def get_argument(self, value, slot):
if value in self.value2argument:
return self.value2argument[value]
else:
# I must add the argument and the return it
if self.n_arguments_per_slot > 1:
# for i in range(0, self.n_arguments_per_slot * 3):
# n = randint(0, self.n_arguments_per_slot - 1)
for i in range(0, self.n_arguments_per_slot):
ARGUMENT = slot + '_' + str(i)
if ARGUMENT in self.argument2value:
continue
else:
self.add(value, ARGUMENT)
return ARGUMENT
# overwrite an existing argument !!!
m = LogMessage()
m.add('-' * 120)
m.add('WARNING: overwriting an existing argument ARGUMENT = {a}, VALUE = {v}'.format(a=ARGUMENT,
v=value))
m.add('ARGUMENTS: {a}'.format(a=self.argument2value))
m.add('')
m.log()
self.add(value, ARGUMENT)
else:
# there is only one argument per slot
ARGUMENT = slot + '_X'
self.add(value, ARGUMENT)
return ARGUMENT
def get_argument(self, value, slot):
if value in self.value2argument:
return self.value2argument[value]
else:
# I must add the argument and the return it
if self.n_arguments_per_slot > 1:
# for i in range(0, self.n_arguments_per_slot * 3):
# n = randint(0, self.n_arguments_per_slot - 1)
for i in range(0, self.n_arguments_per_slot):
ARGUMENT = slot + '_' + str(i)
if ARGUMENT in self.argument2value:
continue
else:
self.add(value, ARGUMENT)
return ARGUMENT
# overwrite an existing argument !!!
m = LogMessage()
m.add('-' * 120)
m.add(
'WARNING: overwriting an existing argument ARGUMENT = {a}, VALUE = {v}'
.format(a=ARGUMENT, v=value))
m.add('ARGUMENTS: {a}'.format(a=self.argument2value))
m.add('')
m.log()
self.add(value, ARGUMENT)
else:
# there is only one argument per slot
ARGUMENT = slot + '_X'
self.add(value, ARGUMENT)
return ARGUMENT
def log_predictions_dataset(self, log_fn, actions_template, actions_arguments, batch_indexes):
m = LogMessage(log_fn=log_fn)
m.add('Shape of action template predictions: {s}'.format(s=actions_template.shape))
m.add('Shape of action arguments predictions: {s}'.format(s=actions_arguments.shape))
m.add()
m.add('Predictions')
m.add()
# print(self.data.batch_histories.shape)
# print(self.data.batch_actions_template.shape)
# print(self.data.batch_actions_arguments.shape)
# print(actions_template.shape)
# print(actions_arguments.shape)
# print(len(batch_indexes))
for prediction_batch_idx, batch_idx in enumerate(batch_indexes):
for history in range(0, self.data.batch_histories.shape[1]):
m.add('History {h}'.format(h=prediction_batch_idx * self.FLAGS.batch_size + history))
for j in range(self.data.batch_histories.shape[2]):
utterance = []
for k in range(self.data.batch_histories.shape[3]):
w = self.data.idx2word_history[self.data.batch_histories[batch_idx, history, j, k]]
if w not in ['_SOS_', '_EOS_']:
utterance.append(w)
if utterance:
m.add('U {j:2} : {c:80}'.format(j=j, c=' '.join(utterance)))
w_histories_arguments = []
for j in range(self.data.batch_histories_arguments.shape[2]):
w = self.data.idx2word_history_arguments[self.data.batch_histories_arguments[batch_idx, history, j]]
w_histories_arguments.append(w)
m.add('ArgsH: {t:80}'.format(t=', '.join(w_histories_arguments)))
m.add('P : {t:80}'.format(
t=self.data.idx2word_action_template[actions_template[prediction_batch_idx, history]])
)
w_actions_arguments = []
for j in range(actions_arguments.shape[2]):
w = self.data.idx2word_action_arguments[actions_arguments[prediction_batch_idx, history, j]]
w_actions_arguments.append(w)
m.add('ArgsP: {t:80}'.format(t=', '.join(w_actions_arguments)))
