class DeepQNetwork:
def __init__(self, num_actions, args):
# remember parameters
self.num_actions = num_actions
self.batch_size = args.batch_size
self.discount_rate = args.discount_rate
self.history_length = args.history_length
self.screen_dim = (args.screen_height, args.screen_width)
self.clip_error = args.clip_error
# create Neon backend
self.be = gen_backend(backend=args.backend,
batch_size=args.batch_size,
rng_seed=args.random_seed,
device_id=args.device_id,
default_dtype=np.dtype(args.datatype).type,
stochastic_round=args.stochastic_round)
# prepare tensors once and reuse them
self.input_shape = (self.history_length, ) + self.screen_dim + (
self.batch_size, )
self.tensor = self.be.empty(self.input_shape)
self.tensor.lshape = self.input_shape # needed for convolutional networks
self.targets = self.be.empty((self.num_actions, self.batch_size))
# create model
layers = self.createLayers(num_actions)
self.model = Model(layers=layers)
self.cost = GeneralizedCost(costfunc=SumSquared())
self.model.initialize(self.tensor.shape[:-1], self.cost)
self.optimizer = RMSProp(learning_rate=args.learning_rate,
decay_rate=args.rmsprop_decay_rate,
stochastic_round=args.stochastic_round)
# create target model
self.target_steps = args.target_steps
self.train_iterations = 0
if self.target_steps:
self.target_model = Model(layers=self.createLayers(num_actions))
self.target_model.initialize(self.tensor.shape[:-1])
self.save_weights_path = args.save_weights_path
else:
self.target_model = self.model
self.callback = None
def createLayers(self, num_actions):
# create network
init_norm = Gaussian(loc=0.0, scale=0.01)
layers = []
# The first hidden layer convolves 32 filters of 8x8 with stride 4 with the input image and applies a rectifier nonlinearity.
layers.append(
Conv((8, 8, 32), strides=4, init=init_norm, activation=Rectlin()))
# The second hidden layer convolves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.
layers.append(
Conv((4, 4, 64), strides=2, init=init_norm, activation=Rectlin()))
# This is followed by a third convolutional layer that convolves 64 filters of 3x3 with stride 1 followed by a rectifier.
layers.append(
Conv((3, 3, 64), strides=1, init=init_norm, activation=Rectlin()))
# The final hidden layer is fully-connected and consists of 512 rectifier units.
layers.append(Affine(nout=512, init=init_norm, activation=Rectlin()))
# The output layer is a fully-connected linear layer with a single output for each valid action.
layers.append(Affine(nout=num_actions, init=init_norm))
return layers
def setTensor(self, states):
# change order of axes to match what Neon expects
states = np.transpose(states, axes=(1, 2, 3, 0))
# copy() shouldn't be necessary here, but Neon doesn't work otherwise
self.tensor.set(states.copy())
# normalize network input between 0 and 1
self.be.divide(self.tensor, 255, self.tensor)
def train(self, minibatch, epoch):
# expand components of minibatch
prestates, actions, rewards, poststates, terminals = minibatch
assert len(prestates.shape) == 4
assert len(poststates.shape) == 4
assert len(actions.shape) == 1
assert len(rewards.shape) == 1
assert len(terminals.shape) == 1
assert prestates.shape == poststates.shape
assert prestates.shape[0] == actions.shape[0] == rewards.shape[
0] == poststates.shape[0] == terminals.shape[0]
if self.target_steps and self.train_iterations % self.target_steps == 0:
# HACK: push something through network, so that weights exist
self.model.fprop(self.tensor)
# HACK: serialize network to disk and read it back to clone
filename = os.path.join(self.save_weights_path,
"target_network.pkl")
save_obj(self.model.serialize(keep_states=False), filename)
self.target_model.load_weights(filename)
# feed-forward pass for poststates to get Q-values
self.setTensor(poststates)
postq = self.target_model.fprop(self.tensor, inference=True)
assert postq.shape == (self.num_actions, self.batch_size)
# calculate max Q-value for each poststate
maxpostq = self.be.max(postq, axis=0).asnumpyarray()
assert maxpostq.shape == (1, self.batch_size)
# feed-forward pass for prestates
self.setTensor(prestates)
preq = self.model.fprop(self.tensor, inference=False)
assert preq.shape == (self.num_actions, self.batch_size)
# make copy of prestate Q-values as targets
targets = preq.asnumpyarray()
# update Q-value targets for actions taken
for i, action in enumerate(actions):
if terminals[i]:
targets[action, i] = float(rewards[i])
else:
targets[action, i] = float(
rewards[i]) + self.discount_rate * maxpostq[0, i]
# copy targets to GPU memory
self.targets.set(targets)
# calculate errors
deltas = self.cost.get_errors(preq, self.targets)
assert deltas.shape == (self.num_actions, self.batch_size)
#assert np.count_nonzero(deltas.asnumpyarray()) == 32
# calculate cost, just in case
cost = self.cost.get_cost(preq, self.targets)
assert cost.shape == (1, 1)
# clip errors
if self.clip_error:
self.be.clip(deltas, -self.clip_error, self.clip_error, out=deltas)
# perform back-propagation of gradients
self.model.bprop(deltas)
# perform optimization
self.optimizer.optimize(self.model.layers_to_optimize, epoch)
# increase number of weight updates (needed for target clone interval)
self.train_iterations += 1
# calculate statistics
if self.callback:
self.callback.on_train(cost.asnumpyarray()[0, 0])
def predict(self, states):
# minibatch is full size, because Neon doesn't let change the minibatch size
assert states.shape == ((
self.batch_size,
self.history_length,
) + self.screen_dim)
# calculate Q-values for the states
self.setTensor(states)
qvalues = self.model.fprop(self.tensor, inference=True)
assert qvalues.shape == (self.num_actions, self.batch_size)
if logger.isEnabledFor(logging.DEBUG):
logger.debug("Q-values: " + str(qvalues.asnumpyarray()[:, 0]))
# find the action with highest q-value
actions = self.be.argmax(qvalues, axis=0)
assert actions.shape == (1, self.batch_size)
# take only the first result
return actions.asnumpyarray()[0, 0]
def getMeanQ(self, states):
assert states.shape == ((
self.batch_size,
self.history_length,
) + self.screen_dim)
# calculate Q-values for the states
self.setTensor(states)
qvalues = self.model.fprop(self.tensor, inference=True)
assert qvalues.shape == (self.num_actions, self.batch_size)
# take maximum Q-value for each state
actions = self.be.max(qvalues, axis=0)
assert actions.astensor().shape == (1, self.batch_size)
# calculate mean Q-value of all states
meanq = self.be.mean(actions, axis=1)
assert meanq.astensor().shape == (1, 1)
# return the mean
return meanq.asnumpyarray()[0, 0]
def load_weights(self, load_path):
self.model.load_weights(load_path)
def save_weights(self, save_path):
save_obj(self.model.serialize(keep_states=True), save_path)
class DQNNeon(Learner):
""" This class is an implementation of the DQN network based on Neon.
The modules that interact with the agent, the replay memory and the
statistic calls are implemented here, taking the individual requirements
of the Lasagne framework into account. The code is adapted from:
https://github.com/tambetm/simple_dqn
Attributes:
input_shape (tuple[int]): Dimension of the network input.
dummy_batch (numpy.ndarray): Dummy batche used to calculate Q-values for single states.
batch_norm (bool): Indicates if normalization is wanted for a certain layer (default=False).
be (neon.backends.nervanagpu.NervanaGPU): Describes the backend for the Neon implementation.
input (neon.backends.nervanagpu.GPUTensor): Definition of network input shape.
targets(neon.backends.nervanagpu.GPUTensor): Definition of network output shape.
model (neon.models.model.Model): Generated Neon model.
target_model (neon.models.model.Model): Generated target Neon model.
cost_func (neon.layers.layer.GeneralizedCost): Cost function for model training.
callback (Statistics): Hook for the statistics object to pass train and test information.
Note:
More attributes of this class are defined in the base class Learner.
"""
def __init__(self, env, args, rng, name = "DQNNeon"):
""" Initializes a network based on the Neon framework.
Args:
env (AtariEnv): The envirnoment in which the agent actuates.
args (argparse.Namespace): All settings either with a default value or set via command line arguments.
rng (mtrand.RandomState): initialized Mersenne Twister pseudo-random number generator.
name (str): The name of the network object.
Note:
This function should always call the base class first to initialize
the common values for the networks.
