def main():
sess = tf.Session(config=config)
max_seq_len = 40
epoch = 30
batch_size = 128
data = slu_data()
total_intent = data.total_intent
total_word = data.total_word
model = slu_model(max_seq_len, total_intent)
sess.run(model.init_model)
# read in the glove embedding matrix
sess.run(model.init_embedding,
feed_dict={model.read_embedding_matrix: data.embedding_matrix})
test_f1_scores = list()
# Train
for cur_epoch in range(epoch):
pred_outputs = list()
labels = list()
total_loss = 0.0
for cnt in range(50):
# get the data
batch_nl, batch_intent, batch_distance = data.get_train_batch(
batch_size, "Guide")
history_nl, history_nl_len, current_nl, current_nl_len, history_intent, ground_truth, history_distance = \
process_batch(batch_nl, batch_intent, max_seq_len, total_intent-1, total_word-1, total_intent, batch_distance)
_, intent_output, loss = sess.run(
[model.train_op, model.intent_output, model.loss],
feed_dict={
model.labels: ground_truth,
model.current_nl: current_nl,
model.current_nl_len: current_nl_len,
model.history_intent: history_intent,
model.hist_nl: history_nl,
model.hist_nl_len: history_nl_len,
model.dropout_keep_prob: 0.5
})
total_loss += loss
pred, truth = calculate_score(intent_output, ground_truth)
pred_outputs = np.append(pred_outputs, pred)
labels = np.append(labels, truth)
#print "Epoch:", cur_epoch
#print "training f1 score:", f1_score(pred_outputs, labels, average="binary")
#print "training loss:", total_loss
# Test
batch_nl, batch_intent, batch_distance = data.get_test_batch()
history_nl, history_nl_len, current_nl, current_nl_len, history_intent, ground_truth, history_distance = \
process_batch(batch_nl, batch_intent, max_seq_len, total_intent-1, total_word-1, total_intent, batch_distance)
# need to avoid OOM, so test only a batch one time
test_output = None
for i in range(0, len(batch_nl), batch_size):
temp_test_output = sess.run(model.intent_output,
feed_dict={
model.labels:
ground_truth[i:i + batch_size],
model.current_nl:
current_nl[i:i + batch_size],
model.current_nl_len:
current_nl_len[i:i + batch_size],
model.history_intent:
history_intent[i:i + batch_size],
model.hist_nl:
history_nl[i:i + batch_size],
model.hist_nl_len:
history_nl_len[i:i + batch_size],
model.dropout_keep_prob:
1.0
})
if i == 0:
test_output = temp_test_output
else:
test_output = np.vstack((test_output, temp_test_output))
test_pred, test_label = calculate_score(test_output, ground_truth,
"Guide")
f1sc = f1_score(test_pred, test_label, average="binary")
#print "testing f1 score:", f1sc
test_f1_scores.append(f1sc)
#print "max test f1 score:", max(test_f1_scores)
print max(test_f1_scores)
sess.close()
act_logit = one_hot(np.argmax(pred_act), "act")
attribute_logit = binary.fit_transform([pred_attribute])
if np.sum(attribute_logit) == 0:
attribute_logit = one_hot(np.argmax(pred_attribute), "attribute")
label = binary.fit_transform([label])
ret_pred_outputs = np.append(ret_pred_outputs, np.append(act_logit, attribute_logit))
ret_ground_truth = np.append(ret_ground_truth, label)
return ret_pred_outputs, ret_ground_truth
if __name__ == '__main__':
sess = tf.Session(config=config)
max_seq_len = 60
epoch = 30
batch_size = 128
data = slu_data()
total_intent = data.total_intent
total_word = data.total_word
model = slu_model(max_seq_len, total_intent)
sess.run(model.init_model)
# read in the glove embedding matrix
sess.run(model.init_embedding, feed_dict={model.read_embedding_matrix:data.embedding_matrix})
test_f1_scores = list()
# Train
for cur_epoch in range(epoch):
pred_outputs = list()
labels = list()
total_loss = 0.0
for cnt in range(50):
# get the data