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 __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 test_sum_squared_derivative(backend_default):
outputs = np.array([0.5, 1.0, 0.0, 0.0001]).reshape((4, 1))
targets = np.array(([0.5, 0.0, 1.0, 0.2])).reshape((4, 1))
expected_result = (outputs - targets) / outputs.shape[1]
compare_tensors(SumSquared(),
outputs,
targets,
expected_result,
deriv=True,
tol=1e-8)
def __init__(self, rounding, callback_args, epochs):
# setup weight initialization function
self.init = Gaussian(loc=0.0, scale=0.01)
# setup optimizer
self.optimizer = GradientDescentMomentum(learning_rate=0.1, momentum_coef=0.9,
stochastic_round=rounding)
# setup cost function as CrossEntropy
self.cost = GeneralizedCost(costfunc=SumSquared())
self.epochs = epochs
self.model = None
self.callback_args = callback_args
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 __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 train_regressor(orig_wordvecs, w2v_W, w2v_vocab):
"""
Return regressor to map word2vec to RNN word space
Function modified from:
https://github.com/ryankiros/skip-thoughts/blob/master/training/tools.py
"""
# Gather all words from word2vec that appear in wordvecs
d = defaultdict(lambda: 0)
for w in w2v_vocab.keys():
d[w] = 1
shared = OrderedDict()
count = 0
for w in list(orig_wordvecs.keys())[:-2]:
if d[w] > 0:
shared[w] = count
count += 1
# Get the vectors for all words in 'shared'
w2v = np.zeros((len(shared), 300), dtype='float32')
sg = np.zeros((len(shared), 620), dtype='float32')
for w in shared.keys():
w2v[shared[w]] = w2v_W[w2v_vocab[w]]
sg[shared[w]] = orig_wordvecs[w]
train_set = ArrayIterator(X=w2v, y=sg, make_onehot=False)
layers = [
Linear(nout=620, init=Gaussian(loc=0.0, scale=0.1)),
Bias(init=Constant(0.0))
]
clf = Model(layers=layers)
# regression model is trained using default global batch size
cost = GeneralizedCost(costfunc=SumSquared())
opt = GradientDescentMomentum(0.1, 0.9, gradient_clip_value=5.0)
callbacks = Callbacks(clf)
clf.fit(train_set,
num_epochs=20,
optimizer=opt,
cost=cost,
callbacks=callbacks)
return clf
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 __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 test_sum_squared_limits(backend_default):
outputs = np.array([0.5, 1.0, 0.0, 0.0001]).reshape((4, 1))
targets = np.array(([0.5, 0.0, 1.0, 0.2])).reshape((4, 1))
expected_result = np.sum((outputs - targets) ** 2, axis=0, keepdims=True) / 2.
compare_tensors(SumSquared(), outputs, targets, expected_result, tol=1e-7)
# setup model layers
layers = [
Conv((5, 5, 16), init=init_norm, activation=Rectlin()),
Pooling(2),
Conv((5, 5, 32), init=init_norm, activation=Rectlin()),
Pooling(2),
Conv((3, 3, 32), init=init_norm, activation=Rectlin()),
Pooling(2),
Affine(nout=100, init=init_norm, activation=Rectlin()),
Linear(nout=4, init=init_norm)
]
model = Model(layers=layers)
# cost = GeneralizedCost(costfunc=CrossEntropyBinary())
cost = GeneralizedCost(costfunc=SumSquared())
# fit and validate
optimizer = RMSProp()
# configure callbacks
callbacks = Callbacks(model, eval_set=eval_set, eval_freq=1)
model.fit(train_set,
cost=cost,
optimizer=optimizer,
num_epochs=10,
callbacks=callbacks)
y_test = model.get_outputs(test_set)
def test_sum_squared(backend):
outputs = np.array([0.5, 0.9, 0.1, 0.0001]).reshape((4, 1))
targets = np.array([0.5, 0.99, 0.01, 0.2]).reshape((4, 1))
expected_result = np.sum(
(outputs - targets)**2, axis=0, keepdims=True) / 2.
compare_tensors(SumSquared(), outputs, targets, expected_result, tol=1e-8)