"""
_logger.info("Initializing new object of type " + str(type(self).__name__))
super(DQNNeon, self).__init__(env, args, rng, name)
self.input_shape = (self.sequence_length,) + self.frame_dims + (self.batch_size,)
self.dummy_batch = np.zeros((self.batch_size, self.sequence_length) + self.frame_dims, dtype=np.uint8)
self.batch_norm = args.batch_norm
self.be = gen_backend(
backend = args.backend,
batch_size = args.batch_size,
rng_seed = args.random_seed,
device_id = args.device_id,
datatype = np.dtype(args.datatype).type,
stochastic_round = args.stochastic_round)
# prepare tensors once and reuse them
self.input = self.be.empty(self.input_shape)
self.input.lshape = self.input_shape # HACK: needed for convolutional networks
self.targets = self.be.empty((self.output_shape, self.batch_size))
# create model
layers = self._create_layer()
self.model = Model(layers = layers)
self.cost_func = GeneralizedCost(costfunc = SumSquared())
# Bug fix
for l in self.model.layers.layers:
l.parallelism = 'Disabled'
self.model.initialize(self.input_shape[:-1], self.cost_func)
self._set_optimizer()
if not self.args.load_weights == None:
self.load_weights(self.args.load_weights)
# create target model
if self.target_update_frequency:
layers = self._create_layer()
self.target_model = Model(layers)
# Bug fix
for l in self.target_model.layers.layers:
l.parallelism = 'Disabled'
self.target_model.initialize(self.input_shape[:-1])
else:
self.target_model = self.model
self.callback = None
_logger.debug("%s" % self)
def _create_layer(self):
""" Build a network consistent with the DeepMind Nature paper. """
_logger.debug("Output shape = %d" % self.output_shape)
# create network
init_norm = Gaussian(loc=0.0, scale=0.01)
layers = []
# The first hidden layer convolves 32 filters of 8x8 with stride 4 with the input image and applies a rectifier nonlinearity.
layers.append(
Conv((8, 8, 32),
strides=4,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The second hidden layer convolves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.
layers.append(
Conv((4, 4, 64),
strides=2,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# This is followed by a third convolutional layer that convolves 64 filters of 3x3 with stride 1 followed by a rectifier.
layers.append(
Conv((3, 3, 64),
strides=1,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The final hidden layer is fully-connected and consists of 512 rectifier units.
layers.append(
Affine(
nout=512,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The output layer is a fully-connected linear layer with a single output for each valid action.
layers.append(
Affine(
nout= self.output_shape,
init = init_norm))
return layers
def _set_optimizer(self):
""" Initializes the selected optimization algorithm. """
_logger.debug("Optimizer = %s" % str(self.args.optimizer))
if self.args.optimizer == 'rmsprop':
self.optimizer = RMSProp(
learning_rate = self.args.learning_rate,
decay_rate = self.args.decay_rate,
stochastic_round = self.args.stochastic_round)
elif self.args.optimizer == 'adam':
self.optimizer = Adam(
learning_rate = self.args.learning_rate,
stochastic_round = self.args.stochastic_round)
elif self.args.optimizer == 'adadelta':
self.optimizer = Adadelta(
decay = self.args.decay_rate,
stochastic_round = self.args.stochastic_round)
else:
assert false, "Unknown optimizer"
def _prepare_network_input(self, states):
""" Transforms and normalizes the states from one minibatch.
Args:
states (): a set of states with the size of minibatch
"""
_logger.debug("Normalizing and transforming input")
# change order of axes to match what Neon expects
states = np.transpose(states, axes = (1, 2, 3, 0))
# copy() shouldn't be necessary here, but Neon doesn't work otherwise
self.input.set(states.copy())
# normalize network input between 0 and 1
self.be.divide(self.input, self.grayscales, self.input)
def train(self, minibatch, epoch):
""" Prepare, perform and document a complete train step for one minibatch.
Args:
minibatch (numpy.ndarray): Mini-batch of states, shape=(batch_size,sequence_length,frame_width,frame_height)
epoch (int): Current train epoch
"""
_logger.debug("Complete trainig step for one minibatch")
prestates, actions, rewards, poststates, terminals = minibatch
assert len(prestates.shape) == 4
assert len(poststates.shape) == 4
assert len(actions.shape) == 1
assert len(rewards.shape) == 1
assert len(terminals.shape) == 1
assert prestates.shape == poststates.shape
assert prestates.shape[0] == actions.shape[0] == rewards.shape[0] == poststates.shape[0] == terminals.shape[0]
# feed-forward pass for poststates to get Q-values
self._prepare_network_input(poststates)
postq = self.target_model.fprop(self.input, inference = True)
assert postq.shape == (self.output_shape, self.batch_size)
# calculate max Q-value for each poststate
maxpostq = self.be.max(postq, axis=0).asnumpyarray()
assert maxpostq.shape == (1, self.batch_size)
# average maxpostq for stats
maxpostq_avg = maxpostq.mean()
# feed-forward pass for prestates
self._prepare_network_input(prestates)
preq = self.model.fprop(self.input, inference = False)
assert preq.shape == (self.output_shape, self.batch_size)
# make copy of prestate Q-values as targets
targets = preq.asnumpyarray()
# clip rewards between -1 and 1
rewards = np.clip(rewards, self.min_reward, self.max_reward)
# update Q-value targets for each state only at actions taken
for i, action in enumerate(actions):
if terminals[i]:
targets[action, i] = float(rewards[i])
else:
targets[action, i] = float(rewards[i]) + self.discount_rate * maxpostq[0,i]
# copy targets to GPU memory
self.targets.set(targets)
# calculate errors
errors = self.cost_func.get_errors(preq, self.targets)
assert errors.shape == (self.output_shape, self.batch_size)
# average error where there is a error (should be 1 in every row)
#TODO: errors_avg = np.sum(errors)/np.size(errors[errors>0.])
# clip errors
if self.clip_error:
self.be.clip(errors, -self.clip_error, self.clip_error, out = errors)
# calculate cost, just in case
cost = self.cost_func.get_cost(preq, self.targets)
assert cost.shape == (1,1)
# perform back-propagation of gradients
self.model.bprop(errors)
# perform optimization
self.optimizer.optimize(self.model.layers_to_optimize, epoch)
# increase number of weight updates (needed for target clone interval)
self.update_iterations += 1
if self.target_update_frequency and self.update_iterations % self.target_update_frequency == 0:
self._copy_theta()
if isinstance(cost, np.ndarray):
_logger.info("Network update #%d: Cost = %s, Avg Max Q-value = %s" % (self.update_iterations, str(cost[0][0]), str(maxpostq_avg)))
else:
_logger.info("Network update #%d: Cost = %s, Avg Max Q-value = %s" % (self.update_iterations, str(cost.asnumpyarray()[0][0]), str(maxpostq_avg)))
# update statistics
if self.callback:
if isinstance(cost, np.ndarray):
self.callback.from_learner(cost[0,0], maxpostq_avg)
else:
self.callback.from_learner(cost.asnumpyarray()[0,0], maxpostq_avg)
def get_Q(self, state):
""" Calculates the Q-values for one mini-batch.
Args:
state(numpy.ndarray): Single state, shape=(sequence_length,frame_width,frame_height).
Returns:
q_values (numpy.ndarray): Results for first element of mini-batch from one forward pass through the network, shape=(self.output_shape,)
"""
_logger.debug("State shape = %s" % str(state.shape))
# minibatch is full size, because Neon doesn't let change the minibatch size
# so we need to run 32 forward steps to get the one we actually want
self.dummy_batch[0] = state
states = self.dummy_batch
assert states.shape == ((self.batch_size, self.sequence_length,) + self.frame_dims)
# calculate Q-values for the states
self._prepare_network_input(states)
qvalues = self.model.fprop(self.input, inference = True)
assert qvalues.shape == (self.output_shape, self.batch_size)
_logger.debug("Qvalues: %s" % (str(qvalues.asnumpyarray()[:,0])))
return qvalues.asnumpyarray()[:,0]
def _copy_theta(self):
""" Copies the weights of the current network to the target network. """
_logger.debug("Copying weights")
pdict = self.model.get_description(get_weights=True, keep_states=True)
self.target_model.deserialize(pdict, load_states=True)
def save_weights(self, target_dir, epoch):
""" Saves the current network parameters to disk.
Args:
target_dir (str): Directory where the network parameters are stored for each episode.
epoch (int): Current epoch.
"""
filename = "%s_%s_%s_%d.prm" % (str(self.args.game.lower()), str(self.args.learner_type.lower()), str(self.args.optimizer.lower()), (epoch + 1))
self.model.save_params(os.path.join(target_dir, filename))
def load_weights(self, source_file):
""" Loads the network parameters from a given file.
Args:
source_file (str): Complete path to a file with network parameters.
"""
self.model.load_params(source_file)
class DeepQNetwork:
def __init__(self, num_actions, args):
# create Neon backend
self.be = gen_backend(backend = args.backend,
batch_size = args.batch_size,
rng_seed = args.random_seed,
device_id = args.device_id,
default_dtype = np.dtype(args.datatype).type,
stochastic_round = args.stochastic_round)
# create model
layers = self.createLayers(num_actions)
self.model = Model(layers = layers)
self.cost = GeneralizedCost(costfunc = SumSquared())
self.optimizer = RMSProp(learning_rate = args.learning_rate,
decay_rate = args.rmsprop_decay_rate,
stochastic_round = args.stochastic_round)
# create target model
self.target_steps = args.target_steps
self.train_iterations = 0
if self.target_steps:
self.target_model = Model(layers = self.createLayers(num_actions))
self.save_weights_path = args.save_weights_path
else:
self.target_model = self.model
# remember parameters
self.num_actions = num_actions
self.batch_size = args.batch_size
self.discount_rate = args.discount_rate
self.history_length = args.history_length
self.screen_dim = (args.screen_height, args.screen_width)
self.clip_error = args.clip_error
# prepare tensors once and reuse them
self.input_shape = (self.history_length,) + self.screen_dim + (self.batch_size,)
self.tensor = self.be.empty(self.input_shape)
self.tensor.lshape = self.input_shape # needed for convolutional networks
self.targets = self.be.empty((self.num_actions, self.batch_size))
self.callback = None
def createLayers(self, num_actions):
# create network
init_norm = Gaussian(loc=0.0, scale=0.01)
layers = []
# The first hidden layer convolves 32 filters of 8x8 with stride 4 with the input image and applies a rectifier nonlinearity.
layers.append(Conv((8, 8, 32), strides=4, init=init_norm, activation=Rectlin()))
# The second hidden layer convolves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.
layers.append(Conv((4, 4, 64), strides=2, init=init_norm, activation=Rectlin()))
# This is followed by a third convolutional layer that convolves 64 filters of 3x3 with stride 1 followed by a rectifier.
