def create_adam_learner(learn_params,
learning_rate=0.0005,
gradient_clipping_threshold_per_sample=0.001):
"""
Create adam learner
"""
lr_schedule = learners.learning_rate_schedule(learning_rate,
learners.UnitType.sample)
momentum = learners.momentum_schedule(0.90)
gradient_clipping_threshold_per_sample = gradient_clipping_threshold_per_sample
gradient_clipping_with_truncation = True
momentum_var = learners.momentum_schedule(0.999)
lr = learners.adam(
learn_params,
lr_schedule,
momentum,
True,
momentum_var,
gradient_clipping_threshold_per_sample=
gradient_clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
learner_desc = 'Alg: Adam, learning rage: {0}, momentum: {1}, gradient clip: {2}'.format(
learning_rate, momentum[0], gradient_clipping_threshold_per_sample)
logger.log("Create learner. {0}".format(learner_desc))
return lr
def __init__(self, input_shape, nb_actions,
gamma=0.99, explorer=LinearEpsilonAnnealingExplorer(1, 0.1, 1000000),
learning_rate=0.00025, momentum=0.95, minibatch_size=32,
memory_size=500000, train_after=10000, train_interval=4,
target_update_interval=10000, monitor=True):
self.input_shape = input_shape
self.nb_actions = nb_actions
self.gamma = gamma
self._train_after = train_after
self._train_interval = train_interval
self._target_update_interval = target_update_interval
self._explorer = explorer
self._minibatch_size = minibatch_size
self._history = History(input_shape)
self._memory = RepMem(memory_size, input_shape[1:], 4)
self._num_actions_taken = 0
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
with default_options(activation=relu, init=he_uniform()):
self._action_value_net = Sequential([
Dense(input_shape, init=he_uniform(scale=0.01)),
Dense(input_shape),
Dense(nb_actions, activation=None, init=he_uniform(scale=0.01))])
self._action_value_net.update_signature(Tensor[input_shape])
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
@Function
@Signature(post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def compute_q_targets(post_states, rewards, terminals):
return element_select(
terminals,
rewards,
gamma * reduce_max(self._target_net(post_states), axis=0) + rewards,
)
@Function
@Signature(pre_states=Tensor[input_shape], actions=Tensor[nb_actions],
post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def criterion(pre_states, actions, post_states, rewards, terminals):
q_targets = compute_q_targets(post_states, rewards, terminals)
q_acted = reduce_sum(self._action_value_net(pre_states) * actions, axis=0)
return huber_loss(q_targets, q_acted, 1.0)
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_schedule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
l_sgd = adam(self._action_value_net.parameters, lr_schedule,
momentum=m_schedule, variance_momentum=vm_schedule)
self._metrics_writer = TensorBoardProgressWriter(freq=1, log_dir='metrics', model=criterion) if monitor else None
self._learner = l_sgd
self._trainer = Trainer(criterion, (criterion, None), l_sgd, self._metrics_writer)
def build_trainer(self):
# Set the learning rate, and the momentum parameters for the Adam optimizer.
lr = learning_rate_schedule(self.lr, UnitType.minibatch)
beta1 = momentum_schedule(0.9)
beta2 = momentum_schedule(0.99)
# Calculate the losses.
loss_on_v = cntk.squared_error(self.R, self.v)
pi_a_s = cntk.log(cntk.times_transpose(self.pi, self.action))
loss_on_pi = cntk.variables.Constant(-1) * (cntk.plus(
cntk.times(pi_a_s, cntk.minus(self.R, self.v_calc)),
0.01 * cntk.times_transpose(self.pi, cntk.log(self.pi))))
#loss_on_pi = cntk.times(pi_a_s, cntk.minus(self.R, self.v_calc))
self.tensorboard_v_writer = TensorBoardProgressWriter(
freq=10, log_dir="tensorboard_v_logs", model=self.v)
self.tensorboard_pi_writer = TensorBoardProgressWriter(
freq=10, log_dir="tensorboard_pi_logs", model=self.pi)
