def build_q_factor_discrete(self, cfg, goal_embedding):
""""Build the q value network.
Args:
cfg: configuration object
goal_embedding: tensor that contains the goal embedding
Returns:
the q value tensor
"""
n_layer_channel = []
for layer_config in cfg.conv_layer_config:
if layer_config[0] > 0:
n_layer_channel.append(layer_config[0])
layer_film_params = film_params(goal_embedding, n_layer_channel)
out = self.inputs
for cfg, param in zip(cfg.conv_layer_config, layer_film_params):
if cfg[0] < 0:
out = tf.layers.conv2d(out, -cfg[0], cfg[1], cfg[2], padding='SAME')
out = tf.nn.relu(out)
else:
out = tf.layers.conv2d(out, cfg[0], cfg[1], cfg[2], padding='SAME')
out = tf.layers.batch_normalization(
out, center=False, scale=False, training=self.is_training)
gamma, beta = tf.split(param, 2, axis=1)
out *= tf.expand_dims(tf.expand_dims(gamma, 1), 1)
out += tf.expand_dims(tf.expand_dims(beta, 1), 1)
out = tf.nn.relu(out)
out_shape = out.get_shape()
factors = [8, 10, 10]
# [B, s1*s2, s3]
out = tf.reshape(out, (-1, np.prod(out_shape[1:-1]), out_shape[-1]))
projection_mat = tf.get_variable(
name='projection_matrix',
shape=(sum(factors), np.prod(out_shape[1:-1])),
dtype=tf.float32, trainable=True)
projection_mat = tf.expand_dims(projection_mat, axis=0)
projection_mat = tf.tile(projection_mat, [tf.shape(out)[0], 1, 1])
out = tf.matmul(projection_mat, out) # [B, sum(fac), s3]
# [B, factor[0], s3] [B, factor[1], s3] [B, factor[2], s3]
fac1, fac2, fac3 = tf.split(out, factors, axis=1)
out = tensor_concat(fac1, fac2, fac3) # [B, f1, f2, f3, s3]
# [B, 800, s3*3]
out = tf.reshape(out, [-1, np.prod(factors), out_shape[-1]*3])
print('tensor concat: {}'.format(out))
goal_tile = tf.expand_dims(
tf.layers.dense(goal_embedding, out_shape[-1]), 1) # [B, 1, s3]
print('goal: {}'.format(goal_tile))
goal_tile = tf.tile(
goal_tile, multiples=[1, np.prod(factors), 1])
# TODO(ydjiang): include context vector too?
out = tf.concat([out, goal_tile], axis=-1)
out = tf.expand_dims(out, axis=1)
# TODO(ydjiang): wider network here?
out = tf.nn.relu(tf.layers.conv2d(out, 100, 1, 1))
out = tf.nn.relu(tf.layers.conv2d(out, 32, 1, 1))
out = tf.layers.conv2d(out, 1, 1, 1)
return tf.squeeze(out, axis=[1, 3])
def call(self, inputs):
goal_embedding = inputs['goal_embedding']
state_inputs = inputs['state_input']
projection_mat = tf.tile(self._projection_mat,
[tf.shape(state_inputs)[0], 1, 1])
out = tf.reshape(
state_inputs,
(-1, np.prod(self.out_shape[1:-1]), self.out_shape[-1]))
out = tf.matmul(projection_mat, out)
# [B, factor[0], s3] [B, factor[1], s3] [B, factor[2], s3]
fac1, fac2, fac3 = tf.split(out, self.factors, axis=1)
out = tensor_concat(fac1, fac2, fac3) # [B, f1, f2, f3, s3]
# [B, 800, s3*3]
out = tf.reshape(
out, [-1, np.prod(self.factors), self.out_shape[-1] * 3])
goal_tile = tf.expand_dims(self._dense_layer(goal_embedding), 1)
goal_tile = tf.tile(goal_tile, multiples=[1, np.prod(self.factors), 1])
out = tf.concat([out, goal_tile], axis=-1)
out = tf.expand_dims(out, axis=1)
out = tf.nn.relu(self._conv_layer_1(out))
out = tf.nn.relu(self._conv_layer_2(out))
out = self._conv_layer_3(out)
return tf.squeeze(out, axis=[1, 3])