layers.append(Conv((3, 3, 64), strides=1, init=init_norm, activation=Rectlin()))
# The final hidden layer is fully-connected and consists of 512 rectifier units.
layers.append(Affine(nout=512, init=init_norm, activation=Rectlin()))
# The output layer is a fully-connected linear layer with a single output for each valid action.
layers.append(Affine(nout = num_actions, init = init_norm))
return layers
def setTensor(self, states):
# change order of axes to match what Neon expects
states = np.transpose(states, axes = (1, 2, 3, 0))
# copy() shouldn't be necessary here, but Neon doesn't work otherwise
self.tensor.set(states.copy())
# normalize network input between 0 and 1
self.be.divide(self.tensor, 255, self.tensor)
def train(self, minibatch, epoch):
# expand components of minibatch
prestates, actions, rewards, poststates, terminals = minibatch
assert len(prestates.shape) == 4
assert len(poststates.shape) == 4
assert len(actions.shape) == 1
assert len(rewards.shape) == 1
assert len(terminals.shape) == 1
assert prestates.shape == poststates.shape
assert prestates.shape[0] == actions.shape[0] == rewards.shape[0] == poststates.shape[0] == terminals.shape[0]
if self.target_steps and self.train_iterations % self.target_steps == 0:
# HACK: push something through network, so that weights exist
self.model.fprop(self.tensor)
# HACK: serialize network to disk and read it back to clone
filename = os.path.join(self.save_weights_path, "target_network.pkl")
save_obj(self.model.serialize(keep_states = False), filename)
self.target_model.load_weights(filename)
# feed-forward pass for poststates to get Q-values
self.setTensor(poststates)
postq = self.target_model.fprop(self.tensor, inference = True)
assert postq.shape == (self.num_actions, self.batch_size)
# calculate max Q-value for each poststate
maxpostq = self.be.max(postq, axis=0).asnumpyarray()
assert maxpostq.shape == (1, self.batch_size)
# feed-forward pass for prestates
self.setTensor(prestates)
preq = self.model.fprop(self.tensor, inference = False)
assert preq.shape == (self.num_actions, self.batch_size)
# make copy of prestate Q-values as targets
targets = preq.asnumpyarray()
# update Q-value targets for actions taken
for i, action in enumerate(actions):
if terminals[i]:
targets[action, i] = float(rewards[i])
else:
targets[action, i] = float(rewards[i]) + self.discount_rate * maxpostq[0,i]
# copy targets to GPU memory
self.targets.set(targets)
# calculate errors
deltas = self.cost.get_errors(preq, self.targets)
assert deltas.shape == (self.num_actions, self.batch_size)
#assert np.count_nonzero(deltas.asnumpyarray()) == 32
# calculate cost, just in case
cost = self.cost.get_cost(preq, self.targets)
assert cost.shape == (1,1)
# clip errors
if self.clip_error:
self.be.clip(deltas, -self.clip_error, self.clip_error, out = deltas)
# perform back-propagation of gradients
self.model.bprop(deltas)
# perform optimization
self.optimizer.optimize(self.model.layers_to_optimize, epoch)
# increase number of weight updates (needed for target clone interval)
self.train_iterations += 1
# calculate statistics
if self.callback:
self.callback.on_train(cost.asnumpyarray()[0,0])
def predict(self, states):
# minibatch is full size, because Neon doesn't let change the minibatch size
assert states.shape == ((self.batch_size, self.history_length,) + self.screen_dim)
# calculate Q-values for the states
self.setTensor(states)
qvalues = self.model.fprop(self.tensor, inference = True)
assert qvalues.shape == (self.num_actions, self.batch_size)
if logger.isEnabledFor(logging.DEBUG):
logger.debug("Q-values: " + str(qvalues.asnumpyarray()[:,0]))
# find the action with highest q-value
actions = self.be.argmax(qvalues, axis = 0)
assert actions.shape == (1, self.batch_size)
# take only the first result
return actions.asnumpyarray()[0,0]
def getMeanQ(self, states):
assert states.shape == ((self.batch_size, self.history_length,) + self.screen_dim)
# calculate Q-values for the states
self.setTensor(states)
qvalues = self.model.fprop(self.tensor, inference = True)
assert qvalues.shape == (self.num_actions, self.batch_size)
# take maximum Q-value for each state
actions = self.be.max(qvalues, axis = 0)
assert actions.astensor().shape == (1, self.batch_size)
# calculate mean Q-value of all states
meanq = self.be.mean(actions, axis = 1)
assert meanq.astensor().shape == (1, 1)
# return the mean
return meanq.asnumpyarray()[0,0]
def load_weights(self, load_path):
self.model.load_weights(load_path)
def save_weights(self, save_path):
save_obj(self.model.serialize(keep_states = True), save_path)
class DeepQNetwork:
def __init__(self, num_actions, args):
# remember parameters
self.num_actions = num_actions
self.batch_size = args.batch_size
self.discount_rate = args.discount_rate
self.history_length = args.history_length
self.screen_dim = (args.screen_height, args.screen_width)
self.clip_error = args.clip_error
self.min_reward = args.min_reward
self.max_reward = args.max_reward
self.batch_norm = args.batch_norm
# create Neon backend
self.be = gen_backend(backend = args.backend,
batch_size = args.batch_size,
rng_seed = args.random_seed,
device_id = args.device_id,
datatype = np.dtype(args.datatype).type,
stochastic_round = args.stochastic_round)
# prepare tensors once and reuse them
self.input_shape = (self.history_length,) + self.screen_dim + (self.batch_size,)
self.input = self.be.empty(self.input_shape)
self.input.lshape = self.input_shape # HACK: needed for convolutional networks
self.targets = self.be.empty((self.num_actions, self.batch_size))
# create model
layers = self._createLayers(num_actions)
self.model = Model(layers = layers)
self.cost = GeneralizedCost(costfunc = SumSquared())
# Bug fix
for l in self.model.layers.layers:
l.parallelism = 'Disabled'
self.model.initialize(self.input_shape[:-1], self.cost)
if args.optimizer == 'rmsprop':
self.optimizer = RMSProp(learning_rate = args.learning_rate,
decay_rate = args.decay_rate,
stochastic_round = args.stochastic_round)
elif args.optimizer == 'adam':
self.optimizer = Adam(learning_rate = args.learning_rate,
stochastic_round = args.stochastic_round)
elif args.optimizer == 'adadelta':
self.optimizer = Adadelta(decay = args.decay_rate,
stochastic_round = args.stochastic_round)
else:
assert false, "Unknown optimizer"
# create target model
self.target_steps = args.target_steps
self.train_iterations = 0
if self.target_steps:
self.target_model = Model(layers = self._createLayers(num_actions))
# Bug fix
for l in self.target_model.layers.layers:
l.parallelism = 'Disabled'
self.target_model.initialize(self.input_shape[:-1])
self.save_weights_prefix = args.save_weights_prefix
else:
self.target_model = self.model
self.callback = None
def _createLayers(self, num_actions):
# create network
init_norm = Gaussian(loc=0.0, scale=0.01)
layers = []
# The first hidden layer convolves 32 filters of 8x8 with stride 4 with the input image and applies a rectifier nonlinearity.
layers.append(Conv((8, 8, 32), strides=4, init=init_norm, activation=Rectlin(), batch_norm=self.batch_norm))
# The second hidden layer convolves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.
layers.append(Conv((4, 4, 64), strides=2, init=init_norm, activation=Rectlin(), batch_norm=self.batch_norm))
# This is followed by a third convolutional layer that convolves 64 filters of 3x3 with stride 1 followed by a rectifier.
layers.append(Conv((3, 3, 64), strides=1, init=init_norm, activation=Rectlin(), batch_norm=self.batch_norm))
# The final hidden layer is fully-connected and consists of 512 rectifier units.
layers.append(Affine(nout=512, init=init_norm, activation=Rectlin(), batch_norm=self.batch_norm))
# The output layer is a fully-connected linear layer with a single output for each valid action.
layers.append(Affine(nout=num_actions, init = init_norm))
return layers
def _setInput(self, states):
# change order of axes to match what Neon expects
states = np.transpose(states, axes = (1, 2, 3, 0))
# copy() shouldn't be necessary here, but Neon doesn't work otherwise
self.input.set(states.copy())
# normalize network input between 0 and 1
self.be.divide(self.input, 255, self.input)
def train(self, minibatch, epoch):
# expand components of minibatch
prestates, actions, rewards, poststates, terminals = minibatch
assert len(prestates.shape) == 4
assert len(poststates.shape) == 4
assert len(actions.shape) == 1
assert len(rewards.shape) == 1
assert len(terminals.shape) == 1
assert prestates.shape == poststates.shape
assert prestates.shape[0] == actions.shape[0] == rewards.shape[0] == poststates.shape[0] == terminals.shape[0]
if self.target_steps and self.train_iterations % self.target_steps == 0:
# have to serialize also states for batch normalization to work
pdict = self.model.get_description(get_weights=True, keep_states=True)
self.target_model.deserialize(pdict, load_states=True)
# feed-forward pass for poststates to get Q-values
self._setInput(poststates)
postq = self.target_model.fprop(self.input, inference = True)
assert postq.shape == (self.num_actions, self.batch_size)
# calculate max Q-value for each poststate
maxpostq = self.be.max(postq, axis=0).asnumpyarray()
assert maxpostq.shape == (1, self.batch_size)
# feed-forward pass for prestates
self._setInput(prestates)
preq = self.model.fprop(self.input, inference = False)
assert preq.shape == (self.num_actions, self.batch_size)
