class MostCommonWordSense:
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 build(self):
# setup model layers
layers = [Affine(nout=100, init=self.init, bias=self.init, activation=Rectlin()),
Affine(nout=2, init=self.init, bias=self.init, activation=Softmax())]
# initialize model object
self.model = Model(layers=layers)
def fit(self, valid_set, train_set):
# configure callbacks
callbacks = Callbacks(self.model, eval_set=valid_set, **self.callback_args)
self.model.fit(train_set, optimizer=self.optimizer, num_epochs=self.epochs,
cost=self.cost, callbacks=callbacks)
def save(self, save_path):
self.model.save_params(save_path)
def load(self, model_path):
self.model = Model(model_path)
def eval(self, valid_set):
eval_rate = self.model.eval(valid_set, metric=Misclassification())
return eval_rate
def get_outputs(self, valid_set):
return self.model.get_outputs(valid_set)
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)
momentum_coef=0.9,
wdecay=0.0005)
opt_b = GradientDescentMomentum(learning_rate=0.01, momentum_coef=0.9)
opt = MultiOptimizer({'default': opt_w, 'Bias': opt_b}, name='multiopt')
# configure callbacks
callbacks = Callbacks(model,
eval_set=valid_set,
metric=Misclassification(),
**args.callback_args)
callbacks.add_callback(
TrainByStageCallback(model, valid_set, Misclassification(),
max_patience=5))
num_prune = [5, 10, 25, 10]
callbacks.add_callback(
FuzzyPruneCallback(num_states=100, num_prune=num_prune, model=model))
print('Original Accuracy = %.2f%%' %
(100. - model.eval(valid_set, metric=Misclassification()) * 100))
logger.info('Training ...')
model.fit(train_set,
optimizer=opt,
num_epochs=250,
cost=cost,
callbacks=callbacks)
print('Accuracy = %.2f%%' %
(100. - model.eval(valid_set, metric=Misclassification()) * 100))
model.save_params('./models/mnist/mnistfp.pkl')
def test_model_serialize(backend_default, data):
(X_train, y_train), (X_test, y_test), nclass = load_mnist(path=data)
train_set = ArrayIterator([X_train, X_train],
y_train,
nclass=nclass,
lshape=(1, 28, 28))
init_norm = Gaussian(loc=0.0, scale=0.01)
# initialize model
path1 = Sequential([
Conv((5, 5, 16),
init=init_norm,
bias=Constant(0),
activation=Rectlin()),
Pooling(2),
Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin())
])
path2 = Sequential([
Affine(nout=100,
init=init_norm,
bias=Constant(0),
activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin())
])
layers = [
MergeMultistream(layers=[path1, path2], merge="stack"),
Affine(nout=20, init=init_norm, batch_norm=True, activation=Rectlin()),
Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))
]
tmp_save = 'test_model_serialize_tmp_save.pickle'
mlp = Model(layers=layers)
mlp.optimizer = GradientDescentMomentum(learning_rate=0.1,
momentum_coef=0.9)
mlp.cost = GeneralizedCost(costfunc=CrossEntropyBinary())
mlp.initialize(train_set, cost=mlp.cost)
n_test = 3
num_epochs = 3
# Train model for num_epochs and n_test batches
for epoch in range(num_epochs):
for i, (x, t) in enumerate(train_set):
x = mlp.fprop(x)
delta = mlp.cost.get_errors(x, t)
mlp.bprop(delta)
mlp.optimizer.optimize(mlp.layers_to_optimize, epoch=epoch)
if i > n_test:
break
# Get expected outputs of n_test batches and states of all layers
outputs_exp = []
pdicts_exp = [l.get_params_serialize() for l in mlp.layers_to_optimize]
for i, (x, t) in enumerate(train_set):
outputs_exp.append(mlp.fprop(x, inference=True))
if i > n_test:
break
# Serialize model
mlp.save_params(tmp_save, keep_states=True)
# Load model
mlp = Model(tmp_save)
mlp.initialize(train_set)
outputs = []
pdicts = [l.get_params_serialize() for l in mlp.layers_to_optimize]
