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.compat.v1.placeholder(tf.float32,
shape=(None, None, input_dim),
name='input')
_step_input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, input_dim),
name='input')
_output_nonlinearity = tf.keras.layers.Dense(
units=output_dim,
activation=None,
kernel_initializer=tf.constant_initializer(1))
with tf.compat.v1.variable_scope('LSTM'):
self.lstm = lstm(
all_input_var=_input_var,
name='lstm',
lstm_cell=self.lstm_cell,
step_input_var=_step_input_var,
step_hidden_var=self._step_hidden_var,
step_cell_var=self._step_cell_var,
hidden_state_init=tf.constant_initializer(hidden_init),
cell_state_init=tf.constant_initializer(cell_init),
output_nonlinearity_layer=_output_nonlinearity)
self.sess.run(tf.compat.v1.global_variables_initializer())
# Compute output by doing t step() on the lstm cell
outputs_t, output_t, h_t, c_t, hidden_init, cell_init = self.lstm
hidden = np.full((self.batch_size, self.hidden_dim),
hidden_init.eval())
cell = np.full((self.batch_size, self.hidden_dim), cell_init.eval())
for _ in range(time_step):
output, hidden, cell = self.sess.run(
[output_t, h_t, c_t],
feed_dict={
_step_input_var: obs_input,
self._step_hidden_var: hidden,
self._step_cell_var: cell
})
assert output.shape == (self.batch_size, output_dim)
assert hidden.shape == (self.batch_size, self.hidden_dim)
assert cell.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 _build(self,
all_input_var,
step_input_var,
step_hidden_var,
step_cell_var,
name=None):
return lstm(
name='lstm',
lstm_cell=self._lstm_cell,
all_input_var=all_input_var,
step_input_var=step_input_var,
step_hidden_var=step_hidden_var,
step_cell_var=step_cell_var,
hidden_state_init=self._hidden_state_init,
hidden_state_init_trainable=self._hidden_state_init_trainable,
cell_state_init=self._cell_state_init,
cell_state_init_trainable=self._cell_state_init_trainable,
output_nonlinearity_layer=self._output_nonlinearity_layer)
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.compat.v1.placeholder(tf.float32,
shape=(None, None, input_dim),
name='input')
_step_input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, input_dim),
name='input')
_output_nonlinearity = tf.keras.layers.Dense(
units=output_dim,
activation=None,
kernel_initializer=tf.constant_initializer(1))
with tf.compat.v1.variable_scope('LSTM'):
self.lstm = lstm(
all_input_var=_input_var,
name='lstm',
lstm_cell=self.lstm_cell,
step_input_var=_step_input_var,
step_hidden_var=self._step_hidden_var,
step_cell_var=self._step_cell_var,
hidden_state_init=tf.constant_initializer(hidden_init),
cell_state_init=tf.constant_initializer(cell_init),
output_nonlinearity_layer=_output_nonlinearity)
self.sess.run(tf.compat.v1.global_variables_initializer())
# Create a RNN and compute the entire outputs
rnn_layer = tf.keras.layers.RNN(cell=self.lstm_cell,
return_sequences=True,
return_state=True)
# Set initial state to all 0s
hidden_var = tf.compat.v1.get_variable(
name='initial_hidden',
shape=(self.batch_size, self.hidden_dim),
initializer=tf.constant_initializer(hidden_init),
trainable=False,
dtype=tf.float32)
cell_var = tf.compat.v1.get_variable(
name='initial_cell',
shape=(self.batch_size, self.hidden_dim),
initializer=tf.constant_initializer(cell_init),
trainable=False,
dtype=tf.float32)
outputs, hiddens, cells = rnn_layer(
_input_var, initial_state=[hidden_var, cell_var])
outputs = _output_nonlinearity(outputs)
self.sess.run(tf.compat.v1.global_variables_initializer())
outputs, hiddens, cells = self.sess.run(
[outputs, hiddens, cells], feed_dict={_input_var: obs_inputs})
# Compute output by doing t step() on the lstm cell
hidden = np.full((self.batch_size, self.hidden_dim), hidden_init)
cell = np.full((self.batch_size, self.hidden_dim), cell_init)
_, output_t, hidden_t, cell_t, _, _ = self.lstm
for i in range(time_step):
output, hidden, cell = self.sess.run(
[output_t, hidden_t, cell_t],
feed_dict={
_step_input_var: obs_input,
self._step_hidden_var: hidden,
self._step_cell_var: cell
})
# The output from i-th timestep
assert np.array_equal(output, outputs[:, i, :])
assert np.array_equal(hidden, hiddens)
assert np.array_equal(cell, cells)
# Also the full output from lstm
full_outputs = self.sess.run(self.lstm[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.compat.v1.placeholder(tf.float32,
shape=(None, None, input_dim),
name='input')