m.add('T : {t:80}'.format(
t=self.data.idx2word_action_template[self.data.batch_actions_template[batch_idx, history]])
)
w_actions_arguments = []
for j in range(self.data.batch_actions_arguments.shape[2]):
w = self.data.idx2word_action_arguments[self.data.batch_actions_arguments[batch_idx, history, j]]
w_actions_arguments.append(w)
m.add('ArgsT: {t:80}'.format(t=', '.join(w_actions_arguments)))
m.add()
# m.log(print_console=True, append=False)
m.log(print_console=False, append=False)
def evaluate(self, epoch, learning_rate, sess):
m = LogMessage()
m.add('')
m.add('Epoch: {epoch}'.format(epoch=epoch))
m.add(' - learning rate = {lr:e}'.format(lr=learning_rate.eval()))
def batch_eval(batch_idx):
return sess.run(
[self.predictions, self.loss, self.accuracy],
feed_dict={
self.batch_idx: batch_idx,
self.use_inputs_prob: 1.0,
self.dropout_keep_prob: 1.0,
self.phase_train: False,
})
self.train_predictions_action, train_loss, train_accuracy_action = self.batch_evaluate(
batch_eval, self.data.train_batch_indexes)
self.train_predictions_action_argmax = np.argmax(
self.train_predictions_action, axis=2)
m.add(' Train data')
m.add(' - loss = {lss:f}'.format(lss=train_loss))
m.add(
' - accuracy = {acc:f}'.format(acc=train_accuracy_action))
self.dev_predictions_action, dev_loss, dev_accuracy_action = self.batch_evaluate(
batch_eval, self.data.dev_batch_indexes)
self.dev_predictions_action_argmax = np.argmax(
self.dev_predictions_action, axis=2)
m.add(' Dev data')
m.add(' - loss = {lss:f}'.format(lss=dev_loss))
m.add(' - accuracy = {acc:f}'.format(acc=dev_accuracy_action))
self.test_predictions_action, test_loss, test_accuracy_action = self.batch_evaluate(
batch_eval, self.data.test_batch_indexes)
self.test_predictions_action_argmax = np.argmax(
self.test_predictions_action, axis=2)
m.add(' Test data')
m.add(' - loss = {lss:f}'.format(lss=test_loss))
m.add(' - accuracy = {acc:f}'.format(acc=test_accuracy_action))
m.add()
m.log()
return train_accuracy_action, train_loss, \
dev_accuracy_action, dev_loss, \
test_accuracy_action, test_loss
def evaluate(self, epoch, learning_rate, sess):
m = LogMessage()
m.add('')
m.add('Epoch: {epoch}'.format(epoch=epoch))
m.add(' - learning rate = {lr:e}'.format(lr=learning_rate.eval()))
def batch_eval(batch_idx):
return sess.run(
[self.predictions, self.loss, self.accuracy],
feed_dict={
self.batch_idx: batch_idx,
self.use_inputs_prob: 1.0,
self.dropout_keep_prob: 1.0,
self.phase_train: False,
}
)
self.train_predictions_action, train_loss, train_accuracy_action = self.batch_evaluate(batch_eval, self.data.train_batch_indexes)
self.train_predictions_action_argmax = np.argmax(self.train_predictions_action, axis=2)
m.add(' Train data')
m.add(' - loss = {lss:f}'.format(lss=train_loss))
m.add(' - accuracy = {acc:f}'.format(acc=train_accuracy_action))
self.dev_predictions_action, dev_loss, dev_accuracy_action = self.batch_evaluate(batch_eval, self.data.dev_batch_indexes)
self.dev_predictions_action_argmax = np.argmax(self.dev_predictions_action, axis=2)
m.add(' Dev data')
m.add(' - loss = {lss:f}'.format(lss=dev_loss))
m.add(' - accuracy = {acc:f}'.format(acc=dev_accuracy_action))
self.test_predictions_action, test_loss, test_accuracy_action = self.batch_evaluate(batch_eval, self.data.test_batch_indexes)
self.test_predictions_action_argmax = np.argmax(self.test_predictions_action, axis=2)
m.add(' Test data')
m.add(' - loss = {lss:f}'.format(lss=test_loss))
m.add(' - accuracy = {acc:f}'.format(acc=test_accuracy_action))