# tensorboard --logdir=tensorboard_pi_logs http://localhost:6006/
# tensorboard --logdir=tensorboard_v_logs http://localhost:6006/
# Create the trainiers.
self.trainer_v = cntk.Trainer(self.v, (loss_on_v), [
adam(self.pms_v,
lr,
beta1,
variance_momentum=beta2,
gradient_clipping_threshold_per_sample=2,
l2_regularization_weight=0.01)
], self.tensorboard_v_writer)
self.trainer_pi = cntk.Trainer(self.pi, (loss_on_pi), [
adam(self.pms_pi,
lr,
beta1,
variance_momentum=beta2,
gradient_clipping_threshold_per_sample=2,
l2_regularization_weight=0.01)
], self.tensorboard_pi_writer)
def __init__(self,
name,
observation_space_shape,
num_actions,
pretrained_policy=None,
*args,
**kwargs):
self.name = name
self.observation_space_shape = observation_space_shape
self.num_actions = num_actions
self._build_network(pretrained_policy)
self.trainer = Trainer(
self.value, self.loss,
[adam(self.value.parameters, lr=0.001, momentum=0.9)])
def __init__(self,
name,
num_frames_to_stack,
observation_space_shape,
num_actions,
pretrained_policy=None,
*args,
**kwargs):
self.name = name
self.num_frames_to_stack = num_frames_to_stack
self.observation_space_shape = observation_space_shape
self.num_actions = num_actions
self._build_network(pretrained_policy)
self.trainer = Trainer(
self.log_probability, self.loss,
[adam(self.probabilities.parameters, lr=0.00001, momentum=0.9)])
def __init__(self, input_shape, nb_actions,
gamma=0.95, explorer=LinearEpsilonAnnealingExplorer(1, 0.1, 100000),
learning_rate=0.01, momentum=0.8, minibatch_size=16,
memory_size=15000, train_after=100, train_interval=100, target_update_interval=500,
monitor=True):
self.input_shape = input_shape
self.nb_actions = nb_actions
self.gamma = gamma
self._train_after = train_after
self._train_interval = train_interval
self._target_update_interval = target_update_interval
self._explorer = explorer
self._minibatch_size = minibatch_size
self._history = History(input_shape)
self._memory = ReplayMemory(memory_size, input_shape[1:], 4)
self._num_actions_taken = 0
self._num_trains = 0
# Metrics accumulator
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
'''
# Action Value model (used by agent to interact with the environment)
with default_options(activation=relu, init=he_uniform()):
self._action_value_net = Sequential([
Convolution2D((8, 8), 16, strides=4),
Convolution2D((4, 4), 32, strides=2),
Convolution2D((3, 3), 32, strides=1),
Dense(256, init=he_uniform(scale=0.01)),
Dense(nb_actions, activation=None, init=he_uniform(scale=0.01))
])
'''
with default_options(activation=relu, init=he_uniform()):
self._action_value_net = Sequential([
Dense(7, init=he_uniform(scale=0.01)),
Dense(8, init=he_uniform(scale=0.01)),
#Dense(16, init=he_uniform(scale=0.01)),
#Dense(32, init=he_uniform(scale=0.01)),
Dense(nb_actions, activation=None, init=he_uniform(scale=0.01))
])
self._action_value_net.update_signature(Tensor[input_shape])
# Target model used to compute the target Q-values in training, updated
# less frequently for increased stability.
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
# Function computing Q-values targets as part of the computation graph
@Function
@Signature(post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def compute_q_targets(post_states, rewards, terminals):
return element_select(
terminals,
rewards,
gamma * reduce_max(self._target_net(post_states), axis=0) + rewards,
)
# Define the loss, using Huber Loss (more robust to outliers)
@Function
@Signature(pre_states=Tensor[input_shape], actions=Tensor[nb_actions],
post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def criterion(pre_states, actions, post_states, rewards, terminals):
# Compute the q_targets
q_targets = compute_q_targets(post_states, rewards, terminals)
# actions is a 1-hot encoding of the action done by the agent
q_acted = reduce_sum(self._action_value_net(pre_states) * actions, axis=0)
# Define training criterion as the Huber Loss function
return huber_loss(q_targets, q_acted, 1.0)
# Adam based SGD
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_schedule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
l_sgd = adam(self._action_value_net.parameters, lr_schedule,
momentum=m_schedule, variance_momentum=vm_schedule)
self._metrics_writer = TensorBoardProgressWriter(freq=1, log_dir='metrics', model=criterion) if monitor else None
self._learner = l_sgd
self._trainer = Trainer(criterion, (criterion, None), l_sgd, self._metrics_writer)
def __init__(self,
in_shape,
output_shape,
device_id=None,
learning_rate=0.00025,
momentum=0.9,
minibatch_size=32,
update_interval=10000,
n_workers=1,
visualizer=None):
"""
Q Neural Network following Mnih and al. implementation and default options.