# make copy of prestate Q-values as targets
# It seems neccessary for cpu backend.
targets = preq.asnumpyarray().copy()
# clip rewards between -1 and 1
rewards = np.clip(rewards, self.min_reward, self.max_reward)
# update Q-value targets for actions taken
for i, action in enumerate(actions):
if terminals[i]:
targets[action, i] = float(rewards[i])
else:
targets[action, i] = float(rewards[i]) + self.discount_rate * maxpostq[0,i]
# copy targets to GPU memory
self.targets.set(targets)
# calculate errors
deltas = self.cost.get_errors(preq, self.targets)
assert deltas.shape == (self.num_actions, self.batch_size)
#assert np.count_nonzero(deltas.asnumpyarray()) == 32
# calculate cost, just in case
cost = self.cost.get_cost(preq, self.targets)
assert cost.shape == (1,1)
# clip errors
if self.clip_error:
self.be.clip(deltas, -self.clip_error, self.clip_error, out = deltas)
# perform back-propagation of gradients
self.model.bprop(deltas)
# perform optimization
self.optimizer.optimize(self.model.layers_to_optimize, epoch)
# increase number of weight updates (needed for target clone interval)
self.train_iterations += 1
# calculate statistics
if self.callback:
self.callback.on_train(cost[0,0])
def predict(self, states):
# minibatch is full size, because Neon doesn't let change the minibatch size
assert states.shape == ((self.batch_size, self.history_length,) + self.screen_dim)
# calculate Q-values for the states
self._setInput(states)
qvalues = self.model.fprop(self.input, inference = True)
assert qvalues.shape == (self.num_actions, self.batch_size)
if logger.isEnabledFor(logging.DEBUG):
logger.debug("Q-values: " + str(qvalues.asnumpyarray()[:,0]))
# transpose the result, so that batch size is first dimension
return qvalues.T.asnumpyarray()
def load_weights(self, load_path):
self.model.load_params(load_path)
def save_weights(self, save_path):
self.model.save_params(save_path)
class DeepQNetwork:
def __init__(self, state_size, num_steers, num_speeds, args):
# remember parameters
self.state_size = state_size
self.num_steers = num_steers
self.num_speeds = num_speeds
self.num_actions = num_steers + num_speeds
self.num_layers = args.hidden_layers
self.hidden_nodes = args.hidden_nodes
self.batch_size = args.batch_size
self.discount_rate = args.discount_rate
self.clip_error = args.clip_error
# create Neon backend
self.be = gen_backend(backend = args.backend,
batch_size = args.batch_size,
rng_seed = args.random_seed,
device_id = args.device_id,
datatype = np.dtype(args.datatype).type,
stochastic_round = args.stochastic_round)
# prepare tensors once and reuse them
self.input_shape = (self.state_size, self.batch_size)
self.input = self.be.empty(self.input_shape)
self.targets = self.be.empty((self.num_actions, self.batch_size))
# create model
self.model = Model(layers = self._createLayers())
self.cost = GeneralizedCost(costfunc = SumSquared())
self.model.initialize(self.input_shape[:-1], self.cost)
if args.optimizer == 'rmsprop':
self.optimizer = RMSProp(learning_rate = args.learning_rate,
decay_rate = args.decay_rate,
stochastic_round = args.stochastic_round)
elif args.optimizer == 'adam':
self.optimizer = Adam(learning_rate = args.learning_rate,
stochastic_round = args.stochastic_round)
elif args.optimizer == 'adadelta':
self.optimizer = Adadelta(decay = args.decay_rate,
stochastic_round = args.stochastic_round)
else:
assert false, "Unknown optimizer"
# create target model
self.target_steps = args.target_steps
self.train_iterations = 0
if self.target_steps:
self.target_model = Model(layers = self._createLayers())
self.target_model.initialize(self.input_shape[:-1])
self.save_weights_prefix = args.save_weights_prefix
else:
self.target_model = self.model
def _createLayers(self):
# create network
init_norm = Gaussian(loc=0.0, scale=0.01)
layers = []
for i in xrange(self.num_layers):
layers.append(Affine(nout=self.hidden_nodes, init=init_norm, activation=Rectlin()))
layers.append(Affine(nout=self.num_actions, init = init_norm))
return layers
def _setInput(self, states):
# change order of axes to match what Neon expects
states = np.transpose(states)
# copy() shouldn't be necessary here, but Neon doesn't work otherwise
self.input.set(states.copy())
# normalize network input between 0 and 1
#self.be.divide(self.input, 200, self.input)
def train(self, minibatch, epoch = 0):
# expand components of minibatch
prestates, steers, speeds, rewards, poststates, terminals = minibatch
assert len(prestates.shape) == 2
assert len(poststates.shape) == 2
assert len(steers.shape) == 1
assert len(speeds.shape) == 1
assert len(rewards.shape) == 1
assert len(terminals.shape) == 1
assert prestates.shape == poststates.shape
assert prestates.shape[0] == steers.shape[0] == speeds.shape[0] == rewards.shape[0] == poststates.shape[0] == terminals.shape[0]
if self.target_steps and self.train_iterations % self.target_steps == 0:
# HACK: serialize network to disk and read it back to clone
filename = self.save_weights_prefix + "_target.pkl"
save_obj(self.model.serialize(keep_states = False), filename)
self.target_model.load_weights(filename)
# feed-forward pass for poststates to get Q-values
self._setInput(poststates)
postq = self.target_model.fprop(self.input, inference = True)
assert postq.shape == (self.num_actions, self.batch_size)
# calculate max Q-value for each poststate
postq = postq.asnumpyarray()
maxsteerq = np.max(postq[:self.num_steers,:], axis=0)
assert maxsteerq.shape == (self.batch_size,), "size: %s" % str(maxsteerq.shape)
maxspeedq = np.max(postq[-self.num_speeds:,:], axis=0)
assert maxspeedq.shape == (self.batch_size,)
# feed-forward pass for prestates
self._setInput(prestates)
preq = self.model.fprop(self.input, inference = False)
assert preq.shape == (self.num_actions, self.batch_size)
# make copy of prestate Q-values as targets
# HACK: copy() was needed to make it work on CPU
targets = preq.asnumpyarray().copy()
# update Q-value targets for actions taken
for i, (steer, speed) in enumerate(zip(steers, speeds)):
if terminals[i]:
targets[steer, i] = float(rewards[i])
targets[self.num_steers + speed, i] = float(rewards[i])
else:
targets[steer, i] = float(rewards[i]) + self.discount_rate * maxsteerq[i]
targets[self.num_steers + speed, i] = float(rewards[i]) + self.discount_rate * maxspeedq[i]
# copy targets to GPU memory
self.targets.set(targets)
# calculate errors
deltas = self.cost.get_errors(preq, self.targets)
assert deltas.shape == (self.num_actions, self.batch_size)
#assert np.count_nonzero(deltas.asnumpyarray()) == 2 * self.batch_size, str(np.count_nonzero(deltas.asnumpyarray()))
# calculate cost, just in case
cost = self.cost.get_cost(preq, self.targets)
assert cost.shape == (1,1)
#print "cost:", cost.asnumpyarray()
# clip errors
if self.clip_error:
self.be.clip(deltas, -self.clip_error, self.clip_error, out = deltas)
# perform back-propagation of gradients
self.model.bprop(deltas)
# perform optimization
self.optimizer.optimize(self.model.layers_to_optimize, epoch)
'''
if np.any(rewards < 0):
preqq = preq.asnumpyarray().copy()
self._setInput(prestates)
qvalues = self.model.fprop(self.input, inference = True).asnumpyarray().copy()
indexes = rewards < 0
print "indexes:", indexes
print "preq:", preqq[:, indexes].T
print "preq':", qvalues[:, indexes].T
print "diff:", (qvalues[:, indexes]-preqq[:, indexes]).T
print "steers:", steers[indexes]
print "speeds:", speeds[indexes]
print "rewards:", rewards[indexes]
print "terminals:", terminals[indexes]
print "preq[0]:", preqq[:, 0]
print "preq[0]':", qvalues[:, 0]
print "diff:", qvalues[:, 0] - preqq[:, 0]
print "deltas:", deltas.asnumpyarray()[:, indexes].T
raw_input("Press Enter to continue...")
'''
# increase number of weight updates (needed for target clone interval)
self.train_iterations += 1
def predict(self, states):
# minibatch is full size, because Neon doesn't let change the minibatch size
assert states.shape == (self.batch_size, self.state_size)
# calculate Q-values for the states
self._setInput(states)
qvalues = self.model.fprop(self.input, inference = True)
assert qvalues.shape == (self.num_actions, self.batch_size)
if logger.isEnabledFor(logging.DEBUG):
logger.debug("Q-values: " + str(qvalues.asnumpyarray()[:,0]))
# transpose the result, so that batch size is first dimension
return qvalues.T.asnumpyarray()
def load_weights(self, load_path):
self.model.load_weights(load_path)
def save_weights(self, save_path):
save_obj(self.model.serialize(keep_states = True), save_path)
class DeepQNetwork:
def __init__(self, num_actions, args):
# remember parameters
self.num_actions = num_actions
self.batch_size = args.batch_size
self.discount_rate = args.discount_rate
self.history_length = args.history_length
self.screen_dim = (args.screen_height, args.screen_width)
self.clip_error = args.clip_error
self.min_reward = args.min_reward
self.max_reward = args.max_reward
self.batch_norm = args.batch_norm
# create Neon backend
self.be = gen_backend(backend = args.backend,
batch_size = args.batch_size,
rng_seed = args.random_seed,
device_id = args.device_id,
datatype = np.dtype(args.datatype).type,
stochastic_round = args.stochastic_round)
# prepare tensors once and reuse them
self.input_shape = (self.history_length,) + self.screen_dim + (self.batch_size,)
self.input = self.be.empty(self.input_shape)
self.input.lshape = self.input_shape # HACK: needed for convolutional networks
self.targets = self.be.empty((self.num_actions, self.batch_size))
# create model
layers = self._createLayers(num_actions)
self.model = Model(layers = layers)
self.cost = GeneralizedCost(costfunc = SumSquared())
# Bug fix
for l in self.model.layers.layers:
l.parallelism = 'Disabled'
self.model.initialize(self.input_shape[:-1], self.cost)
if args.optimizer == 'rmsprop':
self.optimizer = RMSProp(learning_rate = args.learning_rate,
decay_rate = args.decay_rate,
stochastic_round = args.stochastic_round)
elif args.optimizer == 'adam':
self.optimizer = Adam(learning_rate = args.learning_rate,
stochastic_round = args.stochastic_round)
elif args.optimizer == 'adadelta':
self.optimizer = Adadelta(decay = args.decay_rate,
stochastic_round = args.stochastic_round)
else:
assert false, "Unknown optimizer"
# create target model
self.train_iterations = 0
if args.target_steps:
self.target_model = Model(layers = self._createLayers(num_actions))
# Bug fix
for l in self.target_model.layers.layers:
l.parallelism = 'Disabled'
self.target_model.initialize(self.input_shape[:-1])
self.save_weights_prefix = args.save_weights_prefix
else:
self.target_model = self.model
self.callback = None
def _createLayers(self, num_actions):
# create network
init_xavier_conv = Xavier(local=True)
init_xavier_affine = Xavier(local=False)
layers = []
# The first hidden layer convolves 32 filters of 8x8 with stride 4 with the input image and applies a rectifier nonlinearity.