for i, (x, t) in enumerate(train_set):
outputs.append(mlp.fprop(x, inference=True))
if i > n_test:
break
# Check outputs, states, and params are the same
for output, output_exp in zip(outputs, outputs_exp):
assert np.allclose(output.get(), output_exp.get())
for pd, pd_exp in zip(pdicts, pdicts_exp):
for s, s_e in zip(pd['states'], pd_exp['states']):
if isinstance(s, list): # this is the batch norm case
for _s, _s_e in zip(s, s_e):
assert np.allclose(_s, _s_e)
else:
assert np.allclose(s, s_e)
for p, p_e in zip(pd['params'], pd_exp['params']):
assert type(p) == type(p_e)
if isinstance(p, list): # this is the batch norm case
for _p, _p_e in zip(p, p_e):
assert np.allclose(_p, _p_e)
elif isinstance(p, np.ndarray):
assert np.allclose(p, p_e)
else:
assert p == p_e
os.remove(tmp_save)
class NpSemanticSegClassifier:
"""
NP Semantic Segmentation classifier model (based on Neon framework).
Args:
num_epochs(int): number of epochs to train the model
**callback_args (dict): callback args keyword arguments to init a Callback for the model
cost: the model's cost function. Default is 'neon.transforms.CrossEntropyBinary' cost
optimizer (:obj:`neon.optimizers`): the model's optimizer. Default is
'neon.optimizers.GradientDescentMomentum(0.07, momentum_coef=0.9)'
"""
def __init__(self, num_epochs, callback_args,
optimizer=GradientDescentMomentum(0.07, momentum_coef=0.9)):
"""
Args:
num_epochs(int): number of epochs to train the model
**callback_args (dict): callback args keyword arguments to init Callback for the model
cost: the model's cost function. Default is 'neon.transforms.CrossEntropyBinary' cost
optimizer (:obj:`neon.optimizers`): the model's optimizer. Default is
`neon.optimizers.GradientDescentMomentum(0.07, momentum_coef=0.9)`
"""
self.model = None
self.cost = GeneralizedCost(costfunc=CrossEntropyBinary())
self.optimizer = optimizer
self.epochs = num_epochs
self.callback_args = callback_args
def build(self):
"""
Build the model's layers
"""
first_layer_dens = 64
second_layer_dens = 64
output_layer_dens = 2
# setup weight initialization function
init_norm = Gaussian(scale=0.01)
# setup model layers
layers = [Affine(nout=first_layer_dens, init=init_norm,
activation=Rectlin()),
Affine(nout=second_layer_dens, init=init_norm,
activation=Rectlin()),
Affine(nout=output_layer_dens, init=init_norm,
activation=Logistic(shortcut=True))]
# initialize model object
self.model = Model(layers=layers)
def fit(self, test_set, train_set):
"""
Train and fit the model on the datasets
Args:
test_set (:obj:`neon.data.ArrayIterators`): The test set
train_set (:obj:`neon.data.ArrayIterators`): The train set
args: callback_args and epochs from ArgParser input
"""
# configure callbacks
callbacks = Callbacks(self.model, eval_set=test_set, **self.callback_args)
self.model.fit(train_set, optimizer=self.optimizer, num_epochs=self.epochs, cost=self.cost,
callbacks=callbacks)
def save(self, model_path):
"""
Save the model's prm file in model_path location
Args:
model_path(str): local path for saving the model
"""
self.model.save_params(model_path)
def load(self, model_path):
"""
Load pre-trained model's .prm file to NpSemanticSegClassifier object
Args:
model_path(str): local path for loading the model
"""
self.model = Model(model_path)
def eval(self, test_set):
"""
Evaluate the model's test_set on error_rate, test_accuracy_rate and precision_recall_rate
Args:
test_set (ArrayIterator): The test set
Returns:
tuple(int): error_rate, test_accuracy_rate and precision_recall_rate
"""
error_rate = self.model.eval(test_set, metric=Misclassification())
