def test_is_pickleable(self):
input_shape = (28, 28, 3)
gcr = GaussianCNNRegressor(input_shape=input_shape,
num_filters=(3, 6),
filter_dims=(3, 3),
strides=(1, 1),
padding='SAME',
hidden_sizes=(32, ),
output_dim=1,
adaptive_std=False,
use_trust_region=False)
with tf.compat.v1.variable_scope(
'GaussianCNNRegressor/GaussianCNNRegressorModel', reuse=True):
bias = tf.compat.v1.get_variable(
'dist_params/mean_network/hidden_0/bias')
bias.load(tf.ones_like(bias).eval())
result1 = gcr.predict([np.ones(input_shape)])
h = pickle.dumps(gcr)
with tf.compat.v1.Session(graph=tf.Graph()):
gcr_pickled = pickle.loads(h)
result2 = gcr_pickled.predict([np.ones(input_shape)])
assert np.array_equal(result1, result2)
def __init__(
self,
env_spec,
subsample_factor=1.,
regressor_args=None,
name='GaussianCNNBaseline',
):
if not isinstance(env_spec.observation_space, akro.Box) or \
not len(env_spec.observation_space.shape) in (2, 3):
raise ValueError(
'{} can only process 2D, 3D akro.Image or'
' akro.Box observations, but received an env_spec with '
'observation_space of type {} and shape {}'.format(
type(self).__name__,
type(env_spec.observation_space).__name__,
env_spec.observation_space.shape))
super().__init__(env_spec)
if regressor_args is None:
regressor_args = dict()
self._regressor = GaussianCNNRegressor(
input_shape=(env_spec.observation_space.shape),
output_dim=1,
subsample_factor=subsample_factor,
name=name,
**regressor_args)
self.name = name
self.env_spec = env_spec
def test_log_likelihood_sym(self, output_dim):
input_shape = (28, 28, 3)
gcr = GaussianCNNRegressor(input_shape=input_shape,
filters=((3, (3, 3)), (6, (3, 3))),
strides=(1, 1),
padding='SAME',
hidden_sizes=(32, ),
output_dim=1,
adaptive_std=False,
use_trust_region=False)
new_input_var = tf.compat.v1.placeholder(tf.float32,
shape=(None, ) + input_shape)
new_ys_var = tf.compat.v1.placeholder(dtype=tf.float32,
name='ys',
shape=(None, output_dim))
data = np.full(input_shape, 0.5)
label = np.ones(output_dim)
outputs = gcr.log_likelihood_sym(new_input_var,
new_ys_var,
name='ll_sym')
ll_from_sym = self.sess.run(outputs,
feed_dict={
new_input_var: [data],
new_ys_var: [label]
})
mean, log_std = gcr._f_pdists([data])
ll = gcr.model.networks['default'].dist.log_likelihood(
[label], dict(mean=mean, log_std=log_std))
assert np.allclose(ll, ll_from_sym, rtol=0, atol=1e-5)
def __init__(
self,
env_spec,
subsample_factor=1.,
regressor_args=None,
name='GaussianCNNBaseline',
):
super().__init__(env_spec)
if regressor_args is None:
regressor_args = dict()
self._regressor = GaussianCNNRegressor(
input_shape=(env_spec.observation_space.shape),
output_dim=1,
subsample_factor=subsample_factor,
name=name,
**regressor_args)
self.name = name
def test_fit_unnormalized(self):
gcr = GaussianCNNRegressor(input_shape=(10, 10, 3),
num_filters=(3, 6),
filter_dims=(3, 3),
strides=(1, 1),
padding='SAME',
hidden_sizes=(32, ),
output_dim=1,
adaptive_std=True,
normalize_inputs=False,
normalize_outputs=False)
train_data, test_data = get_train_test_data()
observations, returns = train_data
for _ in range(20):
gcr.fit(observations, returns)
paths, expected = test_data
prediction = gcr.predict(paths['observations'])
average_error = 0.0
for i in range(len(expected)):
average_error += np.abs(expected[i] - prediction[i])
average_error /= len(expected)
assert average_error <= 0.1
x_mean = self.sess.run(gcr.model.networks['default'].x_mean)
x_mean_expected = np.zeros_like(x_mean)
x_std = self.sess.run(gcr.model.networks['default'].x_std)
x_std_expected = np.ones_like(x_std)
assert np.array_equal(x_mean, x_mean_expected)
assert np.array_equal(x_std, x_std_expected)
y_mean = self.sess.run(gcr.model.networks['default'].y_mean)
y_mean_expected = np.zeros_like(y_mean)
y_std = self.sess.run(gcr.model.networks['default'].y_std)