The network has the following topology:
Convolution(32, (8, 8))
Convolution(64, (4, 4))
Convolution(64, (2, 2))
Dense(512)
:param in_shape: Shape of the observations perceived by the learner (the neural net input)
:param output_shape: Size of the action space (mapped to the number of output neurons)
:param device_id: Use None to let CNTK select the best available device,
-1 for CPU, >= 0 for GPU
(default: None)
:param learning_rate: Learning rate
(default: 0.00025, as per Mnih et al.)
:param momentum: Momentum, provided as momentum value for
averaging gradients without unit gain filter
Note that CNTK does not currently provide an implementation
of Graves' RmsProp with momentum.
It uses AdamSGD optimizer instead.
(default: 0, no momentum with RProp optimizer)
:param minibatch_size: Minibatch size
(default: 32, as per Mnih et al.)
:param n_workers: Number of concurrent worker for distributed training.
(default: 1, not distributed)
:param visualizer: Optional visualizer allowing the model to save summary data
(default: None, no visualization)
Ref: Mnih et al.: "Human-level control through deep reinforcement learning."
Nature 518.7540 (2015): 529-533.
"""
assert learning_rate > 0, 'learning_rate should be > 0'
assert 0. <= momentum < 1, 'momentum should be 0 <= momentum < 1'
QModel.__init__(self, in_shape, output_shape)
CntkModel.__init__(self, device_id, False, n_workers, visualizer)
self._nb_actions = output_shape
self._steps = 0
self._target_update_interval = update_interval
self._target = None
# Input vars
self._environment = input(in_shape,
name='env',
dynamic_axes=(Axis.default_batch_axis()))
self._q_targets = input(1,
name='q_targets',
dynamic_axes=(Axis.default_batch_axis()))
self._actions = input(output_shape,
name='actions',
dynamic_axes=(Axis.default_batch_axis()))
# Define the neural network graph
self._model = self._build_model()(self._environment)
self._target = self._model.clone(
CloneMethod.freeze, {self._environment: self._environment})
# Define the learning rate
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
# AdamSGD optimizer
m_schedule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
l_sgd = adam(self._model.parameters,
lr_schedule,
momentum=m_schedule,
unit_gain=True,
variance_momentum=vm_schedule)
if self.distributed_training:
raise NotImplementedError('ASGD not implemented yet.')
# _actions is a sparse 1-hot encoding of the actions done by the agent
q_acted = reduce_sum(self._model * self._actions, axis=0)
# Define the trainer with Huber Loss function
criterion = huber_loss(q_acted, self._q_targets, 1.0)
self._learner = l_sgd
self._trainer = Trainer(self._model, (criterion, None), l_sgd)
def __init__(self, input_shape, nb_actions,
gamma=0.99, explorer=LinearEpsilonAnnealingExplorer(1, 0.1, 1000000),
learning_rate=0.00025, momentum=0.95, minibatch_size=32,
memory_size=500000, train_after=10000, train_interval=4, target_update_interval=10000,
monitor=True):
self.input_shape = input_shape
self.nb_actions = nb_actions
self.gamma = gamma
self._train_after = train_after
self._train_interval = train_interval
self._target_update_interval = target_update_interval
self._explorer = explorer
self._minibatch_size = minibatch_size
self._history = History(input_shape)
self._memory = ReplayMemory(memory_size, input_shape[1:], 4)
self._num_actions_taken = 0
# Metrics accumulator
self._episode_rewards, self._episode_q_means, self._episode_q_stddev = [], [], []
# Action Value model (used by agent to interact with the environment)
with default_options(activation=relu, init=he_uniform()):
self._action_value_net = Sequential([
Convolution2D((8, 8), 16, strides=4),
Convolution2D((4, 4), 32, strides=2),
Convolution2D((3, 3), 32, strides=1),
Dense(256, init=he_uniform(scale=0.01)),
Dense(nb_actions, activation=None, init=he_uniform(scale=0.01))
])
self._action_value_net.update_signature(Tensor[input_shape])