layers.append(Conv((8, 8, 32), strides=4, init=init_xavier_conv, activation=Rectlin(), batch_norm=self.batch_norm))
# The second hidden layer convolves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.
layers.append(Conv((4, 4, 64), strides=2, init=init_xavier_conv, activation=Rectlin(), batch_norm=self.batch_norm))
# This is followed by a third convolutional layer that convolves 64 filters of 3x3 with stride 1 followed by a rectifier.
layers.append(Conv((3, 3, 64), strides=1, init=init_xavier_conv, activation=Rectlin(), batch_norm=self.batch_norm))
# The final hidden layer is fully-connected and consists of 512 rectifier units.
layers.append(Affine(nout=512, init=init_xavier_affine, activation=Rectlin(), batch_norm=self.batch_norm))
# The output layer is a fully-connected linear layer with a single output for each valid action.
layers.append(Affine(nout=num_actions, init = init_xavier_affine))
return layers
def _setInput(self, states):
# change order of axes to match what Neon expects
states = np.transpose(states, axes = (1, 2, 3, 0))
# copy() shouldn't be necessary here, but Neon doesn't work otherwise
self.input.set(states.copy())
# normalize network input between 0 and 1
self.be.divide(self.input, 255, self.input)
def update_target_network(self):
# have to serialize also states for batch normalization to work
pdict = self.model.get_description(get_weights=True, keep_states=True)
self.target_model.deserialize(pdict, load_states=True)
def train(self, minibatch, epoch):
# expand components of minibatch
prestates, actions, rewards, poststates, terminals = minibatch
assert len(prestates.shape) == 4
assert len(poststates.shape) == 4
assert len(actions.shape) == 1
assert len(rewards.shape) == 1
assert len(terminals.shape) == 1
assert prestates.shape == poststates.shape
assert prestates.shape[0] == actions.shape[0] == rewards.shape[0] == poststates.shape[0] == terminals.shape[0]
# feed-forward pass for poststates to get Q-values
self._setInput(poststates)
postq = self.target_model.fprop(self.input, inference = True)
assert postq.shape == (self.num_actions, self.batch_size)
# calculate max Q-value for each poststate
maxpostq = self.be.max(postq, axis=0).asnumpyarray()
assert maxpostq.shape == (1, self.batch_size)
# feed-forward pass for prestates
self._setInput(prestates)
preq = self.model.fprop(self.input, inference = False)
assert preq.shape == (self.num_actions, self.batch_size)
# make copy of prestate Q-values as targets
targets = preq.asnumpyarray().copy()
# clip rewards between -1 and 1
rewards = np.clip(rewards, self.min_reward, self.max_reward)
# update Q-value targets for actions taken
for i, action in enumerate(actions):
if terminals[i]:
targets[action, i] = float(rewards[i])
else:
targets[action, i] = float(rewards[i]) + self.discount_rate * maxpostq[0,i]
# copy targets to GPU memory
self.targets.set(targets)
# calculate errors
deltas = self.cost.get_errors(preq, self.targets)
assert deltas.shape == (self.num_actions, self.batch_size)
#assert np.count_nonzero(deltas.asnumpyarray()) == 32
# calculate cost, just in case
cost = self.cost.get_cost(preq, self.targets)
assert cost.shape == (1,1)
# clip errors
if self.clip_error:
self.be.clip(deltas, -self.clip_error, self.clip_error, out = deltas)
# perform back-propagation of gradients
self.model.bprop(deltas)
# perform optimization
self.optimizer.optimize(self.model.layers_to_optimize, epoch)
# increase number of weight updates (needed for stats callback)
self.train_iterations += 1
# calculate statistics
if self.callback:
self.callback.on_train(cost[0,0])
def predict(self, states):
# minibatch is full size, because Neon doesn't let change the minibatch size
assert states.shape == ((self.batch_size, self.history_length,) + self.screen_dim)
# calculate Q-values for the states
self._setInput(states)
qvalues = self.model.fprop(self.input, inference = True)
assert qvalues.shape == (self.num_actions, self.batch_size)
if logger.isEnabledFor(logging.DEBUG):
logger.debug("Q-values: " + str(qvalues.asnumpyarray()[:,0]))
# transpose the result, so that batch size is first dimension
return qvalues.T.asnumpyarray()
def load_weights(self, load_path):
self.model.load_params(load_path)
def save_weights(self, save_path):
self.model.save_params(save_path)
class DeepQNetwork:
def __init__(self, state_size, num_actions, args):
# remember parameters
self.state_size = state_size
self.num_actions = num_actions
self.batch_size = args.batch_size
self.discount_rate = args.discount_rate
self.clip_error = args.clip_error
self.action_count = np.zeros(21)
# create Neon backend
self.be = gen_backend(backend = args.backend,
batch_size = args.batch_size,
rng_seed = args.random_seed,
device_id = args.device_id,
datatype = np.dtype(args.datatype).type,
stochastic_round = args.stochastic_round)
# prepare tensors once and reuse them
self.input_shape = (self.state_size, self.batch_size)
self.input = self.be.empty(self.input_shape)
self.targets = self.be.empty((self.num_actions, self.batch_size))
# create model
layers = self._createLayers(num_actions)
self.model = Model(layers = layers)
self.cost = GeneralizedCost(costfunc = SumSquared())
self.model.initialize(self.input_shape[:-1], self.cost)
if args.optimizer == 'rmsprop':
self.optimizer = RMSProp(learning_rate = args.learning_rate,
decay_rate = args.decay_rate,
stochastic_round = args.stochastic_round)
elif args.optimizer == 'adam':
self.optimizer = Adam(learning_rate = args.learning_rate,
stochastic_round = args.stochastic_round)
elif args.optimizer == 'adadelta':
self.optimizer = Adadelta(decay = args.decay_rate,
stochastic_round = args.stochastic_round)
else:
assert False, "Unknown optimizer"
# create target model
self.target_steps = args.target_steps
self.train_iterations = 0
if self.target_steps:
self.target_model = Model(layers = self._createLayers(num_actions))
self.target_model.initialize(self.input_shape[:-1])
self.save_weights_prefix = args.save_weights_prefix
else:
self.target_model = self.model
def _createLayers(self, num_actions):
# create network
init_norm = Gaussian(loc=0.0, scale=0.01)
layers = []
# The final hidden layer is fully-connected and consists of 512 rectifier units.
layers.append(Affine(nout=64, init=init_norm, bias=init_norm, activation=Rectlin()))
