def _build(self, state_input, step_input, hidden_input, name=None):
action_dim = self._output_dim
with tf.variable_scope('dist_params'):
if self._std_share_network:
# mean and std networks share an MLP
(outputs, step_outputs, step_hidden, hidden_init_var) = gru(
name='mean_std_network',
gru_cell=self._mean_std_gru_cell,
all_input_var=state_input,
step_input_var=step_input,
step_hidden_var=hidden_input,
hidden_state_init=self._hidden_state_init,
hidden_state_init_trainable=self.
_hidden_state_init_trainable,
output_nonlinearity_layer=self.
_mean_std_output_nonlinearity_layer)
with tf.variable_scope('mean_network'):
mean_var = outputs[..., :action_dim]
step_mean_var = step_outputs[..., :action_dim]
with tf.variable_scope('log_std_network'):
log_std_var = outputs[..., action_dim:]
step_log_std_var = step_outputs[..., action_dim:]
else:
# separate MLPs for mean and std networks
# mean network
(mean_var, step_mean_var, step_hidden, hidden_init_var) = gru(
name='mean_network',
gru_cell=self._mean_gru_cell,
all_input_var=state_input,
step_input_var=step_input,
step_hidden_var=hidden_input,
hidden_state_init=self._hidden_state_init,
hidden_state_init_trainable=self.
_hidden_state_init_trainable,
output_nonlinearity_layer=self.
_mean_output_nonlinearity_layer)
log_std_var = parameter(state_input,
length=action_dim,
initializer=tf.constant_initializer(
self._init_std_param),
trainable=self._learn_std,
name='log_std_param')
step_log_std_var = parameter(
step_input,
length=action_dim,
initializer=tf.constant_initializer(self._init_std_param),
trainable=self._learn_std,
name='step_log_std_param')
dist = DiagonalGaussian(self._output_dim)
rnd = tf.random.normal(shape=step_mean_var.get_shape().as_list()[1:])
action_var = rnd * tf.exp(step_log_std_var) + step_mean_var
return (action_var, mean_var, step_mean_var, log_std_var,
step_log_std_var, step_hidden, hidden_init_var, dist)
def _build(self,
all_input_var,
step_input_var,
step_hidden_var,
name=None):
return gru(
name='gru',
gru_cell=self._gru_cell,
all_input_var=all_input_var,
step_input_var=step_input_var,
step_hidden_var=step_hidden_var,
hidden_state_init=self._hidden_state_init,
hidden_state_init_trainable=self._hidden_state_init_trainable,
output_nonlinearity_layer=self._output_nonlinearity_layer)
def test_output_shapes(self, time_step, input_dim, output_dim, hidden_init,
cell_init):
obs_inputs = np.full((self.batch_size, time_step, input_dim), 1.)
obs_input = np.full((self.batch_size, input_dim), 1.)
input_var = tf.placeholder(tf.float32,
shape=(None, None, input_dim),
name='input')
step_input_var = tf.placeholder(tf.float32,
shape=(None, input_dim),
name='step_input')
output_nonlinearity = tf.keras.layers.Dense(
units=output_dim,
activation=None,
kernel_initializer=tf.constant_initializer(1))
with tf.variable_scope('GRU'):
self.gru = gru(
all_input_var=input_var,
name='gru',
gru_cell=self.gru_cell,
step_input_var=step_input_var,
step_hidden_var=self.step_hidden_var,
hidden_state_init=tf.constant_initializer(hidden_init),
output_nonlinearity_layer=output_nonlinearity)
self.sess.run(tf.global_variables_initializer())
# Compute output by doing t step() on the gru cell
outputs_t, output_t, h_t, hidden_init = self.gru
hidden = np.full((self.batch_size, self.hidden_dim),
hidden_init.eval())
for i in range(time_step):
output, hidden = self.sess.run([output_t, h_t],
feed_dict={
step_input_var: obs_input,
self.step_hidden_var: hidden,
}) # noqa: E126
assert output.shape == (self.batch_size, output_dim)
assert hidden.shape == (self.batch_size, self.hidden_dim)
full_output = self.sess.run(outputs_t,
feed_dict={input_var: obs_inputs})
assert full_output.shape == (self.batch_size, time_step, output_dim)
def test_output_same_as_rnn(self, time_step, input_dim, output_dim,
hidden_init, cell_init):
obs_inputs = np.full((self.batch_size, time_step, input_dim), 1.)
obs_input = np.full((self.batch_size, input_dim), 1.)
input_var = tf.placeholder(tf.float32,
shape=(None, None, input_dim),
name='input')
step_input_var = tf.placeholder(tf.float32,
shape=(None, input_dim),
name='step_input')
output_nonlinearity = tf.keras.layers.Dense(
units=output_dim,
activation=None,
kernel_initializer=tf.constant_initializer(1))
with tf.variable_scope('GRU'):
self.gru = gru(
all_input_var=input_var,
name='gru',
gru_cell=self.gru_cell,
step_input_var=step_input_var,
step_hidden_var=self.step_hidden_var,
hidden_state_init=tf.constant_initializer(hidden_init),
output_nonlinearity_layer=output_nonlinearity)
self.sess.run(tf.global_variables_initializer())
# Create a RNN and compute the entire outputs
rnn_layer = tf.keras.layers.RNN(cell=self.gru_cell,
return_sequences=True,
return_state=True)
# Set initial state to all 0s
hidden_var = tf.get_variable(
name='initial_hidden',
shape=(self.batch_size, self.hidden_dim),
initializer=tf.constant_initializer(hidden_init),
trainable=False,
dtype=tf.float32)
outputs, hiddens = rnn_layer(input_var, initial_state=[hidden_var])
outputs = output_nonlinearity(outputs)
self.sess.run(tf.global_variables_initializer())
outputs, hiddens = self.sess.run([outputs, hiddens],
feed_dict={input_var: obs_inputs})
# Compute output by doing t step() on the gru cell
hidden = np.full((self.batch_size, self.hidden_dim), hidden_init)
_, output_t, hidden_t, _ = self.gru
for i in range(time_step):
output, hidden = self.sess.run([output_t, hidden_t],
feed_dict={
step_input_var: obs_input,
self.step_hidden_var: hidden,
}) # noqa: E126
# The output from i-th timestep
assert np.array_equal(output, outputs[:, i, :])
assert np.array_equal(hidden, hiddens)
# Also the full output from lstm
full_outputs = self.sess.run(self.gru[0],
feed_dict={input_var: obs_inputs})
assert np.array_equal(outputs, full_outputs)
def test_gradient_paths(self):
time_step = 3
input_dim = 2
output_dim = 4
obs_inputs = np.full((self.batch_size, time_step, input_dim), 1.)
