def __init__(self, data, FLAGS):
super(Model, self).__init__(data, FLAGS)
database_size = data.database.shape[0]
database_column_embedding_size = 8
n_database_columns = len(data.database_columns)
conv_mul = 2
histories_embedding_size = 16
histories_vocabulary_length = len(data.idx2word_history)
histories_utterance_length = data.train_set['histories'].shape[2]
history_length = data.train_set['histories'].shape[1]
histories_arguments_embedding_size = 8
histories_arguments_vocabulary_length = len(data.idx2word_history_arguments)
n_histories_arguments = data.train_set['histories_arguments'].shape[1]
action_templates_vocabulary_length = len(data.idx2word_action_template)
with tf.name_scope('data'):
database = tf.Variable(data.database, name='database',
trainable=False)
batch_histories = tf.Variable(data.batch_histories, name='histories',
trainable=False)
batch_histories_arguments = tf.Variable(data.batch_histories_arguments, name='histories_arguments',
trainable=False)
batch_actions_template = tf.Variable(data.batch_actions_template, name='actions',
trainable=False)
histories = tf.gather(batch_histories, self.batch_idx)
histories_arguments = tf.gather(batch_histories_arguments, self.batch_idx)
actions_template = tf.gather(batch_actions_template, self.batch_idx)
with tf.name_scope('model'):
with tf.variable_scope("batch_size"):
batch_size = tf.shape(histories)[0]
database_embedding = multicolumn_embedding(
columns=database,
lengths=[len(i2w) for i2w in [data.database_idx2word[column] for column in data.database_columns]],
sizes=[database_column_embedding_size for column in data.database_columns],
# all columns have the same size
name='database_embedding'
)
histories_embedding = embedding(
input=histories,
length=histories_vocabulary_length,
size=histories_embedding_size,
name='histories_embedding'
)
histories_arguments_embedding = embedding(
input=histories_arguments,
length=histories_arguments_vocabulary_length,
size=histories_arguments_embedding_size,
name='histories_arguments_embedding'
)
with tf.name_scope("UtterancesEncoder"):
conv3 = histories_embedding
# conv3 = dropout(conv3, pow_1(self.dropout_keep_prob, 2))
conv3 = conv2d_bn(
input=conv3,
filter=[1, 3, conv3.size, conv3.size * conv_mul],
phase_train=self.phase_train,
name='conv_utt_size_3_layer_1'
)
encoded_utterances = reduce_max(conv3, [2], keep_dims=True, name='encoded_utterances')
with tf.name_scope("HistoryEncoder"):
conv3 = encoded_utterances
conv3 = dropout(conv3, pow_1(self.dropout_keep_prob, 2))
conv3 = conv2d_bn(
input=conv3,
filter=[3, 1, conv3.size, conv3.size * conv_mul],
phase_train=self.phase_train,
name='conv_hist_size_3_layer_1'
)
conv3 = dropout(conv3, pow_1(self.dropout_keep_prob, 2))
conv3 = conv2d_bn(
input=conv3,
filter=[3, 1, conv3.size, conv3.size * conv_mul],
phase_train=self.phase_train,
name='conv_hist_size_3_layer_2'
)
encoded_history = reduce_max(conv3, [1, 2], name='encoded_history')
# print(encoded_history)
with tf.name_scope("DatabaseAttention"):
histories_arguments_embedding = tf.reshape(
histories_arguments_embedding,
[-1, n_histories_arguments * histories_arguments_embedding_size],
name='histories_arguments_embedding'
)
# print(histories_arguments_embedding)
history_predicate = tf.concat(
1,
[encoded_history, histories_arguments_embedding],
name='history_predicate'
)
print(history_predicate)
att_W_nx = conv3.size + n_histories_arguments * histories_arguments_embedding_size
att_W_ny = n_database_columns * database_column_embedding_size
att_W = tf.get_variable(
name='attention_W',
shape=[att_W_nx, att_W_ny],
initializer=tf.random_uniform_initializer(
-glorot_mul(att_W_nx, att_W_ny),
glorot_mul(att_W_nx, att_W_ny)
),
)
hp_x_att_W = tf.matmul(history_predicate, att_W)
attention_scores = tf.matmul(hp_x_att_W, database_embedding, transpose_b=True)