test_accuracy_rate = self.model.eval(test_set, metric=Accuracy())
precision_recall_rate = self.model.eval(test_set, metric=PrecisionRecall(2))
return error_rate, test_accuracy_rate, precision_recall_rate
def get_outputs(self, test_set):
"""
Classify the dataset on the model
Args:
test_set (:obj:`neon.data.ArrayIterators`): The test set
Returns:
list(float): model's predictions
"""
return self.model.get_outputs(test_set)
Affine(nout=4, init=init_uni, activation=Softmax())] cost = GeneralizedCost(costfunc=CrossEntropyMulti()) # Create model mlp = Model(layers=layers) callbacks = Callbacks(mlp, eval_set=test) # Track cost function # Train model mlp.fit(train, optimizer=opt_gdm, num_epochs=num_epochs, cost=cost, callbacks=callbacks) # Check performance print 'Misclassification error = %.1f%%' % (mlp.eval(test, metric=Misclassification())*100) # Save trained model mlp.save_params(param_file_name) # Sanity check from PIL import Image import numpy as np from neon.data.dataiterator import ArrayIterator W = img_size H = img_size L = W*H*3 size = H, W x_new = np.zeros((128, L), dtype=np.float32) def load_sample(test_file_name):
def ShiftAdaMax_with_Scale(LR=1):
return ShiftAdaMax(learning_rate=LR_start * LR,
schedule=ShiftSchedule(2, shift_size=1))
optimizer = MultiOptimizer({
'default': ShiftAdaMax_with_Scale(),
'BinaryLinear_0': ShiftAdaMax_with_Scale(57.038),
'BinaryLinear_1': ShiftAdaMax_with_Scale(73.9008),
'BinaryLinear_2': ShiftAdaMax_with_Scale(73.9008),
'BinaryLinear_3': ShiftAdaMax_with_Scale(52.3195)
})
# initialize model object
bnn = Model(layers=layers)
# configure callbacks
callbacks = Callbacks(bnn, eval_set=val_set, **args.callback_args)
# run fit
bnn.fit(train_set,
optimizer=optimizer,
num_epochs=args.epochs,
cost=cost,
callbacks=callbacks)
print('Misclassification error = %.1f%%' %
(bnn.eval(val_set, metric=Misclassification()) * 100))
bnn.save_params("bin_model/final_model.prm")
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)
lunaModel.load_params(args.model_file)
# configure callbacks
if args.callback_args['eval_freq'] is None:
args.callback_args['eval_freq'] = 1
# configure callbacks
callbacks = Callbacks(lunaModel, eval_set=valid_set, **args.callback_args)
# add a callback that saves the best model state
callbacks.add_save_best_state_callback(
'LUNA16_VGG_model_no_batch_sigmoid_pretrained.prm')
if args.deconv:
callbacks.add_deconv_callback(train_set, valid_set)
lunaModel.fit(train_set,
optimizer=opt,
num_epochs=num_epochs,
cost=cost,
callbacks=callbacks)
lunaModel.save_params('LUNA16_VGG_model_no_batch_sigmoid_pretrained.prm')
neon_logger.display(
'Calculating metrics on the test set. This could take a while...')
neon_logger.display('Misclassification error (test) = {:.2f}%'.format(
lunaModel.eval(test_set, metric=Misclassification())[0] * 100))
neon_logger.display('Precision/recall (test) = {}'.format(
lunaModel.eval(test_set, metric=PrecisionRecall(num_classes=2))))
if args.model_file:
import os
assert os.path.exists(args.model_file), '%s not found' % args.model_file
lunaModel.load_params(args.model_file)
# configure callbacks
#callbacks = Callbacks(lunaModel, eval_set=valid_set, **args.callback_args)
callbacks = Callbacks(lunaModel,
eval_set=valid_set,
metric=Misclassification(),
**args.callback_args)
if args.deconv:
callbacks.add_deconv_callback(train_set, valid_set)
lunaModel.fit(train_set,
optimizer=opt_gdm,
num_epochs=num_epochs,
cost=cost,
callbacks=callbacks)
lunaModel.save_params('LUNA16_simpleCNN_model.prm')
neon_logger.display(
'Finished training. Calculating error on the validation set...')
