind_2_chr), hidden_dims=[512, 512, 512])
lstm_net.set_parameters('seqgen_lstm_512_512_512.pkl')
# having a look at connectioneros from the cellsinas to the outputsos
params = lstm_net.cost_model.get_parameter_values()
for param in params:
print param
# this section deals with prediction probabilities
"""
readouts = VariableFilter(theano_name="readout_readout_output_0")(lstm_net.cost_model.variables)[0]
char_probs = lstm_net.generator.readout.emitter.probs(readouts)
prob_function = function([lstm_net.x, lstm_net.mask], char_probs)
padded_markers = padded_markers.flatten(order="C")
return padded_markers[mask_batch.flatten(order="C") == 1]
if args.function == "len":
corr_function = mark_seq_len_batch
elif args.function == "bound":
corr_function = mark_word_boundaries_batch
else:
raise ValueError("Invalid correlation function specified!")
map_chr_2_ind = cPickle.load(open("char_to_ind.pkl"))
map_ind_2_chr = cPickle.load(open("ind_to_char.pkl"))
lstm_net = Network(NetworkType.SIMPLE_RNN, len(map_ind_2_chr), hidden_dims=[1024])
lstm_net.set_parameters('seqgen_simple_1024.pkl')
# having a look at connectioneros from the cellsinas to the outputsos
params = lstm_net.cost_model.get_parameter_values()
for param in params:
print param
# define a function that gets the overall "sum of scores" at a given time step
readouts = VariableFilter(theano_name="readout_readout_output_0")(lstm_net.cost_model.variables)[0]
score_function = function([lstm_net.x, lstm_net.mask], readouts.sum(axis=2))
# this section of the playground has some fun rides that revolve around various correlation stuff. uncomment to access
# =)
sc = StateComputer(lstm_net.cost_model, map_chr_2_ind)