attention_scores = batch_norm_lin(attention_scores, database_size, self.phase_train, name='attention_scores_bn')
attention = tf.nn.softmax(attention_scores, name="attention_softmax")
print(attention)
attention_max = tf.reduce_max(attention, reduction_indices=1, keep_dims=True)
attention_min = tf.reduce_min(attention, reduction_indices=1, keep_dims=True)
attention_mean = tf.reduce_mean(attention_scores, reduction_indices=1, keep_dims=True)
attention_feat = tf.concat(1, [attention_max, attention_mean, attention_min], name='attention_feat')
attention_feat_size = 3
print(attention_feat)
db_result = tf.matmul(attention, database_embedding, name='db_result')
db_result_size = att_W_ny
print(db_result)
with tf.name_scope("Decoder"):
second_to_last_user_utterance = encoded_utterances[:, history_length - 3, 0, :]
last_system_utterance = encoded_utterances[:, history_length - 2, 0, :]
last_user_utterance = encoded_utterances[:, history_length - 1, 0, :]
dialogue_state = tf.concat(
1,
[
encoded_history,
last_user_utterance,
last_system_utterance,
second_to_last_user_utterance,
attention_feat,
db_result
],
name='dialogue_state'
)
dialogue_state_size = conv3.size + \
3 * histories_embedding_size * conv_mul + \
attention_feat_size + \
db_result_size
activation = tf.nn.relu(dialogue_state)
activation = dropout(activation, self.dropout_keep_prob)
projection = linear(
input=activation,
input_size=dialogue_state_size,
output_size=dialogue_state_size,
name='linear_projection_1'
)
projection = batch_norm_lin(projection, dialogue_state_size, self.phase_train, name='linear_projection_1_bn')
activation = tf.nn.relu(projection)
activation = dropout(activation, self.dropout_keep_prob)
projection = linear(
input=activation,
input_size=dialogue_state_size,
output_size=dialogue_state_size,
name='linear_projection_2'
)
projection = batch_norm_lin(projection, dialogue_state_size, self.phase_train, name='linear_projection_2_bn')
activation = tf.nn.relu(projection)
activation = dropout(activation, self.dropout_keep_prob)
projection = linear(
input=activation,
input_size=dialogue_state_size,
output_size=action_templates_vocabulary_length,
name='linear_projection_3'
)
self.predictions = tf.nn.softmax(projection, name="predictions")
# print(self.predictions)
if FLAGS.print_variables:
for v in tf.trainable_variables():
print(v.name)
with tf.name_scope('loss'):
one_hot_labels = dense_to_one_hot(actions_template, action_templates_vocabulary_length)
self.loss = tf.reduce_mean(- one_hot_labels * tf.log(tf.clip_by_value(self.predictions, 1e-10, 1.0)), name='loss')
tf.scalar_summary('loss', self.loss)
with tf.name_scope('accuracy'):
correct_prediction = tf.equal(tf.argmax(one_hot_labels, 1), tf.argmax(self.predictions, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
tf.scalar_summary('accuracy', self.accuracy)
def __init__(self, data, FLAGS):
super(Model, self).__init__(data, FLAGS)
database_column_embedding_size = 8
n_database_columns = len(data.database_columns)
conv_mul = 2
histories_embedding_size = 16
histories_vocabulary_length = len(data.idx2word_history)
histories_utterance_length = data.train_set['histories'].shape[2]
history_length = data.train_set['histories'].shape[1]
histories_arguments_embedding_size = 8
histories_arguments_vocabulary_length = len(data.idx2word_history_arguments)
n_histories_arguments = data.train_set['histories_arguments'].shape[1]
action_templates_vocabulary_length = len(data.idx2word_action_template)
with tf.name_scope('data'):
database = tf.Variable(data.database, name='database',
trainable=False)
batch_histories = tf.Variable(data.batch_histories, name='histories',
trainable=False)
batch_histories_arguments = tf.Variable(data.batch_histories_arguments, name='histories_arguments',
trainable=False)
batch_actions_template = tf.Variable(data.batch_actions_template, name='actions',
trainable=False)
