wassname/keras-language-modeling
1
0
Fork 0
mirror of https://github.com/wassname/keras-language-modeling.git synced 2026-09-10 12:15:18 +08:00
Code Issues Packages Projects Releases Wiki Activity

added noise

Browse Source
This commit is contained in:
codekansas
2016-05-06 03:36:00 -04:00
parent a853407a3a
commit 547e32acee
1 changed files with 10 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
answer_to_question.py
+10 -8
View File
@@ -45,8 +45,9 @@ class InsuranceQA:
indices[i, self.words_indices[w]] = 1
return indices
def decode(self, indices, calc_argmax=True):
def decode(self, indices, calc_argmax=True, noise=0):
if calc_argmax:
indices = indices + np.random.rand(*indices.shape) * noise
indices = indices.argmax(axis=-1)
return ' '.join(self.indices_words[x] for x in indices)
@@ -54,13 +55,12 @@ def get_model(question_maxlen, answer_maxlen, vocab_len, n_hidden):
answer = Input(shape=(answer_maxlen, vocab_len))
# answer = Masking(mask_value=0.)(answer)
# for i in range(2):
# answer = LSTM(n_hidden, return_sequences=True)(answer)
# encoder rnn
encode_rnn = LSTM(n_hidden, return_sequences=False)(answer)
# can add more layers
for i in range(2):
encode_rnn = LSTM(n_hidden, return_sequences=True)(encode_rnn)
# repeat it maxlen times
repeat_encoding = RepeatVector(question_maxlen)(encode_rnn)
@@ -125,6 +125,8 @@ if __name__ == '__main__':
x, y = next(test_gen)
y = y[0]
pred = model.predict(x, verbose=0)
for i in range(n_test):
print('Expected: {}'.format(qa.table.decode(y[i])))
print('Predicted: {}'.format(qa.table.decode(pred[i])))
for noise in [0, 0.1, 0.2]: # not sure what noise values would be good
print(' Noise: {}'.format(noise))
for i in range(n_test):
print(' Expected: {}'.format(qa.table.decode(y[i])))
print(' Predicted: {}'.format(qa.table.decode(pred[i], noise=noise)))