m.add()
m.log()
return train_accuracy_action, train_loss, \
dev_accuracy_action, dev_loss, \
test_accuracy_action, test_loss
def log_predictions_dataset(self, log_fn, actions, batch_indexes):
m = LogMessage(log_fn=log_fn)
m.add('Shape of action predictions: {s}'.format(s=actions.shape))
m.add('Argmax predictions')
m.add()
for prediction_batch_idx, batch_idx in enumerate(batch_indexes):
for history in range(0, self.data.batch_histories.shape[1]):
m.add('History {h}'.format(
h=prediction_batch_idx * self.FLAGS.batch_size + history))
for j in range(self.data.batch_histories.shape[2]):
utterance = []
for k in range(self.data.batch_histories.shape[3]):
w = self.data.idx2word_history[
self.data.batch_histories[batch_idx, history, j,
k]]
if w not in ['_SOS_', '_EOS_']:
utterance.append(w)
if utterance:
m.add('U {j}: {c:80}'.format(j=j,
c=' '.join(utterance)))
prediction = []
for j in range(actions.shape[2]):
w = self.data.idx2word_action[actions[prediction_batch_idx,
history, j]]
if w not in ['_SOS_', '_EOS_']:
prediction.append(w)
m.add('P : {t:80}'.format(t=' '.join(prediction)))
target = []
for j in range(self.data.batch_actions.shape[2]):
w = self.data.idx2word_action[self.data.batch_actions[
batch_idx, history, j]]
if w not in ['_SOS_', '_EOS_']:
target.append(w)
m.add('T : {t:80}'.format(t=' '.join(target)))
m.add()
# m.log(print_console=True, append=False)
m.log(print_console=False, append=False)
def log_predictions_dataset(self, log_fn, actions_template, batch_indexes):
m = LogMessage(log_fn=log_fn)
m.add('Shape of action template predictions: {s}'.format(
s=actions_template.shape))
m.add()
m.add('Predictions')
m.add()
# print(self.data.batch_histories.shape)
# print(self.data.batch_actions_template.shape)
# print(actions_template.shape)
# print(len(batch_indexes))
for prediction_batch_idx, batch_idx in enumerate(batch_indexes):
for history in range(0, self.data.batch_histories.shape[1]):
m.add('History {h}'.format(
h=prediction_batch_idx * self.FLAGS.batch_size + history))
for j in range(self.data.batch_histories.shape[2]):
utterance = []
for k in range(self.data.batch_histories.shape[3]):
w = self.data.idx2word_history[
self.data.batch_histories[batch_idx, history, j,
k]]
if w not in ['_SOS_', '_EOS_']:
utterance.append(w)
if utterance:
m.add('U {j}: {c:80}'.format(j=j,
c=' '.join(utterance)))
m.add(
'P : {t:80}'.format(t=self.data.idx2word_action_template[
actions_template[prediction_batch_idx, history]]))
m.add(
'T : {t:80}'.format(t=self.data.idx2word_action_template[
self.data.batch_actions_template[batch_idx, history]]))
m.add()
# m.log()
m.log(print_console=False, append=False)
def log_predictions_dataset(self, log_fn, actions, batch_indexes):
m = LogMessage(log_fn=log_fn)
m.add('Shape of action predictions: {s}'.format(s=actions.shape))
m.add('Argmax predictions')
m.add()
for prediction_batch_idx, batch_idx in enumerate(batch_indexes):
for history in range(0, self.data.batch_histories.shape[1]):
m.add('History {h}'.format(h=prediction_batch_idx * self.FLAGS.batch_size + history))
for j in range(self.data.batch_histories.shape[2]):
utterance = []
for k in range(self.data.batch_histories.shape[3]):
w = self.data.idx2word_history[self.data.batch_histories[batch_idx, history, j, k]]
if w not in ['_SOS_', '_EOS_']:
utterance.append(w)
if utterance:
m.add('U {j}: {c:80}'.format(j=j, c=' '.join(utterance)))
prediction = []
for j in range(actions.shape[2]):