histories = tf.gather(batch_histories, self.batch_idx)
histories_arguments = tf.gather(batch_histories_arguments, self.batch_idx)
actions_template = tf.gather(batch_actions_template, self.batch_idx)
with tf.name_scope('model'):
with tf.variable_scope("batch_size"):
batch_size = tf.shape(histories)[0]
database_embedding = multicolumn_embedding(
columns=database,
lengths=[len(i2w) for i2w in [data.database_idx2word[column] for column in data.database_columns]],
sizes=[database_column_embedding_size for column in data.database_columns],
# all columns have the same size
name='database_embedding'
)
histories_embedding = embedding(
input=histories,
length=histories_vocabulary_length,
size=histories_embedding_size,
name='histories_embedding'
)
histories_arguments_embedding = embedding(
input=histories_arguments,
length=histories_arguments_vocabulary_length,
size=histories_arguments_embedding_size,
name='histories_arguments_embedding'
)
with tf.name_scope("UtterancesEncoder"):
conv3 = histories_embedding
# conv3 = dropout(conv3, pow_1(self.dropout_keep_prob, 2))
conv3 = conv2d_bn(
input=conv3,
filter=[1, 3, conv3.size, conv3.size * conv_mul],
phase_train=self.phase_train,
name='conv_utt_size_3_layer_1'
)
encoded_utterances = reduce_max(conv3, [2], keep_dims=True, name='encoded_utterances')
with tf.name_scope("HistoryEncoder"):
conv3 = encoded_utterances
conv3 = dropout(conv3, pow_1(self.dropout_keep_prob, 2))
conv3 = conv2d_bn(
input=conv3,
filter=[3, 1, conv3.size, conv3.size * conv_mul],
phase_train=self.phase_train,
name='conv_hist_size_3_layer_1'
)
conv3 = dropout(conv3, pow_1(self.dropout_keep_prob, 2))
conv3 = conv2d_bn(
input=conv3,
filter=[3, 1, conv3.size, conv3.size * conv_mul],
phase_train=self.phase_train,
name='conv_hist_size_3_layer_2'
)
encoded_history = reduce_max(conv3, [1, 2], name='encoded_history')
# print(encoded_history)
with tf.name_scope("DatabaseAttention"):
histories_arguments_embedding = tf.reshape(
histories_arguments_embedding,
[-1, n_histories_arguments * histories_arguments_embedding_size],
name='histories_arguments_embedding'
)
# print(histories_arguments_embedding)
history_predicate = tf.concat(
1,
[encoded_history, histories_arguments_embedding],
name='history_predicate'
)
print(history_predicate)
att_W_nx = conv3.size + n_histories_arguments * histories_arguments_embedding_size
att_W_ny = n_database_columns * database_column_embedding_size
att_W = tf.get_variable(
name='attention_W',
shape=[att_W_nx, att_W_ny],
initializer=tf.random_uniform_initializer(
-glorot_mul(att_W_nx, att_W_ny),
glorot_mul(att_W_nx, att_W_ny)
),
)
hp_x_att_W = tf.matmul(history_predicate, att_W)
attention_scores = tf.matmul(hp_x_att_W, database_embedding, transpose_b=True)
attention = tf.nn.softmax(attention_scores, name="attention_softmax")
print(attention)
attention_max = tf.reduce_max(attention, reduction_indices=1, keep_dims=True)
attention_min = tf.reduce_min(attention, reduction_indices=1, keep_dims=True)
attention_mean = tf.reduce_mean(attention_scores, reduction_indices=1, keep_dims=True)
attention_feat = tf.concat(1, [attention_max, attention_mean, attention_min], name='attention_feat')
attention_feat_size = 3
print(attention_feat)
db_result = tf.matmul(attention, database_embedding, name='db_result')
db_result_size = att_W_ny
print(db_result)
with tf.name_scope("Decoder"):
second_to_last_user_utterance = encoded_utterances[:, history_length - 3, 0, :]
last_system_utterance = encoded_utterances[:, history_length - 2, 0, :]
last_user_utterance = encoded_utterances[:, history_length - 1, 0, :]
dialogue_state = tf.concat(
1,
[
encoded_history,
last_user_utterance,
last_system_utterance,
second_to_last_user_utterance,
attention_feat,
db_result
],
name='dialogue_state'
)
dialogue_state_size = (
conv3.size +
3 * histories_embedding_size * conv_mul +
attention_feat_size +