neon_logger.display('Misclassification error = {:.2f}%'.format(
lunaModel.eval(valid_set, metric=Misclassification())[0] * 100))
neon_logger.display('Precision/recall = {}'.format(
lunaModel.eval(valid_set, metric=PrecisionRecall(num_classes=2))))
if args.callback_args['eval_freq'] is None:
args.callback_args['eval_freq'] = 1
# configure callbacks
callbacks = Callbacks(mlp, eval_set=valid_set, **args.callback_args)
#callbacks.add_early_stop_callback(stop_func)
#callbacks.add_save_best_state_callback(os.path.join(args.data_dir, "early_stop-best_state.pkl"))
callbacks.add_early_stop_callback(stop_func)
callbacks.add_save_best_state_callback(os.path.join(args.data_dir, "early_stop-best_state.pkl"))
# run fit
mlp.fit(train_set, optimizer=optimizer, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
#evaluate model
print('Evaluation Error = %.4f'%(mlp.eval(valid_set, metric=SmoothL1Metric())))
print('Test set error = %.4f'%(mlp.eval(test_set, metric=SmoothL1Metric())))
# Saving the model
print 'Saving model parameters!'
mlp.save_params("jobwait_model.prm")
# Reloading saved model
# This should go in run.py
mlp=Model("jobwait_model.prm")
print('Test set error = %.4f'%(mlp.eval(test_set, metric=SmoothL1Metric())))
# save the preprocessor vectors:
np.savez("jobwait_preproc", mean=std_scale.mean_, std=std_scale.scale_)
Y = data.train_label-1
X_test = data.test_data
Y_test = data.test_label-1
train_set = ArrayIterator(X=X, y=Y, nclass=11, lshape=(1,200,200))
test_set = ArrayIterator(X_test, None, nclass=11, lshape=(1,200,200))
init_uni = Uniform(low=-0.1, high=0.1)
layers = [Conv(fshape=(4,4,16), init=init_uni, activation=Rectlin()),
Pooling(fshape=2, strides=2),
Conv(fshape=(4,4,32), init=init_uni, activation=Rectlin()),
Pooling(fshape=2, strides=2),
Conv(fshape=(4,4,32), init=init_uni, activation=Rectlin()),
Pooling(fshape=2, strides=2),
Affine(nout=500, init=init_uni, activation=Rectlin()),
Affine(nout=11, init=init_uni, activation=Softmax())]
model = Model(layers)
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
optimizer = GradientDescentMomentum(learning_rate=0.005,
momentum_coef=0.9)
callbacks = Callbacks(model, train_set)
model.fit(dataset=train_set, cost=cost, optimizer=optimizer, num_epochs=40, callbacks=callbacks)
model.save_params('model.pkl')
# out = model.get_outputs(test_set)
# row = len(Y_test)
# result = np.zeros((row,1))
# i=0
# while i<row:
# result[i] = out[i].argmax()
# i=i+1
# np.save('result.npy', result)
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
from neon.optimizers import GradientDescentMomentum, RMSProp
optimizer = GradientDescentMomentum(learning_rate=0.005, momentum_coef=0.9)
# Set up callbacks. By default sets up a progress bar
from neon.callbacks.callbacks import Callbacks
callbacks = Callbacks(model, train_set)
model.fit(dataset=train_set,
cost=cost,
optimizer=optimizer,
num_epochs=num_epochs,
callbacks=callbacks)
model.save_params("cifar10_model.prm")
# Evaluate performance
from neon.transforms import Misclassification
error_pct = 100 * model.eval(test_set, metric=Misclassification())
print 'Misclassification error = %.1f%%' % error_pct
# Sanity check 1
# an image of a frog from wikipedia
# img_source = "https://upload.wikimedia.org/wikipedia/commons/thumb/5/55/Atelopus_zeteki1.jpg/440px-Atelopus_zeteki1.jpg"
# import urllib
# urllib.urlretrieve(img_source, filename="image.jpg")
from PIL import Image
import numpy as np
class NpSemanticSegClassifier:
"""
NP Semantic Segmentation classifier model (based on Neon framework).
