def apply_sampling(embeddings, hid, start='S', end='E', t=1.0):
x = embeddings[start]
e = embeddings[end]
S_x = helper.index_to_vec(start, embeddings)
S_h = np.zeros(hid, dtype=theano.config.floatX)
S_c = np.zeros(hid, dtype=theano.config.floatX)
sampled = [x]
counter = 0
while x != e:
S_x = np.reshape(S_x, (1, -1))
S_h, S_c, S_y = sampling(S_x, S_h.flatten(), S_c.flatten())
S_y = helper.t_softmax(S_y.flatten(), t=t)
S_x = np.random.multinomial(n=1, pvals=S_y)
x = helper.vec_to_index(S_x)
sampled.append(x)
if counter == 1000:
break
counter += 1
return sampled[1:-1]
def apply_sampling(embeddings, hid, start='S', end='E', t=1.0):
x = embeddings[start]
e = embeddings[end]
S_x = helper.index_to_vec(start, embeddings)
S_h = np.zeros(hid, dtype=theano.config.floatX)
S_c = np.zeros(hid, dtype=theano.config.floatX)
sampled = [x]
counter = 0
while x != e:
S_x = np.reshape(S_x, (1, -1))
S_h, S_c, S_y = sampling(S_x, S_h.flatten(), S_c.flatten())
S_y = helper.t_softmax(S_y.flatten(), t=t)
S_x = np.random.multinomial(n=1, pvals=S_y)
x = helper.vec_to_index(S_x)
sampled.append(x)
if counter == 1000:
break
counter += 1
return sampled[1:-1]
def apply_sampling(token_embeddings, hid, start='S', end='E', t=1.0, max_tokens=50):
token_list = list(token_embeddings.token_to_index.keys())
S_x = token_embeddings.token_to_vec(start)
S_h = np.zeros(hid, dtype=theano.config.floatX)
S_c = np.zeros(hid, dtype=theano.config.floatX)
sampled_tokens = [start]
counter = 0
while sampled_tokens[-1] != end:
if counter == max_tokens:
sampled_tokens.append(end)
break
S_x = np.reshape(S_x, (1, -1))
S_h, S_c, S_y = sampling(S_x, S_h.flatten(), S_c.flatten())
S_y = S_y.flatten()
distribution = helper.t_softmax(S_y, t=t)
S_x = np.random.multinomial(n=1, pvals=distribution)
idx = helper.vec_to_index(S_x)
sampled_token = token_list[idx]
sampled_tokens.append(sampled_token)
counter += 1
return sampled_tokens[1:-1]
def char_sequence_probability(sentence, predictions, embeddings, t=1.0):
log_p = 0
sequence = list(' '.join(sentence)) + ['E']
for c, p in zip(sequence, predictions):
probabilities = helper.t_softmax(p, t=t)
index = embeddings[c]
log_p += np.log(probabilities[index])
return log_p
def char_sequence_probability(sentence, predictions, embeddings, t=1.0):
log_p = 0
sequence = list(' '.join(sentence)) + ['E']
for c, p in zip(sequence, predictions):
probabilities = helper.t_softmax(p, t=t)
index = embeddings[c]
log_p += np.log(probabilities[index])
return log_p