db_result_size +
0
)
activation = tf.nn.relu(dialogue_state)
activation = dropout(activation, self.dropout_keep_prob)
projection = linear(
input=activation,
input_size=dialogue_state_size,
output_size=dialogue_state_size,
name='linear_projection_1'
)
projection = batch_norm_lin(projection, dialogue_state_size, self.phase_train, name='linear_projection_1_bn')
activation = tf.nn.relu(projection)
activation = dropout(activation, self.dropout_keep_prob)
projection = linear(
input=activation,
input_size=dialogue_state_size,
output_size=dialogue_state_size,
name='linear_projection_2'
)
projection = batch_norm_lin(projection, dialogue_state_size, self.phase_train, name='linear_projection_2_bn')
activation = tf.nn.relu(projection)
activation = dropout(activation, self.dropout_keep_prob)
projection = linear(
input=activation,
input_size=dialogue_state_size,
output_size=action_templates_vocabulary_length,
name='linear_projection_3'
)
self.predictions = tf.nn.softmax(projection, name="predictions")
# print(self.predictions)
if FLAGS.print_variables:
for v in tf.trainable_variables():
print(v.name)
with tf.name_scope('loss'):
one_hot_labels = dense_to_one_hot(actions_template, action_templates_vocabulary_length)
self.loss = tf.reduce_mean(- one_hot_labels * tf.log(tf.clip_by_value(self.predictions, 1e-10, 1.0)), name='loss')
tf.scalar_summary('loss', self.loss)
with tf.name_scope('accuracy'):
correct_prediction = tf.equal(tf.argmax(one_hot_labels, 1), tf.argmax(self.predictions, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
tf.scalar_summary('accuracy', self.accuracy)
def __init__(self, data, FLAGS):
super(Model, self).__init__(data, FLAGS)
database_column_embedding_size = 8
n_database_columns = len(data.database_columns)
conv_mul = 2
histories_embedding_size = 16
histories_vocabulary_length = len(data.idx2word_history)
history_length = data.train_set["histories"].shape[1]
histories_arguments_embedding_size = 8
histories_arguments_vocabulary_length = len(data.idx2word_history_arguments)
n_histories_arguments = data.train_set["histories_arguments"].shape[1]
action_templates_vocabulary_length = len(data.idx2word_action_template)
action_templates_embedding_size = 8
num_actions_arguments = data.batch_actions_arguments.shape[2]
actions_arguments_vocabulary_length = len(data.idx2word_action_arguments)
with tf.name_scope("data"):
database = tf.Variable(data.database, name="database", trainable=False)
batch_histories = tf.Variable(data.batch_histories, name="histories", trainable=False)
batch_histories_arguments = tf.Variable(
data.batch_histories_arguments, name="histories_arguments", trainable=False
)
batch_actions_template = tf.Variable(data.batch_actions_template, name="actions", trainable=False)
batch_action_arguments = tf.Variable(
data.batch_actions_arguments, name="actions_arguments", trainable=False
)
histories = tf.gather(batch_histories, self.batch_idx)
histories_arguments = tf.gather(batch_histories_arguments, self.batch_idx)
actions_template = tf.gather(batch_actions_template, self.batch_idx)
actions_arguments = tf.gather(batch_action_arguments, self.batch_idx)
with tf.name_scope("model"):
database_embedding = multicolumn_embedding(
columns=database,
lengths=[len(i2w) for i2w in [data.database_idx2word[column] for column in data.database_columns]],
sizes=[database_column_embedding_size for column in data.database_columns],
# all columns have the same size
name="database_embedding",
)
histories_embedding = embedding(
input=histories,
length=histories_vocabulary_length,
size=histories_embedding_size,
name="histories_embedding",
)
histories_arguments_embedding = embedding(
input=histories_arguments,
length=histories_arguments_vocabulary_length,
size=histories_arguments_embedding_size,
name="histories_arguments_embedding",
)
with tf.name_scope("UtterancesEncoder"):