# Target model used to compute the target Q-values in training, updated
# less frequently for increased stability.
self._target_net = self._action_value_net.clone(CloneMethod.freeze)
# Function computing Q-values targets as part of the computation graph
@Function
@Signature(post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def compute_q_targets(post_states, rewards, terminals):
return element_select(
terminals,
rewards,
gamma * reduce_max(self._target_net(post_states), axis=0) + rewards,
)
# Define the loss, using Huber Loss (more robust to outliers)
@Function
@Signature(pre_states=Tensor[input_shape], actions=Tensor[nb_actions],
post_states=Tensor[input_shape], rewards=Tensor[()], terminals=Tensor[()])
def criterion(pre_states, actions, post_states, rewards, terminals):
# Compute the q_targets
q_targets = compute_q_targets(post_states, rewards, terminals)
# actions is a 1-hot encoding of the action done by the agent
q_acted = reduce_sum(self._action_value_net(pre_states) * actions, axis=0)
# Define training criterion as the Huber Loss function
return huber_loss(q_targets, q_acted, 1.0)
# Adam based SGD
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_schedule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
l_sgd = adam(self._action_value_net.parameters, lr_schedule,
momentum=m_schedule, variance_momentum=vm_schedule)
self._metrics_writer = TensorBoardProgressWriter(freq=1, log_dir='metrics', model=criterion) if monitor else None
self._learner = l_sgd
self._trainer = Trainer(criterion, (criterion, None), l_sgd, self._metrics_writer)
# define loss / metrics
# like Tensorflow the softmax is done
# internally (if needed), so all we need are the logits
ce = cross_entropy_with_softmax(logits, labels)
pe = classification_error(logits, labels)
# training config
batch_size = 32
epochs = 15
n_batches = len(Xtrain) // batch_size
# do the training
# specify the training algorithm
trainer = Trainer(logits, (ce, pe),
adam(logits.parameters, lr=1e-2, momentum=0.9))
# helper function
def get_output(node, X, Y):
ret = node.forward(dict(inputs=X, labels=Y))
return list(ret[1].values())[0].mean()
costs = []
errors = []
test_costs = []
test_errors = []
for i in range(epochs):
cost = 0
err = 0
# internally (if needed), so all we need are the logits ce = cross_entropy_with_softmax(logits, labels) pe = classification_error(logits, labels) # training config batch_size = 32 epochs = 15 n_batches = len(Xtrain) // batch_size # do the training # specify the training algorithm trainer = Trainer(logits, (ce, pe), adam(logits.parameters, lr=1e-2, momentum=0.9)) # helper function def get_output(node, X, Y): ret = node.forward(dict(inputs=X, labels=Y)) return list(ret[1].values())[0].mean() costs = [] errors = [] test_costs = [] test_errors = [] for i in range(epochs): cost = 0 err = 0
def train_and_evaluate(reader_train,
reader_test,
network_name,
epoch_size,
max_epochs,
minibatch_size,
model_dir=None,
log_dir=None,
tensorboard_logdir=None,
gen_heartbeat=False,
fp16=False):
"""
:param reader_train:
:param reader_test:
:param network_name:
:param epoch_size: 一个epoch有多少样本
:param max_epochs: 训练多少个epoch
:param model_dir:
:param log_dir:
:param tensorboard_logdir:
:param gen_heartbeat:
:param fp16:
:return:准确率,用时
"""
set_computation_network_trace_level(0)
# Input variables denoting the features and label data
input_var = C.input_variable((num_channels, image_height, image_width),
name='features')
label_var = C.input_variable((num_classes))
with C.default_options(dtype=np.float32):
# create model, and configure learning parameters
model = create_cifar10_model(input_var, 3, num_classes)
# loss and metric
loss = cross_entropy_with_softmax(model, label_var)
error_rate = classification_error(model, label_var)
# shared training parameters
# Set learning parameters
lr_per_sample = []
check_point = [80, 120, 160, 180]
lrs = [3e-2, 3e-3, 3e-4, 3e-4, 5e-5]
for i in range(max_epochs + 1):
if i in range(0, check_point[0]):
lr_per_sample.append(lrs[0])
if i in range(check_point[0], check_point[1]):
lr_per_sample.append(lrs[1])
if i in range(check_point[1], check_point[2]):
lr_per_sample.append(lrs[2])
if i in range(check_point[2], check_point[3]):
lr_per_sample.append(lrs[3])
if i > check_point[3]:
lr_per_sample.append(lrs[4])
lr_schedule = learning_parameter_schedule(lr_per_sample,
minibatch_size=minibatch_size,
epoch_size=epoch_size)
mm_schedule = momentum_schedule(0.9, minibatch_size) #动量
# progress writers
progress_writers = [
ProgressPrinter(tag='Training',
num_epochs=max_epochs,
gen_heartbeat=gen_heartbeat)
]
tensorboard_writer = None
if tensorboard_logdir is not None:
tensorboard_writer = TensorBoardProgressWriter(
freq=10, log_dir=tensorboard_logdir, model=model)
progress_writers.append(tensorboard_writer)
# trainer object
l2_reg_weight = 0.0001
learner = adam(model.parameters,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight)
trainer = Trainer(model, (loss, error_rate), learner, progress_writers)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(model)
print("*********Training Start*********")
start = time.clock()
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(
min(minibatch_size, epoch_size - sample_count),
input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
trainer.summarize_training_progress()