obs_input = np.full((self.batch_size, input_dim), 1.)
input_var = tf.placeholder(tf.float32,
shape=(None, None, input_dim),
name='input')
step_input_var = tf.placeholder(tf.float32,
shape=(None, input_dim),
name='step_input')
output_nonlinearity = tf.keras.layers.Dense(
units=output_dim,
activation=None,
kernel_initializer=tf.constant_initializer(1))
with tf.variable_scope('GRU'):
self.gru = gru(all_input_var=input_var,
name='gru',
gru_cell=self.gru_cell,
step_input_var=step_input_var,
step_hidden_var=self.step_hidden_var,
output_nonlinearity_layer=output_nonlinearity)
self.sess.run(tf.global_variables_initializer())
# Compute output by doing t step() on the gru cell
outputs_t, output_t, h_t, hidden_init = self.gru
hidden = np.full((self.batch_size, self.hidden_dim),
hidden_init.eval())
grads_step_o_i = tf.gradients(output_t, step_input_var)
grads_step_o_h = tf.gradients(output_t, self.step_hidden_var)
grads_step_h = tf.gradients(h_t, step_input_var)
self.sess.run([grads_step_o_i, grads_step_o_h, grads_step_h],
feed_dict={
step_input_var: obs_input,
self.step_hidden_var: hidden,
}) # noqa: E126
grads_full = tf.gradients(outputs_t, input_var)
self.sess.run(grads_full, feed_dict={input_var: obs_inputs})
grads_step_o_i = tf.gradients(outputs_t, step_input_var)
grads_step_o_h = tf.gradients(outputs_t, self.step_hidden_var)
grads_step_h = tf.gradients(h_t, input_var)
# No gradient flow
with self.assertRaises(TypeError):
self.sess.run(grads_step_o_i,
feed_dict={
step_input_var: obs_input,
self.step_hidden_var: hidden,
})
with self.assertRaises(TypeError):
self.sess.run(grads_step_o_h,
feed_dict={
step_input_var: obs_input,
self.step_hidden_var: hidden,
})
with self.assertRaises(TypeError):
self.sess.run(grads_step_h, feed_dict={input_var: obs_inputs})
def test_output_value_trainable_hidden_and_cell(self, time_step, input_dim,
output_dim):
obs_inputs = np.full((self.batch_size, time_step, input_dim), 1.)
obs_input = np.full((self.batch_size, input_dim), 1.)
input_var = tf.placeholder(tf.float32,
shape=(None, None, input_dim),
name='input')
step_input_var = tf.placeholder(tf.float32,
shape=(None, input_dim),
name='step_input')
output_nonlinearity = tf.keras.layers.Dense(
units=output_dim,
activation=None,
kernel_initializer=tf.constant_initializer(1))
with tf.variable_scope('GRU'):
self.gru = gru(all_input_var=input_var,
name='gru',
gru_cell=self.gru_cell,
step_input_var=step_input_var,
step_hidden_var=self.step_hidden_var,
hidden_state_init_trainable=True,
output_nonlinearity_layer=output_nonlinearity)
self.sess.run(tf.global_variables_initializer())
# Compute output by doing t step() on the gru cell
outputs_t, output_t, h_t, hidden_init = self.gru
hidden = np.full((self.batch_size, self.hidden_dim),
hidden_init.eval())
output, hidden = self.sess.run([output_t, h_t],
feed_dict={
step_input_var: obs_input,
self.step_hidden_var: hidden,
}) # noqa: E126
with tf.variable_scope('GRU/gru', reuse=True):
hidden_init_var = tf.get_variable(name='initial_hidden')
assert hidden_init_var in tf.trainable_variables()
full_output1 = self.sess.run(outputs_t,
feed_dict={input_var: obs_inputs})
hidden2 = np.full((self.batch_size, self.hidden_dim),
hidden_init.eval())
stack_hidden = None
for i in range(time_step):
hidden2 = recurrent_step_gru(input_val=obs_inputs[:, i, :],
num_units=self.hidden_dim,
step_hidden=hidden2,
w_x_init=1.,
w_h_init=1.,
b_init=0.,
nonlinearity=np.tanh,
gate_nonlinearity=lambda x: 1. /
(1. + np.exp(-x)))
if stack_hidden is None:
stack_hidden = hidden2[:, np.newaxis, :]
else:
stack_hidden = np.concatenate(
(stack_hidden, hidden2[:, np.newaxis, :]), axis=1)
output_nonlinearity = np.full((np.prod(hidden2.shape[1:]), output_dim),
1.)
full_output2 = np.matmul(stack_hidden, output_nonlinearity)
assert np.allclose(full_output1, full_output2)