conv3 = histories_embedding
# conv3 = dropout(conv3, pow_1(self.dropout_keep_prob, 2))
conv3 = conv2d_bn(
input=conv3,
filter=[1, 3, conv3.size, conv3.size * conv_mul],
phase_train=self.phase_train,
name="conv_utt_size_3_layer_1",
)
encoded_utterances = reduce_max(conv3, [2], keep_dims=True, name="encoded_utterances")
with tf.name_scope("HistoryEncoder"):
conv3 = encoded_utterances
conv3 = dropout(conv3, pow_1(self.dropout_keep_prob, 2))
conv3 = conv2d_bn(
input=conv3,
filter=[3, 1, conv3.size, conv3.size * conv_mul],
phase_train=self.phase_train,
name="conv_hist_size_3_layer_1",
)
conv3 = dropout(conv3, pow_1(self.dropout_keep_prob, 2))
conv3 = conv2d_bn(
input=conv3,
filter=[3, 1, conv3.size, conv3.size * conv_mul],
phase_train=self.phase_train,
name="conv_hist_size_3_layer_2",
)
encoded_history = reduce_max(conv3, [1, 2], name="encoded_history")
# print(encoded_history)
with tf.name_scope("DatabaseAttention"):
histories_arguments_embedding = tf.reshape(
histories_arguments_embedding,
[-1, n_histories_arguments * histories_arguments_embedding_size],
name="histories_arguments_embedding",
)
# print(histories_arguments_embedding)
history_predicate = tf.concat(
1, [encoded_history, histories_arguments_embedding], name="history_predicate"
)
# print(history_predicate)
att_W_nx = conv3.size + n_histories_arguments * histories_arguments_embedding_size
att_W_ny = n_database_columns * database_column_embedding_size
att_W = tf.get_variable(
name="attention_W",
shape=[att_W_nx, att_W_ny],
initializer=tf.random_uniform_initializer(
-glorot_mul(att_W_nx, att_W_ny), glorot_mul(att_W_nx, att_W_ny)
),
)
hp_x_att_W = tf.matmul(history_predicate, att_W)
attention_scores = tf.matmul(hp_x_att_W, database_embedding, transpose_b=True)
attention = tf.nn.softmax(attention_scores, name="attention_softmax")
print(attention)
attention_max = tf.reduce_max(attention, reduction_indices=1, keep_dims=True)
attention_min = tf.reduce_min(attention, reduction_indices=1, keep_dims=True)
attention_mean = tf.reduce_mean(attention_scores, reduction_indices=1, keep_dims=True)
attention_feat = tf.concat(1, [attention_max, attention_mean, attention_min], name="attention_feat")
attention_feat_size = 3
# print(attention_feat)
db_result = tf.matmul(attention, database_embedding, name="db_result")
db_result_size = att_W_ny
# print(db_result)
with tf.name_scope("Decoder"):
second_to_last_user_utterance = encoded_utterances[:, history_length - 3, 0, :]
last_system_utterance = encoded_utterances[:, history_length - 2, 0, :]
last_user_utterance = encoded_utterances[:, history_length - 1, 0, :]
dialogue_state = tf.concat(
1,
[
encoded_history,
last_user_utterance,
last_system_utterance,
second_to_last_user_utterance,
attention_feat,
db_result,
],
name="dialogue_state",
)
dialogue_state_size = (
conv3.size + 3 * histories_embedding_size * conv_mul + attention_feat_size + db_result_size + 0
)
dialogue_state = tf.nn.relu(dialogue_state)
dialogue_state = dropout(dialogue_state, self.dropout_keep_prob)
# action prediction
projection = linear(
input=dialogue_state,
input_size=dialogue_state_size,
output_size=dialogue_state_size,
name="linear_projection_1",
)
projection = batch_norm_lin(
projection, dialogue_state_size, self.phase_train, name="linear_projection_1_bn"
)
activation = tf.nn.relu(projection)
activation = dropout(activation, self.dropout_keep_prob)
projection = linear(
input=activation,
input_size=dialogue_state_size,
output_size=dialogue_state_size,
name="linear_projection_2",
)
projection = batch_norm_lin(
projection, dialogue_state_size, self.phase_train, name="linear_projection_2_bn"
)
activation = tf.nn.relu(projection)
activation = dropout(activation, self.dropout_keep_prob)
projection = linear(
input=activation,
input_size=dialogue_state_size,
output_size=action_templates_vocabulary_length,