Args:
num_epochs(int): number of epochs to train the model
**callback_args (dict): callback args keyword arguments to init a Callback for the model
cost: the model's cost function. Default is 'neon.transforms.CrossEntropyBinary' cost
optimizer (:obj:`neon.optimizers`): the model's optimizer. Default is
'neon.optimizers.GradientDescentMomentum(0.07, momentum_coef=0.9)'
"""
def __init__(self,
num_epochs,
callback_args,
optimizer=GradientDescentMomentum(0.07, momentum_coef=0.9)):
"""
Args:
num_epochs(int): number of epochs to train the model
**callback_args (dict): callback args keyword arguments to init Callback for the model
cost: the model's cost function. Default is 'neon.transforms.CrossEntropyBinary' cost
optimizer (:obj:`neon.optimizers`): the model's optimizer. Default is
`neon.optimizers.GradientDescentMomentum(0.07, momentum_coef=0.9)`
"""
self.model = None
self.cost = GeneralizedCost(costfunc=CrossEntropyBinary())
self.optimizer = optimizer
self.epochs = num_epochs
self.callback_args = callback_args
def build(self):
"""
Build the model's layers
"""
first_layer_dens = 64
second_layer_dens = 64
output_layer_dens = 2
# setup weight initialization function
init_norm = Gaussian(scale=0.01)
# setup model layers
layers = [
Affine(nout=first_layer_dens, init=init_norm,
activation=Rectlin()),
Affine(nout=second_layer_dens,
init=init_norm,
activation=Rectlin()),
Affine(nout=output_layer_dens,
init=init_norm,
activation=Logistic(shortcut=True))
]
# initialize model object
self.model = Model(layers=layers)
def fit(self, test_set, train_set):
"""
Train and fit the model on the datasets
Args:
test_set (:obj:`neon.data.ArrayIterators`): The test set
train_set (:obj:`neon.data.ArrayIterators`): The train set
args: callback_args and epochs from ArgParser input
"""
# configure callbacks
callbacks = Callbacks(self.model,
eval_set=test_set,
**self.callback_args)
self.model.fit(train_set,
optimizer=self.optimizer,
num_epochs=self.epochs,
cost=self.cost,
callbacks=callbacks)
def save(self, model_path):
"""
Save the model's prm file in model_path location
Args:
model_path(str): local path for saving the model
"""
self.model.save_params(model_path)
def load(self, model_path):
"""
Load pre-trained model's .prm file to NpSemanticSegClassifier object
Args:
model_path(str): local path for loading the model
"""
self.model = Model(model_path)
def eval(self, test_set):
"""
Evaluate the model's test_set on error_rate, test_accuracy_rate and precision_recall_rate
Args:
test_set (ArrayIterator): The test set
Returns:
tuple(int): error_rate, test_accuracy_rate and precision_recall_rate
"""
error_rate = self.model.eval(test_set, metric=Misclassification())
test_accuracy_rate = self.model.eval(test_set, metric=Accuracy())
precision_recall_rate = self.model.eval(test_set,
metric=PrecisionRecall(2))
return error_rate, test_accuracy_rate, precision_recall_rate
def get_outputs(self, test_set):
"""
Classify the dataset on the model
Args:
test_set (:obj:`neon.data.ArrayIterators`): The test set
Returns:
list(float): model's predictions
"""
return self.model.get_outputs(test_set)
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 WordseqRegressor():
def __init__(self, pickle_model="", datadir=None):
self.maxlen = 100
self.n_words = 100000
parser = NeonArgparser(__doc__)
self.args = parser.parse_args()
self.args.batch_size = self.batch_size = 2048 #