# The output layer is a fully-connected linear layer with a single output for each valid action.
layers.append(Affine(nout=num_actions, init=init_norm, bias=init_norm))
return layers
def _setInput(self, states):
# change order of axes to match what Neon expects
states = np.transpose(states)
# copy() shouldn't be necessary here, but Neon doesn't work otherwise
self.input.set(states.copy())
# normalize network input between 0 and 1
# self.be.divide(self.input, 255, self.input)
def train(self, minibatch, epoch):
# expand components of minibatch
prestates, actions, speed_actions, rewards, poststates, terminals = minibatch
assert len(prestates.shape) == 2
assert len(poststates.shape) == 2
assert len(actions.shape) == 1
assert len(rewards.shape) == 1
assert len(terminals.shape) == 1
assert prestates.shape == poststates.shape
assert prestates.shape[0] == actions.shape[0] == rewards.shape[0] == poststates.shape[0] == terminals.shape[0]
#print "WE ARE ACTUALLY TRAINING IN HERE"
if self.target_steps and self.train_iterations % self.target_steps == 0:
# HACK: serialize network to disk and read it back to clone
filename = self.save_weights_prefix + "_target.pkl"
save_obj(self.model.serialize(keep_states = False), filename)
self.target_model.load_weights(filename)
# feed-forward pass for poststates to get Q-values
self._setInput(poststates)
postq = self.target_model.fprop(self.input, inference = True)
assert postq.shape == (self.num_actions, self.batch_size)
# calculate max Q-value for each poststate
postq = postq.asnumpyarray()
maxpostq = np.max(postq, axis=0)
#print maxpostq.shape
assert maxpostq.shape == (self.batch_size,)
# feed-forward pass for prestates
self._setInput(prestates)
preq = self.model.fprop(self.input, inference = False)
assert preq.shape == (self.num_actions, self.batch_size)
# make copy of prestate Q-values as targets
targets = preq.asnumpyarray().copy()
# update Q-value targets for actions taken
for i, action in enumerate(actions):
self.action_count[action] += 1
if terminals[i]:
targets[action, i] = float(rewards[i])
if rewards[i] == -1000:
print "######################### action ", action, "should never be sampled again"
print "sampled_terminal"
else:
targets[action, i] = float(rewards[i]) + self.discount_rate * maxpostq[i]
#targets[i,action] = float(rewards[i]) + self.discount_rate * maxpostq[i]
#print "action count", self.action_count
# copy targets to GPU memory
self.targets.set(targets)
# calculate errors
deltas = self.cost.get_errors(preq, self.targets)
assert deltas.shape == (self.num_actions, self.batch_size)
#assert np.count_nonzero(deltas.asnumpyarray()) == 32
print "nonzero deltas", np.count_nonzero(deltas.asnumpyarray())
# calculate cost, just in case
cost = self.cost.get_cost(preq, self.targets)
assert cost.shape == (1,1)
print "cost:", cost.asnumpyarray()
# clip errors
#if self.clip_error:
# self.be.clip(deltas, -self.clip_error, self.clip_error, out = deltas)
# perform back-propagation of gradients
self.model.bprop(deltas)
# perform optimization
self.optimizer.optimize(self.model.layers_to_optimize, epoch)
# increase number of weight updates (needed for target clone interval)
self.train_iterations += 1
def predict(self, states):
# minibatch is full size, because Neon doesn't let change the minibatch size
assert states.shape == (self.batch_size, self.state_size)
# calculate Q-values for the states
self._setInput(states)
qvalues = self.model.fprop(self.input, inference = True)
assert qvalues.shape == (self.num_actions, self.batch_size)
if logger.isEnabledFor(logging.DEBUG):
logger.debug("Q-values: " + str(qvalues.asnumpyarray()[:,0]))
# transpose the result, so that batch size is first dimension
return qvalues.T.asnumpyarray()
def load_weights(self, load_path):
self.model.load_weights(load_path)
def save_weights(self, save_path):
save_obj(self.model.serialize(keep_states = True), save_path)
class ModelRunnerNeon():
def __init__(self, args, max_action_no, batch_dimension):
self.args = args
self.train_batch_size = args.train_batch_size
self.discount_factor = args.discount_factor
self.use_gpu_replay_mem = args.use_gpu_replay_mem
self.be = gen_backend(backend='gpu', batch_size=self.train_batch_size)
self.input_shape = (batch_dimension[1], batch_dimension[2],
batch_dimension[3], batch_dimension[0])
self.input = self.be.empty(self.input_shape)
self.input.lshape = self.input_shape # HACK: needed for convolutional networks
self.targets = self.be.empty((max_action_no, self.train_batch_size))
if self.use_gpu_replay_mem:
self.history_buffer = self.be.zeros(batch_dimension,
dtype=np.uint8)
self.input_uint8 = self.be.empty(self.input_shape, dtype=np.uint8)
else:
self.history_buffer = np.zeros(batch_dimension, dtype=np.float32)
self.train_net = Model(self.create_layers(max_action_no))
self.cost = GeneralizedCost(costfunc=SumSquared())
# Bug fix
for l in self.train_net.layers.layers:
l.parallelism = 'Disabled'
self.train_net.initialize(self.input_shape[:-1], self.cost)
self.target_net = Model(self.create_layers(max_action_no))
# Bug fix
for l in self.target_net.layers.layers:
l.parallelism = 'Disabled'
self.target_net.initialize(self.input_shape[:-1])
if self.args.optimizer == 'Adam': # Adam
self.optimizer = Adam(beta_1=args.rms_decay,
beta_2=args.rms_decay,
learning_rate=args.learning_rate)
else: # Neon RMSProp
self.optimizer = RMSProp(decay_rate=args.rms_decay,
learning_rate=args.learning_rate)
self.max_action_no = max_action_no
self.running = True
def get_initializer(self, input_size):
dnnInit = self.args.dnn_initializer
if dnnInit == 'xavier':
initializer = Xavier()
elif dnnInit == 'fan_in':
std_dev = 1.0 / math.sqrt(input_size)
initializer = Uniform(low=-std_dev, high=std_dev)
else:
initializer = Gaussian(0, 0.01)
return initializer
def create_layers(self, max_action_no):
layers = []
initializer = self.get_initializer(input_size=4 * 8 * 8)
layers.append(
Conv(fshape=(8, 8, 32),
strides=4,
init=initializer,
bias=initializer,
activation=Rectlin()))
initializer = self.get_initializer(input_size=32 * 4 * 4)
layers.append(
Conv(fshape=(4, 4, 64),
strides=2,
init=initializer,
bias=initializer,
activation=Rectlin()))
initializer = self.get_initializer(input_size=64 * 3 * 3)
layers.append(
Conv(fshape=(3, 3, 64),
strides=1,
init=initializer,
bias=initializer,
activation=Rectlin()))
initializer = self.get_initializer(input_size=7 * 7 * 64)
layers.append(
Affine(nout=512,
init=initializer,
bias=initializer,
activation=Rectlin()))
initializer = self.get_initializer(input_size=512)
layers.append(
Affine(nout=max_action_no, init=initializer, bias=initializer))
return layers
def clip_reward(self, reward):
if reward > self.args.clip_reward_high:
return self.args.clip_reward_high
elif reward < self.args.clip_reward_low:
return self.args.clip_reward_low
else:
return reward
def set_input(self, data):
if self.use_gpu_replay_mem:
self.be.copy_transpose(data, self.input_uint8, axes=(1, 2, 3, 0))
self.input[:] = self.input_uint8 / 255
else:
self.input.set(data.transpose(1, 2, 3, 0).copy())
self.be.divide(self.input, 255, self.input)
def predict(self, history_buffer):
self.set_input(history_buffer)
output = self.train_net.fprop(self.input, inference=True)
return output.T.asnumpyarray()[0]
def print_weights(self):
pass
def train(self, minibatch, replay_memory, learning_rate, debug):
if self.args.prioritized_replay == True:
prestates, actions, rewards, poststates, terminals, replay_indexes, heap_indexes, weights = minibatch
else:
prestates, actions, rewards, poststates, terminals = minibatch
# Get Q*(s, a) with targetNet
self.set_input(poststates)
post_qvalue = self.target_net.fprop(self.input,
inference=True).T.asnumpyarray()
if self.args.double_dqn == True:
# Get Q*(s, a) with trainNet
post_qvalue2 = self.train_net.fprop(
self.input, inference=True).T.asnumpyarray()
# Get Q(s, a) with trainNet
self.set_input(prestates)
pre_qvalue = self.train_net.fprop(self.input, inference=False)
label = pre_qvalue.asnumpyarray().copy()
for i in range(0, self.train_batch_size):
if self.args.clip_reward:
reward = self.clip_reward(rewards[i])
else:
reward = rewards[i]
if terminals[i]:
label[actions[i], i] = reward
else:
if self.args.double_dqn == True:
max_index = np.argmax(post_qvalue2[i])
label[actions[i],
i] = reward + self.discount_factor * post_qvalue[i][
max_index]
else:
label[actions[i],
i] = reward + self.discount_factor * np.max(
post_qvalue[i])
# copy targets to GPU memory
self.targets.set(label)
delta = self.cost.get_errors(pre_qvalue, self.targets)
if self.args.prioritized_replay == True:
delta_value = delta.asnumpyarray()
for i in range(self.train_batch_size):
if debug:
print 'weight[%s]: %.5f, delta: %.5f, newDelta: %.5f' % (
i, weights[i], delta_value[actions[i], i],
weights[i] * delta_value[actions[i], i])
replay_memory.update_td(heap_indexes[i],
abs(delta_value[actions[i], i]))
delta_value[actions[i],
i] = weights[i] * delta_value[actions[i], i]
delta.set(delta_value.copy())
if self.args.clip_loss:
self.be.clip(delta, -1.0, 1.0, out=delta)
self.train_net.bprop(delta)
self.optimizer.optimize(self.train_net.layers_to_optimize, epoch=0)
def update_model(self):
# have to serialize also states for batch normalization to work
pdict = self.train_net.get_description(get_weights=True,
keep_states=True)
self.target_net.deserialize(pdict, load_states=True)
#print ('Updated target model')
def finish_train(self):
self.running = False
def load(self, file_name):
self.train_net.load_params(file_name)
self.update_model()
def save(self, file_name):
self.train_net.save_params(file_name)
class DQNNeon(Learner):
""" This class is an implementation of the DQN network based on Neon.
The modules that interact with the agent, the replay memory and the
statistic calls are implemented here, taking the individual requirements
of the Lasagne framework into account. The code is adapted from:
https://github.com/tambetm/simple_dqn
Attributes:
input_shape (tuple[int]): Dimension of the network input.
dummy_batch (numpy.ndarray): Dummy batche used to calculate Q-values for single states.
batch_norm (bool): Indicates if normalization is wanted for a certain layer (default=False).
be (neon.backends.nervanagpu.NervanaGPU): Describes the backend for the Neon implementation.
input (neon.backends.nervanagpu.GPUTensor): Definition of network input shape.
targets(neon.backends.nervanagpu.GPUTensor): Definition of network output shape.
model (neon.models.model.Model): Generated Neon model.
target_model (neon.models.model.Model): Generated target Neon model.
cost_func (neon.layers.layer.GeneralizedCost): Cost function for model training.
callback (Statistics): Hook for the statistics object to pass train and test information.
Note:
More attributes of this class are defined in the base class Learner.
"""
def __init__(self, env, args, rng, name = "DQNNeon"):
""" Initializes a network based on the Neon framework.
Args:
env (AtariEnv): The envirnoment in which the agent actuates.
args (argparse.Namespace): All settings either with a default value or set via command line arguments.
rng (mtrand.RandomState): initialized Mersenne Twister pseudo-random number generator.
name (str): The name of the network object.