y_std_expected = np.ones_like(y_std)
assert np.allclose(y_mean, y_mean_expected)
assert np.allclose(y_std, y_std_expected)
def test_fit_normalized(self):
gcr = GaussianCNNRegressor(input_shape=(10, 10, 3),
filters=((3, (3, 3)), (6, (3, 3))),
strides=(1, 1),
padding='SAME',
hidden_sizes=(32, ),
output_dim=1,
adaptive_std=False,
use_trust_region=True)
train_data, test_data = get_train_test_data()
observations, returns = train_data
for _ in range(20):
gcr.fit(observations, returns)
paths, expected = test_data
prediction = gcr.predict(paths['observations'])
average_error = 0.0
for i, exp in enumerate(expected):
average_error += np.abs(exp - prediction[i])
average_error /= len(expected)
assert average_error <= 0.1
x_mean = self.sess.run(gcr.model.networks['default'].x_mean)
x_mean_expected = np.mean(observations, axis=0, keepdims=True)
x_std = self.sess.run(gcr.model.networks['default'].x_std)
x_std_expected = np.std(observations, axis=0, keepdims=True)
assert np.allclose(x_mean, x_mean_expected)
assert np.allclose(x_std, x_std_expected)
y_mean = self.sess.run(gcr.model.networks['default'].y_mean)
y_mean_expected = np.mean(returns, axis=0, keepdims=True)
y_std = self.sess.run(gcr.model.networks['default'].y_std)
y_std_expected = np.std(returns, axis=0, keepdims=True)
assert np.allclose(y_mean, y_mean_expected)
assert np.allclose(y_std, y_std_expected)
def test_fit_without_trusted_region(self):
gcr = GaussianCNNRegressor(input_shape=(10, 10, 3),
filters=((3, (3, 3)), (6, (3, 3))),
strides=(1, 1),
padding='SAME',
hidden_sizes=(32, ),
output_dim=1,
adaptive_std=False,
use_trust_region=False)
train_data, test_data = get_train_test_data()
observations, returns = train_data
for _ in range(20):
gcr.fit(observations, returns)
paths, expected = test_data
prediction = gcr.predict(paths['observations'])
average_error = 0.0
for i, exp in enumerate(expected):
average_error += np.abs(exp - prediction[i])
average_error /= len(expected)
assert average_error <= 0.1
def test_optimizer_args(self, mock_lbfgs):
lbfgs_args = dict(max_opt_itr=25)
gcr = GaussianCNNRegressor(input_shape=(10, 10, 3),
filters=((3, (3, 3)), (6, (3, 3))),
strides=(1, 1),
padding='SAME',
hidden_sizes=(32, ),
output_dim=1,
optimizer=LbfgsOptimizer,
optimizer_args=lbfgs_args,
use_trust_region=True)
assert mock_lbfgs.return_value is gcr._optimizer
mock_lbfgs.assert_called_with(max_opt_itr=25)
def test_fit_smaller_subsample_factor(self):
gcr = GaussianCNNRegressor(input_shape=(10, 10, 3),
num_filters=(3, 6),
filter_dims=(3, 3),
strides=(1, 1),
padding='SAME',
hidden_sizes=(32, ),
output_dim=1,
subsample_factor=0.9,
adaptive_std=False)
train_data, test_data = get_train_test_data()
observations, returns = train_data
for _ in range(20):
gcr.fit(observations, returns)
paths, expected = test_data
prediction = gcr.predict(paths['observations'])
average_error = 0.0
for i in range(len(expected)):
average_error += np.abs(expected[i] - prediction[i])
average_error /= len(expected)
assert average_error <= 0.05
def test_is_pickleable2(self):
input_shape = (28, 28, 3)
gcr = GaussianCNNRegressor(input_shape=input_shape,
filters=((3, (3, 3)), (6, (3, 3))),
strides=(1, 1),
padding='SAME',
hidden_sizes=(32, ),
output_dim=1,
adaptive_std=False,
use_trust_region=False)
with tf.compat.v1.variable_scope(
'GaussianCNNRegressor/GaussianCNNRegressorModel', reuse=True):
x_mean = tf.compat.v1.get_variable('normalized_vars/x_mean')
x_mean.load(tf.ones_like(x_mean).eval())
x1 = gcr.model.networks['default'].x_mean.eval()
h = pickle.dumps(gcr)
with tf.compat.v1.Session(graph=tf.Graph()):
gcr_pickled = pickle.loads(h)
x2 = gcr_pickled.model.networks['default'].x_mean.eval()
assert np.array_equal(x1, x2)
class GaussianCNNBaseline(Baseline):
"""GaussianCNNBaseline With Model.