self.args.deterministic = None
self.args.rng_seed = 0
print extract_valid_args(self.args, gen_backend)
self.be = gen_backend(**extract_valid_args(self.args, gen_backend))
embedding_dim = 100
init_emb = Uniform(-0.1 / embedding_dim, 0.1 / embedding_dim)
init_glorot = GlorotUniform()
self.layers = [
LookupTable(vocab_size=self.n_words,
embedding_dim=embedding_dim,
init=init_emb,
pad_idx=0,
update=True,
name="LookupTable"),
Dropout(keep=0.5),
BiLSTM(100,
init=init_glorot,
activation=Tanh(),
gate_activation=Logistic(),
reset_cells=True,
split_inputs=False,
name="BiLSTM"),
RecurrentMean(),
Affine(1,
init_glorot,
bias=init_glorot,
activation=Identity(),
name="Affine")
]
self.wordbatch = wordbatch.WordBatch(normalize_text,
n_words=self.n_words,
extractors=[(wordbatch.WordSeq, {
"seq_maxlen":
self.maxlen
})])
if datadir == None:
self.model = Model(self.layers)
self.model.load_params(pickle_model)
self.wordbatch = pkl.load(gzip.open(pickle_model + ".wb", 'rb'))
else:
self.train(datadir, pickle_model)
def remove_unks(self, x):
return [[self.n_words if w >= self.n_words else w for w in sen]
for sen in x]
def format_texts(self, texts):
return self.remove_unks(self.wordbatch.transform(texts))
class ThreadWithReturnValue(Thread):
def __init__(self,
group=None,
target=None,
name=None,
args=(),
kwargs={},
Verbose=None):
Thread.__init__(self, group, target, name, args, kwargs, Verbose)
self._return = None
def run(self):
if self._Thread__target is not None:
self._return = self._Thread__target(*self._Thread__args,
**self._Thread__kwargs)
def join(self):
Thread.join(self)
return self._return
def train(self, datadir, pickle_model=""):
texts = []
labels = []
training_data = os.listdir(datadir)
rcount = 0
texts2 = []
batchsize = 100000
t = None
for jsonfile in training_data:
with open(datadir + "/" + jsonfile, u'r') as inputfile:
for line in inputfile:
#if rcount > 1000000: break
try:
line = json.loads(line.strip())
except:
continue
for review in line["Reviews"]:
rcount += 1
if rcount % 100000 == 0: print rcount
if rcount % 8 != 0: continue
if "Overall" not in review["Ratings"]: continue
texts.append(review["Content"])
labels.append(
(float(review["Ratings"]["Overall"]) - 3) * 0.5)
if len(texts) % batchsize == 0:
if t != None: texts2.append(t.join())
t = self.ThreadWithReturnValue(
target=self.wordbatch.transform,
args=(texts, ))
t.start()
texts = []
texts2.append(t.join())
texts2.append(self.wordbatch.transform(texts))
del (texts)
texts = sp.vstack(texts2)
self.wordbatch.dictionary_freeze = True
train = [
np.asarray(texts, dtype='int32'),
np.asanyarray(labels, dtype='float32')
]
train[1].shape = (train[1].shape[0], 1)
num_epochs = 10
cost = GeneralizedCost(costfunc=SumSquared())
self.model = Model(layers=self.layers)
optimizer = Adam(learning_rate=0.01)
index_shuf = list(range(len(train[0])))
random.shuffle(index_shuf)
train[0] = np.asarray([train[0][x] for x in index_shuf], dtype='int32')
train[1] = np.asarray([train[1][x] for x in index_shuf],
dtype='float32')
train_iter = ArrayIterator(train[0],
train[1],
nclass=1,
make_onehot=False)
self.model.fit(train_iter,
optimizer=optimizer,
num_epochs=num_epochs,
cost=cost,
callbacks=Callbacks(self.model,
**self.args.callback_args))
if pickle_model != "":
self.model.save_params(pickle_model)
with gzip.open(pickle_model + ".wb", 'wb') as model_file:
pkl.dump(self.wordbatch, model_file, protocol=2)
def predict_batch(self, texts):