Note:
This function should always call the base class first to initialize
the common values for the networks.
"""
_logger.info("Initializing new object of type " + str(type(self).__name__))
super(DQNNeon, self).__init__(env, args, rng, name)
self.input_shape = (self.sequence_length,) + self.frame_dims + (self.batch_size,)
self.dummy_batch = np.zeros((self.batch_size, self.sequence_length) + self.frame_dims, dtype=np.uint8)
self.batch_norm = args.batch_norm
self.be = gen_backend(
backend = args.backend,
batch_size = args.batch_size,
rng_seed = args.random_seed,
device_id = args.device_id,
datatype = np.dtype(args.datatype).type,
stochastic_round = args.stochastic_round)
# prepare tensors once and reuse them
self.input = self.be.empty(self.input_shape)
self.input.lshape = self.input_shape # HACK: needed for convolutional networks
self.targets = self.be.empty((self.output_shape, self.batch_size))
# create model
layers = self._create_layer()
self.model = Model(layers = layers)
self.cost_func = GeneralizedCost(costfunc = SumSquared())
# Bug fix
for l in self.model.layers.layers:
l.parallelism = 'Disabled'
self.model.initialize(self.input_shape[:-1], self.cost_func)
self._set_optimizer()
if not self.args.load_weights == None:
self.load_weights(self.args.load_weights)
# create target model
if self.target_update_frequency:
layers = self._create_layer()
self.target_model = Model(layers)
# Bug fix
for l in self.target_model.layers.layers:
l.parallelism = 'Disabled'
self.target_model.initialize(self.input_shape[:-1])
else:
self.target_model = self.model
self.callback = None
_logger.debug("%s" % self)
def _create_layer(self):
""" Build a network consistent with the DeepMind Nature paper. """
_logger.debug("Output shape = %d" % self.output_shape)
# create network
init_norm = Gaussian(loc=0.0, scale=0.01)
layers = []
# The first hidden layer convolves 32 filters of 8x8 with stride 4 with the input image and applies a rectifier nonlinearity.
layers.append(
Conv((8, 8, 32),
strides=4,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The second hidden layer convolves 64 filters of 4x4 with stride 2, again followed by a rectifier nonlinearity.
layers.append(
Conv((4, 4, 64),
strides=2,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# This is followed by a third convolutional layer that convolves 64 filters of 3x3 with stride 1 followed by a rectifier.
layers.append(
Conv((3, 3, 64),
strides=1,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The final hidden layer is fully-connected and consists of 512 rectifier units.
layers.append(
Affine(
nout=512,
init=init_norm,
activation=Rectlin(),
batch_norm=self.batch_norm))
# The output layer is a fully-connected linear layer with a single output for each valid action.
layers.append(
Affine(
nout= self.output_shape,
init = init_norm))
return layers
def _set_optimizer(self):
""" Initializes the selected optimization algorithm. """
_logger.debug("Optimizer = %s" % str(self.args.optimizer))
if self.args.optimizer == 'rmsprop':
self.optimizer = RMSProp(
learning_rate = self.args.learning_rate,
decay_rate = self.args.decay_rate,
stochastic_round = self.args.stochastic_round)
elif self.args.optimizer == 'adam':
self.optimizer = Adam(
learning_rate = self.args.learning_rate,
stochastic_round = self.args.stochastic_round)
elif self.args.optimizer == 'adadelta':
self.optimizer = Adadelta(
decay = self.args.decay_rate,
stochastic_round = self.args.stochastic_round)
else:
assert false, "Unknown optimizer"
def _prepare_network_input(self, states):
""" Transforms and normalizes the states from one minibatch.
Args:
states (): a set of states with the size of minibatch
"""
_logger.debug("Normalizing and transforming input")
# change order of axes to match what Neon expects
states = np.transpose(states, axes = (1, 2, 3, 0))
# copy() shouldn't be necessary here, but Neon doesn't work otherwise
self.input.set(states.copy())
# normalize network input between 0 and 1
self.be.divide(self.input, self.grayscales, self.input)
def train(self, minibatch, epoch):
""" Prepare, perform and document a complete train step for one minibatch.
Args:
minibatch (numpy.ndarray): Mini-batch of states, shape=(batch_size,sequence_length,frame_width,frame_height)
epoch (int): Current train epoch
"""
_logger.debug("Complete trainig step for one minibatch")
prestates, actions, rewards, poststates, terminals = minibatch
assert len(prestates.shape) == 4
assert len(poststates.shape) == 4
assert len(actions.shape) == 1
assert len(rewards.shape) == 1
assert len(terminals.shape) == 1
assert prestates.shape == poststates.shape
assert prestates.shape[0] == actions.shape[0] == rewards.shape[0] == poststates.shape[0] == terminals.shape[0]
# feed-forward pass for poststates to get Q-values
self._prepare_network_input(poststates)
postq = self.target_model.fprop(self.input, inference = True)
assert postq.shape == (self.output_shape, self.batch_size)
# calculate max Q-value for each poststate
maxpostq = self.be.max(postq, axis=0).asnumpyarray()
assert maxpostq.shape == (1, self.batch_size)
# average maxpostq for stats
maxpostq_avg = maxpostq.mean()
# feed-forward pass for prestates
self._prepare_network_input(prestates)
preq = self.model.fprop(self.input, inference = False)
assert preq.shape == (self.output_shape, self.batch_size)
# make copy of prestate Q-values as targets
targets = preq.asnumpyarray()
# clip rewards between -1 and 1
rewards = np.clip(rewards, self.min_reward, self.max_reward)
# update Q-value targets for each state only at actions taken
for i, action in enumerate(actions):
if terminals[i]:
targets[action, i] = float(rewards[i])
else:
targets[action, i] = float(rewards[i]) + self.discount_rate * maxpostq[0,i]
# copy targets to GPU memory
self.targets.set(targets)
# calculate errors
errors = self.cost_func.get_errors(preq, self.targets)
assert errors.shape == (self.output_shape, self.batch_size)
# average error where there is a error (should be 1 in every row)
#TODO: errors_avg = np.sum(errors)/np.size(errors[errors>0.])
# clip errors
if self.clip_error:
self.be.clip(errors, -self.clip_error, self.clip_error, out = errors)
# calculate cost, just in case
cost = self.cost_func.get_cost(preq, self.targets)
assert cost.shape == (1,1)
# perform back-propagation of gradients
self.model.bprop(errors)
# perform optimization
self.optimizer.optimize(self.model.layers_to_optimize, epoch)
# increase number of weight updates (needed for target clone interval)
self.update_iterations += 1
if self.target_update_frequency and self.update_iterations % self.target_update_frequency == 0:
self._copy_theta()
_logger.info("Network update #%d: Cost = %s, Avg Max Q-value = %s" % (self.update_iterations, str(cost.asnumpyarray()[0][0]), str(maxpostq_avg)))
# update statistics
if self.callback:
self.callback.from_learner(cost.asnumpyarray()[0,0], maxpostq_avg)
def get_Q(self, state):
""" Calculates the Q-values for one mini-batch.
Args:
state(numpy.ndarray): Single state, shape=(sequence_length,frame_width,frame_height).
Returns:
q_values (numpy.ndarray): Results for first element of mini-batch from one forward pass through the network, shape=(self.output_shape,)
"""
_logger.debug("State shape = %s" % str(state.shape))
# minibatch is full size, because Neon doesn't let change the minibatch size
# so we need to run 32 forward steps to get the one we actually want
self.dummy_batch[0] = state
states = self.dummy_batch
assert states.shape == ((self.batch_size, self.sequence_length,) + self.frame_dims)
# calculate Q-values for the states
self._prepare_network_input(states)
qvalues = self.model.fprop(self.input, inference = True)
assert qvalues.shape == (self.output_shape, self.batch_size)
_logger.debug("Qvalues: %s" % (str(qvalues.asnumpyarray()[:,0])))
return qvalues.asnumpyarray()[:,0]
def _copy_theta(self):
""" Copies the weights of the current network to the target network. """
_logger.debug("Copying weights")
pdict = self.model.get_description(get_weights=True, keep_states=True)
self.target_model.deserialize(pdict, load_states=True)
def save_weights(self, target_dir, epoch):
""" Saves the current network parameters to disk.
Args:
target_dir (str): Directory where the network parameters are stored for each episode.
epoch (int): Current epoch.
"""
filename = "%s_%s_%s_%d.prm" % (str(self.args.game.lower()), str(self.args.net_type.lower()), str(self.args.optimizer.lower()), (epoch + 1))
self.model.save_params(os.path.join(target_dir, filename))
def load_weights(self, source_file):
""" Loads the network parameters from a given file.
Args:
source_file (str): Complete path to a file with network parameters.
