83 lines
3 KiB
Python
83 lines
3 KiB
Python
|
from nolearn.lasagne import NeuralNet, BatchIterator
|
||
|
|
from lasagne.layers import InputLayer as Input
|
||
|
|
from lasagne.layers import DropoutLayer as Dropout
|
||
|
|
from lasagne.layers import DenseLayer as Dense
|
||
|
|
from lasagne.layers import RecurrentLayer as RNN
|
||
|
|
from lasagne.layers import LSTMLayer as LSTM
|
||
|
|
from lasagne.layers import GRULayer as GRU
|
||
|
|
from lasagne.layers import ReshapeLayer as Reshape
|
||
|
|
from lasagne.layers import EmbeddingLayer as Embedding
|
||
|
|
from lasagne.layers import NonlinearityLayer as Nonlinearity
|
||
|
|
from lasagne.nonlinearities import softmax
|
||
|
|
from lasagne.objectives import categorical_crossentropy
|
||
|
|
from lasagne.updates import adam
|
||
|
|
|
||
|
|
|
||
|
|
elif 0:
|
||
|
|
# TODO: consider using to_categorical instead
|
||
|
|
#from keras.utils.np_utils import to_categorical
|
||
|
|
x = np.array(seq_text[:newlen], dtype='int32').reshape((-1, seq_length))
|
||
|
|
# Generate input and output per substring, as an indicator matrix.
|
||
|
|
y = np.zeros((x.shape[0], x.shape[1], self.vocab_size), dtype='bool')
|
||
|
|
for i in np.arange(x.shape[0]):
|
||
|
|
for j in np.arange(x.shape[1]):
|
||
|
|
y[i, j, x[i, j]] = 1
|
||
|
|
|
||
|
|
# push a padding character to the front of the inputs.
|
||
|
|
# this effectively maps x[a,b,c]'s next character to y[a,b,c]
|
||
|
|
x = np.roll(y, 1, axis=1)
|
||
|
|
x[:, 0, :] = 0
|
||
|
|
x[:, 0, 0] = 1
|
||
|
|
|
||
|
|
vx = x
|
||
|
|
vy = y
|
||
|
|
else:
|
||
|
|
# TODO: handle dtype more elegantly
|
||
|
|
#vx = seq_text[0:newlen + 0]
|
||
|
|
#vy = seq_text[1:newlen + 1]
|
||
|
|
vy = seq_text[:newlen]
|
||
|
|
vx = np.roll(vy, 1)
|
||
|
|
vx[0] = 0 # remember we assert that 0 corresponds to PADDING in Translator
|
||
|
|
|
||
|
|
# stuff in the original LanguageModel.lua that we don't need
|
||
|
|
#vx = vx.reshape(batch_size, -1, self.seq_length)
|
||
|
|
#vy = vy.reshape(batch_size, -1, self.seq_length)
|
||
|
|
#vx = vx.transpose([1, 0, 2])
|
||
|
|
#vy = vy.transpose([1, 0, 2])
|
||
|
|
|
||
|
|
# this is good enough
|
||
|
|
vx = vx.reshape(-1, self.seq_length)
|
||
|
|
vy = vy.reshape(-1, self.seq_length)
|
||
|
|
vx = np.array(vx, dtype=np.uint8)
|
||
|
|
vy = np.array(vy, dtype=np.uint8)
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
sentence = text[start_index:start_index + seq_length]
|
||
|
|
generated = sentence
|
||
|
|
lament('## Generating')
|
||
|
|
sys.stdout.write(generated)
|
||
|
|
sys.stdout.write('~') # PADDING
|
||
|
|
|
||
|
|
was_pad = True
|
||
|
|
for i in range(sample_chars):
|
||
|
|
x = np.zeros((1, seq_length, textmap.vocab_size), dtype=np.bool)
|
||
|
|
for t, char in enumerate(sentence):
|
||
|
|
x[0, t, textmap.map(char)] = 1
|
||
|
|
|
||
|
|
preds = model.predict(x, batch_size=1, verbose=0)[0]
|
||
|
|
next_index = asample(preds, temperature)
|
||
|
|
next_char = textmap.unmap(next_index)[0]
|
||
|
|
|
||
|
|
generated += next_char
|
||
|
|
sentence = sentence[1:] + next_char
|
||
|
|
|
||
|
|
is_pad = next_char == PADDING
|
||
|
|
sys.stdout.write(next_char)
|
||
|
|
if is_pad and not was_pad:
|
||
|
|
sys.stdout.write('\n')
|
||
|
|
sys.stdout.flush()
|
||
|
|
|
||
|
|
was_pad = is_pad
|
||
|
|
lament()
|