It fits the input data to a gaussian distribution estimated by a CNN.
Args:
env_spec (metarl.envs.env_spec.EnvSpec): Environment specification.
subsample_factor (float): The factor to subsample the data. By
default it is 1.0, which means using all the data.
regressor_args (dict): Arguments for regressor.
name (str): Name of baseline.
"""
def __init__(
self,
env_spec,
subsample_factor=1.,
regressor_args=None,
name='GaussianCNNBaseline',
):
if not isinstance(env_spec.observation_space, akro.Box) or \
not len(env_spec.observation_space.shape) in (2, 3):
raise ValueError(
'{} can only process 2D, 3D akro.Image or'
' akro.Box observations, but received an env_spec with '
'observation_space of type {} and shape {}'.format(
type(self).__name__,
type(env_spec.observation_space).__name__,
env_spec.observation_space.shape))
super().__init__(env_spec)
if regressor_args is None:
regressor_args = dict()
self._regressor = GaussianCNNRegressor(
input_shape=(env_spec.observation_space.shape),
output_dim=1,
subsample_factor=subsample_factor,
name=name,
**regressor_args)
self.name = name
self.env_spec = env_spec
def fit(self, paths):
"""Fit regressor based on paths.
Args:
paths (dict[numpy.ndarray]): Sample paths.
"""
observations = np.concatenate([p['observations'] for p in paths])
if isinstance(self.env_spec.observation_space, akro.Image):
observations = normalize_pixel_batch(observations)
returns = np.concatenate([p['returns'] for p in paths])
self._regressor.fit(observations, returns.reshape((-1, 1)))
def predict(self, path):
"""Predict value based on paths.
Args:
path (dict[numpy.ndarray]): Sample paths.
Returns:
numpy.ndarray: Predicted value.
"""
observations = path['observations']
if isinstance(self.env_spec.observation_space, akro.Image):
observations = normalize_pixel_batch(observations)
return self._regressor.predict(observations).flatten()
def get_param_values(self):
"""Get parameter values.
Returns:
List[np.ndarray]: A list of values of each parameter.
"""
return self._regressor.get_param_values()
def set_param_values(self, flattened_params):
"""Set param values.
Args:
flattened_params (np.ndarray): A numpy array of parameter values.
"""
self._regressor.set_param_values(flattened_params)
def get_params_internal(self):
"""Get the params, which are the trainable variables.
Returns:
List[tf.Variable]: A list of trainable variables in the current
variable scope.
"""
return self._regressor.get_params_internal()
class GaussianCNNBaseline(Baseline):
"""GaussianCNNBaseline With Model.
It fits the input data to a gaussian distribution estimated by a CNN.
Args:
env_spec (metarl.envs.env_spec.EnvSpec): Environment specification.
subsample_factor (float): The factor to subsample the data. By
default it is 1.0, which means using all the data.
regressor_args (dict): Arguments for regressor.
name (str): Name of baseline.
"""
def __init__(
self,
env_spec,
subsample_factor=1.,
regressor_args=None,
name='GaussianCNNBaseline',
):
super().__init__(env_spec)
if regressor_args is None:
regressor_args = dict()
self._regressor = GaussianCNNRegressor(
input_shape=(env_spec.observation_space.shape),
output_dim=1,
subsample_factor=subsample_factor,
name=name,
**regressor_args)
self.name = name
def fit(self, paths):
"""Fit regressor based on paths.
Args:
paths (dict[numpy.ndarray]): Sample paths.
"""
observations = np.concatenate([p['observations'] for p in paths])
returns = np.concatenate([p['returns'] for p in paths])
self._regressor.fit(observations, returns.reshape((-1, 1)))
def predict(self, path):
"""Predict value based on paths.
Args:
path (dict[numpy.ndarray]): Sample paths.
Returns:
numpy.ndarray: Predicted value.
"""
return self._regressor.predict(path['observations']).flatten()
def get_param_values(self):
"""Get parameter values.
Returns:
List[np.ndarray]: A list of values of each parameter.
"""
return self._regressor.get_param_values()
def set_param_values(self, flattened_params):
"""Set param values.
Args:
flattened_params (np.ndarray): A numpy array of parameter values.
"""
self._regressor.set_param_values(flattened_params)
def get_params_internal(self):
"""Get the params, which are the trainable variables.
Returns:
List[tf.Variable]: A list of trainable variables in the current
variable scope.
"""
return self._regressor.get_params_internal()