_step_input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, input_dim),
name='input')
_output_nonlinearity = tf.keras.layers.Dense(
units=output_dim,
activation=None,
kernel_initializer=tf.constant_initializer(1))
with tf.compat.v1.variable_scope('LSTM'):
self.lstm = lstm(all_input_var=_input_var,
name='lstm',
lstm_cell=self.lstm_cell,
step_input_var=_step_input_var,
step_hidden_var=self._step_hidden_var,
step_cell_var=self._step_cell_var,
output_nonlinearity_layer=_output_nonlinearity)
self.sess.run(tf.compat.v1.global_variables_initializer())
# Compute output by doing t step() on the lstm cell
outputs_t, output_t, h_t, c_t, hidden_init, cell_init = self.lstm
hidden = np.full((self.batch_size, self.hidden_dim),
hidden_init.eval())
cell = np.full((self.batch_size, self.hidden_dim), cell_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_o_c = tf.gradients(output_t, self._step_cell_var)
grads_step_h = tf.gradients(h_t, _step_input_var)
grads_step_c = tf.gradients(c_t, _step_input_var)
self.sess.run(
[
grads_step_o_i, grads_step_o_h, grads_step_o_c, grads_step_h,
grads_step_c
],
feed_dict={
_step_input_var: obs_input,
self._step_hidden_var: hidden,
self._step_cell_var: cell
})
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_o_c = tf.gradients(outputs_t, self._step_cell_var)
grads_step_h = tf.gradients(h_t, _input_var)
grads_step_c = tf.gradients(c_t, _input_var)
# No gradient flow
with pytest.raises(TypeError):
self.sess.run(grads_step_o_i,
feed_dict={
_step_input_var: obs_input,
self._step_hidden_var: hidden,
self._step_cell_var: cell
})
with pytest.raises(TypeError):
self.sess.run(grads_step_o_h,
feed_dict={
_step_input_var: obs_input,
self._step_hidden_var: hidden,
self._step_cell_var: cell
})
with pytest.raises(TypeError):
self.sess.run(grads_step_o_c,
feed_dict={
_step_input_var: obs_input,
self._step_hidden_var: hidden,
self._step_cell_var: cell
})
with pytest.raises(TypeError):
self.sess.run(grads_step_h, feed_dict={_input_var: obs_inputs})
with pytest.raises(TypeError):
self.sess.run(grads_step_c, 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.compat.v1.placeholder(tf.float32,
shape=(None, None, input_dim),
name='input')
_step_input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, input_dim),
name='input')
_output_nonlinearity = tf.keras.layers.Dense(
units=output_dim,
activation=None,
kernel_initializer=tf.constant_initializer(1))
with tf.compat.v1.variable_scope('LSTM'):
self.lstm = lstm(all_input_var=_input_var,
name='lstm',
lstm_cell=self.lstm_cell,
step_input_var=_step_input_var,
step_hidden_var=self._step_hidden_var,
step_cell_var=self._step_cell_var,
hidden_state_init_trainable=True,
cell_state_init_trainable=True,
output_nonlinearity_layer=_output_nonlinearity)
self.sess.run(tf.compat.v1.global_variables_initializer())
# Compute output by doing t step() on the lstm cell
outputs_t, _, h_t, c_t, hidden_init, cell_init = self.lstm
hidden = np.full((self.batch_size, self.hidden_dim),
hidden_init.eval())
cell = np.full((self.batch_size, self.hidden_dim), cell_init.eval())
hidden, cell = self.sess.run(
[h_t, c_t],
feed_dict={
_step_input_var: obs_input,
self._step_hidden_var: hidden,
self._step_cell_var: cell
})
with tf.compat.v1.variable_scope('LSTM/lstm', reuse=True):
hidden_init_var = tf.compat.v1.get_variable(name='initial_hidden')
cell_init_var = tf.compat.v1.get_variable(name='initial_cell')
assert hidden_init_var in tf.compat.v1.trainable_variables()
assert cell_init_var in tf.compat.v1.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())
cell2 = np.full((self.batch_size, self.hidden_dim), cell_init.eval())
stack_hidden = None
for i in range(time_step):
hidden2, cell2 = recurrent_step_lstm(
input_val=obs_inputs[:, i, :],
num_units=self.hidden_dim,
step_hidden=hidden2,
step_cell=cell2,
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)
def _build(self,
state_input,
step_input,
step_hidden,
step_cell,
name=None):
"""Build model.
Args:
state_input (tf.Tensor): Entire time-series observation input,
with shape :math:`(N, T, S^*)`.
step_input (tf.Tensor): Single timestep observation input,
with shape :math:`(N, S^*)`.
step_hidden (tf.Tensor): Hidden state for step, with shape
:math:`(N, S^*)`.
step_cell (tf.Tensor): Cell state for step, with shape
:math:`(N, S^*)`.
name (str): Inner model name, also the variable scope of the
inner model, if exist. One example is
metarl.tf.models.Sequential.