w = self.data.idx2word_action[actions[prediction_batch_idx, history, j]]
if w not in ['_SOS_', '_EOS_']:
prediction.append(w)
m.add('P : {t:80}'.format(t=' '.join(prediction)))
target = []
for j in range(self.data.batch_actions.shape[2]):
w = self.data.idx2word_action[self.data.batch_actions[batch_idx, history, j]]
if w not in ['_SOS_', '_EOS_']:
target.append(w)
m.add('T : {t:80}'.format(t=' '.join(target)))
m.add()
# m.log(print_console=True, append=False)
m.log(print_console=False, append=False)
def train(data, model):
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
inter_op_parallelism_threads=FLAGS.threads,
intra_op_parallelism_threads=FLAGS.threads,
use_per_session_threads=True)) as sess:
# Merge all the summaries and write them out to ./log
merged_summaries = tf.merge_all_summaries()
writer = tf.train.SummaryWriter(logging.exp_dir, sess.graph_def)
saver = tf.train.Saver()
# training
t_vars = tf.trainable_variables()
# t_vars = [v for v in t_vars if 'embedding_table' not in v.name] # all variables except embeddings
learning_rate = tf.Variable(float(FLAGS.learning_rate), trainable=False)
# train_op = tf.train.AdagradOptimizer(
# learning_rate=learning_rate,
# )
# train_op = AdamPlusCovOptimizer(
train_op = AdamPlusOptimizer(
learning_rate=learning_rate,
beta1=FLAGS.beta1,
beta2=FLAGS.beta2,
epsilon=FLAGS.epsilon,
pow=FLAGS.pow,
dense_regularization=FLAGS.dense_regularization,
sparse_regularization=FLAGS.sparse_regularization,
use_locking=False,
name='trainer')
learning_rate_decay_op = learning_rate.assign(learning_rate * FLAGS.decay)
global_step = tf.Variable(0, trainable=False)
gradients = tf.gradients(model.loss, t_vars)
clipped_gradients, _ = tf.clip_by_global_norm(gradients, FLAGS.max_gradient_norm)
train_op = train_op.apply_gradients(zip(clipped_gradients, t_vars), global_step=global_step)
tf.initialize_all_variables().run()
# prepare batch indexes
m = LogMessage()
train_set_size = model.data.train_set_size
m.add('Train set size: {d}'.format(d=train_set_size))
batch_size = FLAGS.batch_size
m.add('Batch size: {d}'.format(d=batch_size))
m.add('#Batches: {d}'.format(d=len(model.data.train_batch_indexes)))
m.log()
train_accuracies, train_losses = [], []
dev_accuracies, dev_losses = [], []
test_accuracies, test_losses = [], []
max_accuracy_epoch = 0
use_inputs_prob = 1.0
for epoch in range(FLAGS.max_epochs):
# update the model
LogMessage.write('Batch: ')
for b, batch_idx in enumerate(model.data.train_batch_indexes):
LogMessage.write(b)
LogMessage.write(' ')
sess.run(
[train_op],
feed_dict={
model.batch_idx: batch_idx,
model.use_inputs_prob: use_inputs_prob,
model.dropout_keep_prob: FLAGS.dropout_keep_prob,
model.phase_train: True,
}
)
shuffle(model.data.train_batch_indexes)
LogMessage.write('\n\n')
LogMessage.write('Used inputs prob = {uip:f}'.format(uip=use_inputs_prob))
LogMessage.write('\n')
# evaluate the model
train_acc, train_lss, \
dev_acc, dev_lss, \
test_acc, test_lss = \
model.evaluate(epoch, learning_rate, sess)
if epoch == 0 or dev_acc > max(dev_accuracies):
max_accuracy_epoch = epoch
model_fn = saver.save(sess, os.path.join(logging.exp_dir, "model.ckpt"))
m = LogMessage()
m.add('New max accuracy achieved on the dev data.')
m.add("Model saved in file: {s}".format(s=model_fn))
m.log()
# save predictions on train, dev, and test sets
model.log_predictions()
m = LogMessage()
m.add()
m.add("Epoch with max accuracy on dev data: {d}".format(d=max_accuracy_epoch))
m.add()
m.log()