# Log mean of each parameter tensor, so that we can confirm that the parameters change indeed.
if tensorboard_writer:
for parameter in model.parameters:
tensorboard_writer.write_value(parameter.uid + "/mean",
reduce_mean(parameter).eval(),
epoch)
if model_dir:
model.save(
os.path.join(model_dir,
network_name + "_{}.dnn".format(epoch)))
enable_profiler() # begin to collect profiler data after first epoch
# Evaluation parameters
test_epoch_size = 10000
minibatch_size = 32
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
while sample_count < test_epoch_size:
current_minibatch = min(minibatch_size, test_epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch,
input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += data[label_var].num_samples
print("")
trainer.summarize_test_progress()
print("")
elapsed = (time.clock() - start)
return 1 - metric_numer / metric_denom, elapsed
# internally (if needed), so all we need are the logits ce = cross_entropy_with_softmax(logits, labels) pe = classification_error(logits, labels) # training config batch_size = 32 epochs = 15 n_batches = len(Xtrain) // batch_size # do the training # specify the training algorithm trainer = Trainer(logits, (ce, pe), adam(logits.parameters, lr=1e-2, momentum=0.9)) # helper function def get_output(node, X, Y): ret = node.forward(dict(inputs=X, labels=Y)) return list(ret[1].values())[0].mean() costs = [] errors = [] test_costs = [] test_errors = [] for i in range(epochs): cost = 0 err = 0
input_sequences, labels = create_model_placeholders()
if USE_SAVED_MODEL:
model = load_model(MODEL_FILE_PATH)
else:
model = create_model()
z = model(input_sequences)
ce = cross_entropy_with_softmax(z, labels)
errs = classification_error(z, labels)
momentum_schedule = momentum_schedule_per_sample(0.9990913221888589)
clipping_threshold_per_sample = 5.0
gradient_clipping_with_truncation = True
learner = adam(z.parameters, 0.001, momentum_schedule)
trainer = Trainer(z, (ce, errs), [learner])
for e in range(EPOCHS):
arguments = get_random_batch()
if TRAIN:
trainer.train_minibatch(arguments)
if e % LOG_FREQUENCY == 0:
print('Epoch: ' + str(e) +
', Average Classification Error: {:,.0%}'.format(
pd.DataFrame(errs.eval(arguments))[0].mean()))
if TRAIN:
z.save(MODEL_FILE_PATH)
print("Saved model to '%s'" % MODEL_FILE_PATH)
def __init__(self,
state_dim,
action_dim,
gamma=0.99,
learning_rate=1e-4,
momentum=0.95):
self.state_dim = state_dim
self.action_dim = action_dim
self.gamma = gamma
with default_options(activation=relu, init=he_uniform()):
# Convolution filter counts were halved to save on memory, no gpu :(
self.model = Sequential([
Convolution2D((8, 8), 16, strides=4, name='conv1'),
Convolution2D((4, 4), 32, strides=2, name='conv2'),
Convolution2D((3, 3), 32, strides=1, name='conv3'),
Dense(256, init=he_uniform(scale=0.01), name='dense1'),
Dense(action_dim,
activation=None,
init=he_uniform(scale=0.01),
name='actions')
])
self.model.update_signature(Tensor[state_dim])
# Create the target model as a copy of the online model
self.target_model = None
self.update_target()
self.pre_states = input_variable(state_dim, name='pre_states')
self.actions = input_variable(action_dim, name='actions')
self.post_states = input_variable(state_dim, name='post_states')
self.rewards = input_variable((), name='rewards')
self.terminals = input_variable((), name='terminals')
self.is_weights = input_variable((), name='is_weights')
predicted_q = reduce_sum(self.model(self.pre_states) * self.actions,
axis=0)
# DQN - calculate target q values
# post_q = reduce_max(self.target_model(self.post_states), axis=0)
# DDQN - calculate target q values
online_selection = one_hot(
argmax(self.model(self.post_states), axis=0), self.action_dim)