name="linear_projection_3_predictions_action",
)
self.predictions_action = tf.nn.softmax(projection, name="softmax_output_prediction_action")
# argument prediction
# first encode decoded action template and teh true action template
choice = tf.floor(tf.random_uniform([1], self.use_inputs_prob, 1 + self.use_inputs_prob, tf.float32))
prediction_action_argmax = tf.stop_gradient(tf.argmax(self.predictions_action, 1))
predicted_action_templates_embedding = embedding(
input=prediction_action_argmax,
length=action_templates_vocabulary_length,
size=action_templates_embedding_size,
name="action_templates_embedding",
)
true_action_template_embedding = tf.gather(
predicted_action_templates_embedding.embedding_table, actions_template
)
predicted_action_templates_embedding = tf.stop_gradient(predicted_action_templates_embedding)
action_templates_embedding = (
choice * true_action_template_embedding + (1.0 - choice) * predicted_action_templates_embedding
)
dialogue_state_action_template = tf.concat(
1, [dialogue_state, action_templates_embedding], name="dialogue_state_action_template"
)
dialogue_state_action_template_size = dialogue_state_size + action_templates_embedding_size
# condition on the dialogue state and the decoded template
projection = linear(
input=dialogue_state_action_template,
input_size=dialogue_state_action_template_size,
output_size=dialogue_state_action_template_size,
name="linear_projection_1_predictions_arguments",
)
projection = batch_norm_lin(
projection,
dialogue_state_action_template_size,
self.phase_train,
name="linear_projection_1_predictions_arguments_bn",
)
activation = tf.nn.relu(projection)
activation = dropout(activation, self.dropout_keep_prob)
projection = linear(
input=activation,
input_size=dialogue_state_action_template_size,
output_size=dialogue_state_action_template_size,
name="linear_projection_2_predictions_arguments",
)
projection = batch_norm_lin(
projection,
dialogue_state_action_template_size,
self.phase_train,
name="linear_projection_2_predictions_arguments_bn",
)
activation = tf.nn.relu(projection)
activation = dropout(activation, self.dropout_keep_prob)
projection = linear(
input=activation,
input_size=dialogue_state_action_template_size,
output_size=num_actions_arguments * actions_arguments_vocabulary_length,
name="linear_projection_3_predictions_arguments",
)
self.predictions_arguments = softmax_2d(
input=projection,
n_classifiers=num_actions_arguments,
n_classes=actions_arguments_vocabulary_length,
name="softmax_2d_predictions_arguments",
)
if FLAGS.print_variables:
for v in tf.trainable_variables():
print(v.name)
with tf.name_scope("loss"):
one_hot_labels_action = dense_to_one_hot(actions_template, action_templates_vocabulary_length)
one_hot_labels_arguments = dense_to_one_hot(actions_arguments, actions_arguments_vocabulary_length)
loss_action = tf.reduce_mean(
-one_hot_labels_action * tf.log(tf.clip_by_value(self.predictions_action, 1e-10, 1.0)), name="loss"
)
loss_arguments = tf.reduce_mean(
-one_hot_labels_arguments * tf.log(tf.clip_by_value(self.predictions_arguments, 1e-10, 1.0)),
name="loss",
)
self.loss = (loss_action + loss_arguments) / 2
tf.scalar_summary("loss", self.loss)
with tf.name_scope("accuracy"):
correct_prediction_action = tf.equal(
tf.argmax(one_hot_labels_action, 1), tf.argmax(self.predictions_action, 1)
)
self.accuracy_action = tf.reduce_mean(tf.cast(correct_prediction_action, "float"))
tf.scalar_summary("accuracy_action", self.accuracy_action)
correct_prediction_arguments = tf.equal(
tf.argmax(one_hot_labels_arguments, 2), tf.argmax(self.predictions_arguments, 2)
)
self.accuracy_arguments = tf.reduce_mean(tf.cast(correct_prediction_arguments, "float"))
tf.scalar_summary("accuracy_arguments", self.accuracy_arguments)
self.accuracy = (self.accuracy_action + self.accuracy_arguments) / 2
tf.scalar_summary("accuracy", self.accuracy)