# decrease learning rate if no improvement was seen over last 4 episodes.
if len(train_losses) > 6 and train_lss >= max(train_losses[-4:]) + 1e-10:
sess.run(learning_rate_decay_op)
train_losses.append(train_lss)
train_accuracies.append(train_acc)
dev_losses.append(dev_lss)
dev_accuracies.append(dev_acc)
test_losses.append(test_lss)
test_accuracies.append(test_acc)
# stop when reached a threshold maximum or when no improvement of accuracy in the last 100 steps
if train_acc > .999 or epoch > max_accuracy_epoch + 100:
break
use_inputs_prob *= FLAGS.use_inputs_prob_decay
# save the results
results = {
'epoch': epoch,
'max_accuracy_epoch_on_dev_data': max_accuracy_epoch,
'train_loss': str(train_losses[max_accuracy_epoch]),
'train_accuracy': str(train_accuracies[max_accuracy_epoch]),
'dev_loss': str(dev_losses[max_accuracy_epoch]),
'dev_accuracy': str(dev_accuracies[max_accuracy_epoch]),
'test_loss': str(test_losses[max_accuracy_epoch]),
'test_accuracy': str(test_accuracies[max_accuracy_epoch]),
}
LogExperiment(results)
LogMessage(log_fn='.done', msg='done', time=True).log()
def main(run):
start_experiment(run)
if FLAGS.runs == 1:
# set the seed to constant
seed(0)
tf.set_random_seed(1)
graph = tf.Graph()
with graph.as_default():
with graph.device(device_for_node_gpu if FLAGS.gpu else device_for_node_cpu):
if 'w2t' in FLAGS.model:
FLAGS.task = 'w2t'
if 'w2w' in FLAGS.model:
FLAGS.task = 'w2w'
if 'w2targs' in FLAGS.model:
FLAGS.task = 'w2targs'
m = LogMessage(time=True)
m.add('-' * 120)
m.add('End to End Neural Dialogue Manager')
m.add(' runs = {runs}'.format(runs=FLAGS.runs))
m.add(' threads = {threads}'.format(threads=FLAGS.threads))
m.add(' gpu = {gpu}'.format(gpu=FLAGS.gpu))
m.add(' model = {model}'.format(model=FLAGS.model))
m.add(' input = {i}'.format(i=FLAGS.input))
m.add(' data_fraction = {data_fraction}'.format(data_fraction=FLAGS.data_fraction))
m.add(' train_data = {train_data}'.format(train_data=FLAGS.train_data))
m.add(' dev_data = {dev_data}'.format(dev_data=FLAGS.dev_data))
m.add(' test_data = {test_data}'.format(test_data=FLAGS.test_data))
m.add(' ontology = {ontology}'.format(ontology=FLAGS.ontology))
m.add(' database = {database}'.format(database=FLAGS.database))
m.add(' max_epochs = {max_epochs}'.format(max_epochs=FLAGS.max_epochs))
m.add(' batch_size = {batch_size}'.format(batch_size=FLAGS.batch_size))
m.add(' learning_rate = {learning_rate:2e}'.format(learning_rate=FLAGS.learning_rate))
m.add(' decay = {decay}'.format(decay=FLAGS.decay))
m.add(' beta1 = {beta1}'.format(beta1=FLAGS.beta1))
m.add(' beta2 = {beta2}'.format(beta2=FLAGS.beta2))
m.add(' epsilon = {epsilon}'.format(epsilon=FLAGS.epsilon))
m.add(' pow = {pow}'.format(pow=FLAGS.pow))
m.add(' dense_regularization = {regularization}'.format(regularization=FLAGS.dense_regularization))
m.add(' sparse_regularization = {regularization}'.format(regularization=FLAGS.sparse_regularization))
m.add(
' max_gradient_norm = {max_gradient_norm}'.format(max_gradient_norm=FLAGS.max_gradient_norm))
m.add(' use_inputs_prob_decay = {use_inputs_prob_decay}'.format(
use_inputs_prob_decay=FLAGS.use_inputs_prob_decay))
m.add(
' dropout_keep_prob = {dropout_keep_prob}'.format(dropout_keep_prob=FLAGS.dropout_keep_prob))
m.add('-' * 120)