post_q = reduce_sum(self.target_model(self.post_states) *
online_selection,
axis=0)
post_q = (1.0 - self.terminals) * post_q
target_q = stop_gradient(self.rewards + self.gamma * post_q)
# Huber loss
delta = 1.0
self.td_error = minus(predicted_q, target_q, name='td_error')
abs_error = abs(self.td_error)
errors = element_select(less(abs_error, delta),
square(self.td_error) * 0.5,
delta * (abs_error - 0.5 * delta))
loss = errors * self.is_weights
# Adam based SGD
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_scheule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
self._learner = adam(self.model.parameters,
lr_schedule,
m_scheule,
variance_momentum=vm_schedule)
self.writer = TensorBoardProgressWriter(log_dir='metrics',
model=self.model)
self.trainer = Trainer(self.model, (loss, None), [self._learner],
self.writer)
def __init__(self, in_shape, output_shape, device_id=None,
learning_rate=0.00025, momentum=0.9,
minibatch_size=32, update_interval=10000,
n_workers=1, visualizer=None):
"""
Q Neural Network following Mnih and al. implementation and default options.
The network has the following topology:
Convolution(32, (8, 8))
Convolution(64, (4, 4))
Convolution(64, (2, 2))
Dense(512)
:param in_shape: Shape of the observations perceived by the learner (the neural net input)
:param output_shape: Size of the action space (mapped to the number of output neurons)
:param device_id: Use None to let CNTK select the best available device,
-1 for CPU, >= 0 for GPU
(default: None)
:param learning_rate: Learning rate
(default: 0.00025, as per Mnih et al.)
:param momentum: Momentum, provided as momentum value for
averaging gradients without unit gain filter
Note that CNTK does not currently provide an implementation
of Graves' RmsProp with momentum.
It uses AdamSGD optimizer instead.
(default: 0, no momentum with RProp optimizer)
:param minibatch_size: Minibatch size
(default: 32, as per Mnih et al.)
:param n_workers: Number of concurrent worker for distributed training.
(default: 1, not distributed)
:param visualizer: Optional visualizer allowing the model to save summary data
(default: None, no visualization)
Ref: Mnih et al.: "Human-level control through deep reinforcement learning."
Nature 518.7540 (2015): 529-533.
"""
assert learning_rate > 0, 'learning_rate should be > 0'
assert 0. <= momentum < 1, 'momentum should be 0 <= momentum < 1'
QModel.__init__(self, in_shape, output_shape)
CntkModel.__init__(self, device_id, False, n_workers, visualizer)
self._nb_actions = output_shape
self._steps = 0
self._target_update_interval = update_interval
self._target = None
# Input vars
self._environment = input(in_shape, name='env',
dynamic_axes=(Axis.default_batch_axis()))
self._q_targets = input(1, name='q_targets',
dynamic_axes=(Axis.default_batch_axis()))
self._actions = input(output_shape, name='actions',
dynamic_axes=(Axis.default_batch_axis()))
# Define the neural network graph
self._model = self._build_model()(self._environment)
self._target = self._model.clone(
CloneMethod.freeze, {self._environment: self._environment}
)
# Define the learning rate
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
# AdamSGD optimizer
m_schedule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
l_sgd = adam(self._model.parameters, lr_schedule,
momentum=m_schedule,
unit_gain=True,
variance_momentum=vm_schedule)
if self.distributed_training:
raise NotImplementedError('ASGD not implemented yet.')
# _actions is a sparse 1-hot encoding of the actions done by the agent
q_acted = reduce_sum(self._model * self._actions, axis=0)
# Define the trainer with Huber Loss function
criterion = huber_loss(q_acted, self._q_targets, 1.0)
self._learner = l_sgd
self._trainer = Trainer(self._model, (criterion, None), l_sgd)