def make_policy():
"""Returns one copy of the model."""
artifact = {}
if cfg.intruction_repr == 'language':
trainable_encoder = cfg.trainable_encoder
print('The encoder is trainable: {}'.format(trainable_encoder))
embedding = tf.get_variable(
name='word_embedding',
shape=(cfg.vocab_size, cfg.embedding_size),
dtype=tf.float32,
trainable=trainable_encoder)
_, goal_embedding = encoder(
self.word_inputs,
embedding,
cfg.encoder_n_unit,
trainable=trainable_encoder)
artifact['embedding'] = embedding
elif cfg.intruction_repr == 'one_hot':
print('Goal input for one-hot max len {}'.format(
cfg.max_sequence_length))
one_hot_goal = tf.one_hot(self.word_inputs, cfg.max_sequence_length)
one_hot_goal.set_shape([None, cfg.max_sequence_length])
layer_cfg = [cfg.max_sequence_length // 8, cfg.encoder_n_unit]
goal_embedding = stack_dense_layer(one_hot_goal, layer_cfg)
else:
raise ValueError('Unrecognized instruction type: {}'.format(
cfg.instruction_repr))
artifact['goal_embedding'] = goal_embedding
all_q = self.build_q_factor_discrete(cfg, goal_embedding)
predict_action = tf.argmax(all_q, axis=-1)
action = tf.placeholder(shape=None, dtype=tf.int32)
action_onehot = tf.one_hot(
action, cfg.ac_dim[0], dtype=tf.float32)
q = tf.reduce_sum(
tf.multiply(all_q, action_onehot), axis=1)
artifact.update(
{
'all_q': all_q,
'predict_action': predict_action,
'action_ph': action,
'action_onehot': action_onehot,
'q': q
}
)
return artifact
def make_policy():
"""Build one copy of the model."""
artifact = {}
if cfg.intruction_repr == 'language':
trainable_encoder = cfg.trainable_encoder
print('The encoder is trainable: {}'.format(trainable_encoder))
embedding = tf.get_variable(name='word_embedding',
shape=(cfg.vocab_size,
cfg.embedding_size),
dtype=tf.float32,
trainable=trainable_encoder)
_, goal_embedding = encoder(self.word_inputs,
embedding,
cfg.encoder_n_unit,
trainable=trainable_encoder)
artifact['embedding'] = embedding
elif cfg.intruction_repr == 'one_hot':
print('Goal input for one-hot max len {}'.format(
cfg.max_sequence_length))
one_hot_goal = tf.one_hot(self.word_inputs,
cfg.max_sequence_length)
one_hot_goal.set_shape([None, cfg.max_sequence_length])
layer_cfg = [cfg.max_sequence_length // 8, cfg.encoder_n_unit]
goal_embedding = stack_dense_layer(one_hot_goal, layer_cfg)
else:
raise ValueError('Unrecognized instruction type: {}'.format(
cfg.instruction_repr))
artifact['goal_embedding'] = goal_embedding
if cfg.action_type == 'perfect':
print('using perfect action Q function...')
all_q, predict_object, predict_object_action = self.build_q_perfect(
cfg, goal_embedding)
predict_action = tf.stack(
[predict_object, predict_object_action], axis=1)
action = tf.placeholder(shape=(None, 2), dtype=tf.int32)
stacked_indices = tf.concat([
tf.expand_dims(tf.range(0,
tf.shape(action)[0]), axis=1),
action
],
axis=1)
q = tf.gather_nd(all_q, stacked_indices)
artifact.update({
'all_q': all_q,
'predict_object': predict_object,
'predict_object_action': predict_object_action,
'predict_action': predict_action,
'action_ph': action,
'q': q,
})
elif cfg.action_type == 'discrete':
print('using discrete action Q function...')
ac_dim = cfg.per_input_ac_dim[0]
all_q = self.build_q_discrete(goal_embedding, ac_dim)
predict_action = tf.argmax(all_q, axis=-1)
action = tf.placeholder(shape=None, dtype=tf.int32)
action_onehot = tf.one_hot(action, ac_dim, dtype=tf.float32)
q = tf.reduce_sum(tf.multiply(all_q, action_onehot), axis=1)
artifact.update({
'all_q': all_q,
'predict_action': predict_action,
'action_ph': action,
'action_onehot': action_onehot,
'q': q,
})
else:
raise ValueError('Unrecognized action type: {}'.format(
cfg.action_type))
return artifact