"""
self.model.load_params(source_file)
class DeepQNetwork:
def __init__(self,
num_actions,
batch_size=32,
discount_rate=0.99,
history_length=4,
cols=64,
rows=64,
clip_error=1,
min_reward=-1,
max_reward=1,
batch_norm=False):
self.num_actions = num_actions
self.batch_size = batch_size
self.discount_rate = discount_rate
self.history_length = history_length
self.board_dim = (cols, rows)
self.clip_error = clip_error
self.min_reward = min_reward
self.max_reward = max_reward
self.batch_norm = batch_norm
self.be = gen_backend(backend='gpu',
batch_size=self.batch_size,
datatype=np.dtype('float32').type)
self.input_shape = (self.history_length, ) + self.board_dim + (
self.batch_size, )
self.input = self.be.empty(self.input_shape)
self.input.lshape = self.input_shape # hack from simple_dqn "needed for convolutional networks"
self.targets = self.be.empty((self.num_actions, self.batch_size))
layers = self._createLayers(self.num_actions)
self.model = Model(layers=layers)
self.cost = GeneralizedCost(costfunc=SumSquared())
# for l in self.model.layers.layers:
# l.parallelism = 'Disabled'
self.model.initialize(self.input_shape[:-1], cost=self.cost)
self.optimizer = RMSProp(learning_rate=0.002,
decay_rate=0.95,
stochastic_round=True)
self.train_iterations = 0
self.target_model = Model(layers=self._createLayers(num_actions))
# for l in self.target_model.layers.layers:
# l.parallelism = 'Disabled'
self.target_model.initialize(self.input_shape[:-1])
self.callback = None
def _createLayers(self, num_actions):
init_xavier_conv = Xavier(local=True)
init_xavier_affine = Xavier(local=False)
layers = []
layers.append(
Conv((8, 8, 32),
strides=4,
init=init_xavier_conv,
activation=Rectlin(),
batch_norm=self.batch_norm))
layers.append(
Conv((4, 4, 64),
strides=2,
init=init_xavier_conv,
activation=Rectlin(),
batch_norm=self.batch_norm))
layers.append(
Conv((2, 2, 128),
strides=1,
init=init_xavier_conv,
activation=Rectlin(),
batch_norm=self.batch_norm))
layers.append(
Affine(nout=256,
init=init_xavier_affine,
activation=Rectlin(),
batch_norm=self.batch_norm))
layers.append(Affine(nout=num_actions, init=init_xavier_affine))
return layers
def _setInput(self, states):
states = np.transpose(states, axes=(1, 2, 3, 0))
self.input.set(states.copy())
self.be.add(self.input, 1, self.input)
self.be.divide(self.input, 2, self.input)
def update_target_network(self):
pdict = self.model.get_description(get_weights=True, keep_states=True)
self.target_model.deserialize(pdict, load_states=True)
def train(self, minibatch, epoch):
prestates, actions, rewards, poststates, terminals = minibatch
self._setInput(poststates)
postq = self.target_model.fprop(self.input, inference=True)
assert postq.shape == (self.num_actions, self.batch_size)
maxpostq = self.be.max(postq, axis=0).asnumpyarray()
assert maxpostq.shape == (1, self.batch_size)
self._setInput(prestates)
preq = self.model.fprop(self.input, inference=False)
assert preq.shape == (self.num_actions, self.batch_size)
targets = preq.asnumpyarray().copy()
rewards = np.clip(rewards, -1, 1)
for i, action in enumerate(actions):
if terminals[i]:
targets[action, i] = float(rewards[i])
else:
targets[action, i] = float(
rewards[i]) + self.discount_rate * maxpostq[0, i]
self.targets.set(targets)
deltas = self.cost.get_errors(preq, self.targets)
assert deltas.shape == (self.num_actions, self.batch_size)
cost = self.cost.get_cost(preq, self.targets)
assert cost.shape == (1, 1)
if self.clip_error:
self.be.clip(deltas, -self.clip_error, self.clip_error, out=deltas)
self.model.bprop(deltas)
self.optimizer.optimize(self.model.layers_to_optimize, epoch)
self.train_iterations += 1
self.callback.on_train(cost[0, 0])
def predict(self, states):
assert states.shape == ((
self.batch_size,
self.history_length,
) + self.board_dim)
self._setInput(states)
qvalues = self.model.fprop(self.input, inference=True)
assert qvalues.shape == (self.num_actions, self.batch_size)
return qvalues.T.asnumpyarray()
def load_weights(self, load_path):
self.model.load_params(load_path)
def save_weights(self, save_path):
self.model.save_params(save_path)
class ModelRunnerNeon():
def __init__(self, args, max_action_no, batch_dimension):
self.args = args
self.train_batch_size = args.train_batch_size
self.discount_factor = args.discount_factor
self.use_gpu_replay_mem = args.use_gpu_replay_mem
self.be = gen_backend(backend='gpu',
batch_size=self.train_batch_size)
self.input_shape = (batch_dimension[1], batch_dimension[2], batch_dimension[3], batch_dimension[0])
self.input = self.be.empty(self.input_shape)
self.input.lshape = self.input_shape # HACK: needed for convolutional networks
self.targets = self.be.empty((max_action_no, self.train_batch_size))
if self.use_gpu_replay_mem:
self.history_buffer = self.be.zeros(batch_dimension, dtype=np.uint8)
self.input_uint8 = self.be.empty(self.input_shape, dtype=np.uint8)
else:
self.history_buffer = np.zeros(batch_dimension, dtype=np.float32)
self.train_net = Model(self.create_layers(max_action_no))
self.cost = GeneralizedCost(costfunc=SumSquared())
# Bug fix
for l in self.train_net.layers.layers:
l.parallelism = 'Disabled'
self.train_net.initialize(self.input_shape[:-1], self.cost)
self.target_net = Model(self.create_layers(max_action_no))
# Bug fix
for l in self.target_net.layers.layers:
l.parallelism = 'Disabled'
self.target_net.initialize(self.input_shape[:-1])
if self.args.optimizer == 'Adam': # Adam
self.optimizer = Adam(beta_1=args.rms_decay,
beta_2=args.rms_decay,
learning_rate=args.learning_rate)
else: # Neon RMSProp
self.optimizer = RMSProp(decay_rate=args.rms_decay,
learning_rate=args.learning_rate)
self.max_action_no = max_action_no
self.running = True
def get_initializer(self, input_size):
dnnInit = self.args.dnn_initializer
if dnnInit == 'xavier':
initializer = Xavier()
elif dnnInit == 'fan_in':
std_dev = 1.0 / math.sqrt(input_size)
initializer = Uniform(low=-std_dev, high=std_dev)
else:
initializer = Gaussian(0, 0.01)
return initializer
def create_layers(self, max_action_no):
layers = []
initializer = self.get_initializer(input_size = 4 * 8 * 8)
layers.append(Conv(fshape=(8, 8, 32), strides=4, init=initializer, bias=initializer, activation=Rectlin()))
initializer = self.get_initializer(input_size = 32 * 4 * 4)
layers.append(Conv(fshape=(4, 4, 64), strides=2, init=initializer, bias=initializer, activation=Rectlin()))
initializer = self.get_initializer(input_size = 64 * 3 * 3)
layers.append(Conv(fshape=(3, 3, 64), strides=1, init=initializer, bias=initializer, activation=Rectlin()))
initializer = self.get_initializer(input_size = 7 * 7 * 64)
layers.append(Affine(nout=512, init=initializer, bias=initializer, activation=Rectlin()))
initializer = self.get_initializer(input_size = 512)
layers.append(Affine(nout=max_action_no, init=initializer, bias=initializer))
return layers
def clip_reward(self, reward):
if reward > self.args.clip_reward_high:
return self.args.clip_reward_high
elif reward < self.args.clip_reward_low:
return self.args.clip_reward_low
else:
return reward
def set_input(self, data):
if self.use_gpu_replay_mem:
self.be.copy_transpose(data, self.input_uint8, axes=(1, 2, 3, 0))
self.input[:] = self.input_uint8 / 255
else:
self.input.set(data.transpose(1, 2, 3, 0).copy())
self.be.divide(self.input, 255, self.input)
def predict(self, history_buffer):
self.set_input(history_buffer)
output = self.train_net.fprop(self.input, inference=True)
return output.T.asnumpyarray()[0]
def print_weights(self):
pass
def train(self, minibatch, replay_memory, learning_rate, debug):
if self.args.prioritized_replay == True:
prestates, actions, rewards, poststates, terminals, replay_indexes, heap_indexes, weights = minibatch
else:
prestates, actions, rewards, poststates, terminals = minibatch
# Get Q*(s, a) with targetNet
self.set_input(poststates)
post_qvalue = self.target_net.fprop(self.input, inference=True).T.asnumpyarray()
if self.args.double_dqn == True:
# Get Q*(s, a) with trainNet
post_qvalue2 = self.train_net.fprop(self.input, inference=True).T.asnumpyarray()
# Get Q(s, a) with trainNet
self.set_input(prestates)
pre_qvalue = self.train_net.fprop(self.input, inference=False)
label = pre_qvalue.asnumpyarray().copy()
for i in range(0, self.train_batch_size):
if self.args.clip_reward:
reward = self.clip_reward(rewards[i])
else:
reward = rewards[i]
if terminals[i]:
label[actions[i], i] = reward
else:
if self.args.double_dqn == True:
max_index = np.argmax(post_qvalue2[i])
label[actions[i], i] = reward + self.discount_factor* post_qvalue[i][max_index]
else:
label[actions[i], i] = reward + self.discount_factor* np.max(post_qvalue[i])
# copy targets to GPU memory
self.targets.set(label)
delta = self.cost.get_errors(pre_qvalue, self.targets)
if self.args.prioritized_replay == True:
delta_value = delta.asnumpyarray()
for i in range(self.train_batch_size):
if debug:
print 'weight[%s]: %.5f, delta: %.5f, newDelta: %.5f' % (i, weights[i], delta_value[actions[i], i], weights[i] * delta_value[actions[i], i])
replay_memory.update_td(heap_indexes[i], abs(delta_value[actions[i], i]))
delta_value[actions[i], i] = weights[i] * delta_value[actions[i], i]
delta.set(delta_value.copy())
if self.args.clip_loss:
self.be.clip(delta, -1.0, 1.0, out = delta)
self.train_net.bprop(delta)
self.optimizer.optimize(self.train_net.layers_to_optimize, epoch=0)
def update_model(self):
# have to serialize also states for batch normalization to work
pdict = self.train_net.get_description(get_weights=True, keep_states=True)
self.target_net.deserialize(pdict, load_states=True)
#print ('Updated target model')
def finish_train(self):
self.running = False
def load(self, file_name):
self.train_net.load_params(file_name)
self.update_model()
def save(self, file_name):
self.train_net.save_params(file_name)