Returns:
tf.Tensor: Step means, with shape :math:`(N, S^*)`.
tf.Tensor: Step log std, with shape :math:`(N, S^*)`.
tf.Tensor: Step hidden state, with shape :math:`(N, S^*)`.
tf.Tensor: Step cell state, with shape :math:`(N, S^*)`.
tf.Tensor: Initial hidden state, with shape :math:`(S^*)`.
tf.Tensor: Initial cell state, with shape :math:`(S^*)`
tfp.distributions.MultivariateNormalDiag: Policy distribution.
"""
del name
action_dim = self._output_dim
with tf.compat.v1.variable_scope('dist_params'):
if self._std_share_network:
# mean and std networks share an MLP
(outputs, step_outputs, step_hidden, step_cell,
hidden_init_var, cell_init_var) = lstm(
name='mean_std_network',
lstm_cell=self._mean_std_lstm_cell,
all_input_var=state_input,
step_input_var=step_input,
step_hidden_var=step_hidden,
step_cell_var=step_cell,
hidden_state_init=self._hidden_state_init,
hidden_state_init_trainable=self.
_hidden_state_init_trainable,
cell_state_init=self._cell_state_init,
cell_state_init_trainable=self._cell_state_init_trainable,
output_nonlinearity_layer=self.
_mean_std_output_nonlinearity_layer)
with tf.compat.v1.variable_scope('mean_network'):
mean_var = outputs[..., :action_dim]
step_mean_var = step_outputs[..., :action_dim]
with tf.compat.v1.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, step_cell,
hidden_init_var, cell_init_var) = lstm(
name='mean_network',
lstm_cell=self._mean_lstm_cell,
all_input_var=state_input,
step_input_var=step_input,
step_hidden_var=step_hidden,
step_cell_var=step_cell,
hidden_state_init=self._hidden_state_init,
hidden_state_init_trainable=self.
_hidden_state_init_trainable,
cell_state_init=self._cell_state_init,
cell_state_init_trainable=self._cell_state_init_trainable,
output_nonlinearity_layer=self.
_mean_output_nonlinearity_layer)
log_std_var, step_log_std_var = recurrent_parameter(
input_var=state_input,
step_input_var=step_input,
length=action_dim,
initializer=tf.constant_initializer(self._init_std_param),
trainable=self._learn_std,
name='log_std_param')
dist = tfp.distributions.MultivariateNormalDiag(
loc=mean_var, scale_diag=tf.exp(log_std_var))
return (dist, step_mean_var, step_log_std_var, step_hidden, step_cell,
hidden_init_var, cell_init_var)
def _build(self,
state_input,
step_input,
hidden_input,
cell_input,
name=None):
action_dim = self._output_dim
with tf.compat.v1.variable_scope('dist_params'):
if self._std_share_network:
# mean and std networks share an MLP
(outputs, step_outputs, step_hidden, step_cell,
hidden_init_var, cell_init_var) = lstm(
name='mean_std_network',
lstm_cell=self._mean_std_lstm_cell,
all_input_var=state_input,
step_input_var=step_input,
step_hidden_var=hidden_input,
step_cell_var=cell_input,
hidden_state_init=self._hidden_state_init,
hidden_state_init_trainable=self.
_hidden_state_init_trainable,
cell_state_init=self._cell_state_init,
cell_state_init_trainable=self._cell_state_init_trainable,
output_nonlinearity_layer=self.
_mean_std_output_nonlinearity_layer)
with tf.compat.v1.variable_scope('mean_network'):
mean_var = outputs[..., :action_dim]
step_mean_var = step_outputs[..., :action_dim]
with tf.compat.v1.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, step_cell,
hidden_init_var, cell_init_var) = lstm(
name='mean_network',
lstm_cell=self._mean_lstm_cell,
all_input_var=state_input,
step_input_var=step_input,
step_hidden_var=hidden_input,
step_cell_var=cell_input,
hidden_state_init=self._hidden_state_init,
hidden_state_init_trainable=self.
_hidden_state_init_trainable,
cell_state_init=self._cell_state_init,
cell_state_init_trainable=self._cell_state_init_trainable,
output_nonlinearity_layer=self.
_mean_output_nonlinearity_layer)
log_std_var, step_log_std_var = recurrent_parameter(
input_var=state_input,
step_input_var=step_input,
length=action_dim,
initializer=tf.constant_initializer(self._init_std_param),
trainable=self._learn_std,
name='log_std_param')
dist = DiagonalGaussian(self._output_dim)
return (mean_var, step_mean_var, log_std_var, step_log_std_var,
step_hidden, step_cell, hidden_init_var, cell_init_var, dist)