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.q, self.loss,
[sgd(self.q.parameters, lr=5e-4)])
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,
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.frame_stacker = FrameStacker(stack_size=num_frames_to_stack,
frame_shape=observation_space_shape)
self.num_actions = num_actions
self._build_network(pretrained_policy)
self.trainer = Trainer(self.q, self.loss,
[sgd(self.q.parameters, lr=0.000001)])
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)
classificationError = classification_error(outputLayer, labelsShape)
input_map = {
featuresShape: reader.streams.features,
labelsShape: reader.streams.labels
}
numOfEpochs = 10
printer = [ProgressPrinter(
tag = 'Training',
num_epochs = numOfEpochs)]
learningRate = learning_rate_schedule([0.1, 0.01, 0.001], UnitType.sample, 700)
trainer = Trainer(outputLayer,(crossEntropy, classificationError), [adadelta(outputLayer.parameters, learningRate)], printer)
minibatchSize = 50
numberOfSamples = 2208
numberOfSweepsForTraining = 10
trainingSession = training_session(
trainer=trainer,
mb_source=reader,
mb_size=minibatchSize,
model_inputs_to_streams=input_map,
max_samples=numberOfSamples * numberOfSweepsForTraining,
progress_frequency=numberOfSamples
)
trainingSession.train()