input = np.array(self.format_texts(texts))
output = np.zeros((texts.shape[0], 1))
test = ArrayIterator(input, output, nclass=1, make_onehot=False)
results = [row[0] for row in self.model.get_outputs(test)]
return results
from neon.layers import GeneralizedCost
from neon.transforms import CrossEntropyMulti
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
from neon.optimizers import GradientDescentMomentum, RMSProp
optimizer = GradientDescentMomentum(learning_rate=0.005,
momentum_coef=0.9)
# Set up callbacks. By default sets up a progress bar
from neon.callbacks.callbacks import Callbacks
callbacks = Callbacks(model, train_set)
model.fit(dataset=train_set, cost=cost, optimizer=optimizer, num_epochs=num_epochs, callbacks=callbacks)
model.save_params("cifar10_model.prm")
# Evaluate performance
from neon.transforms import Misclassification
error_pct = 100 * model.eval(test_set, metric=Misclassification())
print 'Misclassification error = %.1f%%' % error_pct
# Sanity check 1
# an image of a frog from wikipedia
# img_source = "https://upload.wikimedia.org/wikipedia/commons/thumb/5/55/Atelopus_zeteki1.jpg/440px-Atelopus_zeteki1.jpg"
# import urllib
# urllib.urlretrieve(img_source, filename="image.jpg")
from PIL import Image
import numpy as np
class SequenceChunker(object):
"""
Sequence chunker model (Neon based)
Args:
sentence_length (str): max sentence length
token_vocab_size (int): word vocabulary size
pos_vocab_size (int, optional): POS vocabulary size
char_vocab_size (int, optional): characters vocabulary size
max_char_word_length (int, optional): max word length in characters
token_embedding_size (int, optional): word embedding dims
pos_embedding_size (int, optional): POS embedding dims
char_embedding_size (int, optional): character embedding dims
num_labels (int, optional): number of output labels possible per token
lstm_hidden_size (int, optional): LSTM hidden size
num_lstm_layers (int, optional): number of LSTM layers
use_external_embedding (bool, optional): input is provided as external word embedding
dropout (float, optional): dropout rate
"""
def __init__(self, sentence_length,
token_vocab_size,
pos_vocab_size=None,
char_vocab_size=None,
max_char_word_length=20,
token_embedding_size=None,
pos_embedding_size=None,
char_embedding_size=None,
num_labels=None,
lstm_hidden_size=100,
num_lstm_layers=1,
use_external_embedding=None,
dropout=0.5
):
init = GlorotUniform()
tokens = []
if use_external_embedding is None:
tokens.append(LookupTable(vocab_size=token_vocab_size,
embedding_dim=token_embedding_size,
init=init,
pad_idx=0))
else:
tokens.append(DataInput())
tokens.append(Reshape((-1, sentence_length)))
f_layers = [tokens]
# add POS tag input
if pos_vocab_size is not None and pos_embedding_size is not None:
f_layers.append([
LookupTable(vocab_size=pos_vocab_size,
embedding_dim=pos_embedding_size,
init=init,
pad_idx=0),
Reshape((-1, sentence_length))
])
# add Character RNN input
if char_vocab_size is not None and char_embedding_size is not None:
char_lut_layer = LookupTable(vocab_size=char_vocab_size,
embedding_dim=char_embedding_size,
init=init,
pad_idx=0)
char_nn = [char_lut_layer,
TimeDistBiLSTM(char_embedding_size, init, activation=Logistic(),
gate_activation=Tanh(),
reset_cells=True, reset_freq=max_char_word_length),
TimeDistributedRecurrentLast(timesteps=max_char_word_length),
Reshape((-1, sentence_length))]