m.log()
data = dataset.DSTC2(
input=FLAGS.input,
data_fraction=FLAGS.data_fraction,
train_data_fn=FLAGS.train_data,
dev_data_fn=FLAGS.dev_data,
test_data_fn=FLAGS.test_data,
ontology_fn=FLAGS.ontology,
database_fn=FLAGS.database,
batch_size=FLAGS.batch_size
)
m = LogMessage()
m.add('Database # rows: {d}'.format(d=len(data.database)))
m.add('Database # columns: {d}'.format(d=len(data.database_word2idx.keys())))
m.add('History vocabulary size: {d}'.format(d=len(data.idx2word_history)))
m.add('History args. vocabulary size: {d}'.format(d=len(data.idx2word_history_arguments)))
m.add('State vocabulary size: {d}'.format(d=len(data.idx2word_state)))
m.add('Action vocabulary size: {d}'.format(d=len(data.idx2word_action)))
m.add('Action args. vocabulary size: {d}'.format(d=len(data.idx2word_action_arguments)))
m.add('Action tmpl. vocabulary size: {d}'.format(d=len(data.idx2word_action_template)))
m.add('-' * 120)
m.log()
if FLAGS.model == 'cnn-w2w':
model = cnn_w2w.Model(data, FLAGS)
elif FLAGS.model == 'rnn-w2w':
model = rnn_w2w.Model(data, FLAGS)
elif FLAGS.model == 'cnn12-w2t':
model = cnn12_w2t.Model(data, FLAGS)
elif FLAGS.model == 'cnn12-bn-w2targs':
model = cnn12_bn_w2targs.Model(data, FLAGS)
elif FLAGS.model == 'cnn12-bn-w2t':
model = cnn12_bn_w2t.Model(data, FLAGS)
elif FLAGS.model == 'cnn12-mp-bn-w2t':
model = cnn12_mp_bn_w2t.Model(data, FLAGS)
elif FLAGS.model == 'cnn13-bn-w2t':
model = cnn13_bn_w2t.Model(data, FLAGS)
elif FLAGS.model == 'cnn13-mp-bn-w2t':
model = cnn13_mp_bn_w2t.Model(data, FLAGS)
elif FLAGS.model == 'cnn12-att-a-w2t':
model = cnn12_att_a_w2t.Model(data, FLAGS)
elif FLAGS.model == 'cnn12-bn-att-a-w2targs':
model = cnn12_bn_att_a_w2targs.Model(data, FLAGS)
elif FLAGS.model == 'cnn12-bn-att-a-w2t':
model = cnn12_bn_att_a_w2t.Model(data, FLAGS)
elif FLAGS.model == 'cnn12-bn-att-a-bn-w2t':
model = cnn12_bn_att_a_bn_w2t.Model(data, FLAGS)
elif FLAGS.model == 'cnn12-mp-bn-att-a-w2t':
model = cnn12_mp_bn_att_a_w2t.Model(data, FLAGS)
elif FLAGS.model == 'cnn12-att-b-w2t':
model = cnn12_att_b_w2t.Model(data, FLAGS)
elif FLAGS.model == 'cnn23-mp-bn-w2t':
model = cnn23_mp_bn_w2t.Model(data, FLAGS)
elif FLAGS.model == 'rnn1-w2t':
model = rnn1_w2t.Model(data, FLAGS)
elif FLAGS.model == 'rnn2-w2t':
model = rnn2_w2t.Model(data, FLAGS)
else:
raise Exception('Error: Unsupported model - {m}'.format(m=FLAGS.model))
train(data, model)
summary_hash = new_summary_hash
# run only if we have some stats
m = LogMessage(time=True)
m.add('-' * 80)
m.add('Experiment summary')
m.add(' runs = {runs}'.format(runs=FLAGS.runs))
m.add()
m.add(' epoch min = {d}'.format(d=min(epoch)))
m.add(' max = {d}'.format(d=max(epoch)))
m.add(' max_accuracy_epoch min = {d}'.format(d=min(max_accuracy_epoch_on_dev_data)))
m.add(' max = {d}'.format(d=max(max_accuracy_epoch_on_dev_data)))
m.add()
m.add(' dev acc max = {f:6f}'.format(f=max(dev_accuracy)))
m.add(' mean = {f:6f}'.format(f=mean(dev_accuracy)))
if FLAGS.runs > 1:
m.add(' stdev = {f:6f}'.format(f=stdev(dev_accuracy)))
m.add(' min = {f:6f}'.format(f=min(dev_accuracy)))
m.add()
m.log()
if done_runs >= len(ps):
# stop this loop when all runs are finished
break
# for i, p in enumerate(ps):
# p.join()
# print('Joining process {d}'.format(d=i))
print('All done')