f_layers.append(char_nn)
layers = []
if len(f_layers) == 1:
layers.append(f_layers[0][0])
else:
layers.append(MergeMultistream(layers=f_layers, merge="stack"))
layers.append(Reshape((-1, sentence_length)))
layers += [DeepBiLSTM(lstm_hidden_size, init, activation=Logistic(),
gate_activation=Tanh(),
reset_cells=True,
depth=num_lstm_layers),
Dropout(keep=dropout),
Affine(num_labels, init, bias=init, activation=Softmax())]
self._model = Model(layers=layers)
def fit(self, dataset, optimizer, cost, callbacks, epochs=10):
"""
fit a model
Args:
dataset: train/test set of CONLL2000 dataset
optimizer: optimizer (Neon based)
cost: cost function (Neon based)
callbacks: callbacks (Neon based)
epochs (int, optional): number of epochs to train
"""
self._model.fit(dataset,
optimizer=optimizer,
num_epochs=epochs,
cost=cost,
callbacks=callbacks)
def predict(self, dataset):
"""
predict output of given dataset
Args:
dataset: Neon based iterator
Returns:
prediction on given dataset
"""
return self._model.get_outputs(dataset)
def save(self, path):
"""
Save model weights to path
Args:
path (str): path to weights file
"""
self._model.save_params(path)
def get_model(self):
"""
Get model
Returns:
Neon model object
"""
return self._model
class SequenceChunker(object):
"""
Sequence chunker model (Neon based)
Args:
sentence_length (str): max sentence length
token_vocab_size (int): word vocabulary size
pos_vocab_size (int, optional): POS vocabulary size
char_vocab_size (int, optional): characters vocabulary size
max_char_word_length (int, optional): max word length in characters
token_embedding_size (int, optional): word embedding dims
pos_embedding_size (int, optional): POS embedding dims
char_embedding_size (int, optional): character embedding dims
num_labels (int, optional): number of output labels possible per token
lstm_hidden_size (int, optional): LSTM hidden size
num_lstm_layers (int, optional): number of LSTM layers
use_external_embedding (bool, optional): input is provided as external word embedding
dropout (float, optional): dropout rate
"""
def __init__(self,
sentence_length,
token_vocab_size,
pos_vocab_size=None,
char_vocab_size=None,
max_char_word_length=20,
token_embedding_size=None,
pos_embedding_size=None,
char_embedding_size=None,
num_labels=None,
lstm_hidden_size=100,
num_lstm_layers=1,
use_external_embedding=None,
dropout=0.5):
init = GlorotUniform()
tokens = []
if use_external_embedding is None:
tokens.append(
LookupTable(vocab_size=token_vocab_size,
embedding_dim=token_embedding_size,
init=init,
pad_idx=0))
else:
tokens.append(DataInput())
tokens.append(Reshape((-1, sentence_length)))
f_layers = [tokens]
# add POS tag input
if pos_vocab_size is not None and pos_embedding_size is not None:
f_layers.append([
LookupTable(vocab_size=pos_vocab_size,
embedding_dim=pos_embedding_size,
init=init,
pad_idx=0),
Reshape((-1, sentence_length))
])
# add Character RNN input
if char_vocab_size is not None and char_embedding_size is not None:
char_lut_layer = LookupTable(vocab_size=char_vocab_size,
embedding_dim=char_embedding_size,
init=init,
pad_idx=0)
char_nn = [
char_lut_layer,
TimeDistBiLSTM(char_embedding_size,
init,
activation=Logistic(),
gate_activation=Tanh(),
reset_cells=True,
reset_freq=max_char_word_length),
TimeDistributedRecurrentLast(timesteps=max_char_word_length),
Reshape((-1, sentence_length))
]
f_layers.append(char_nn)
layers = []
if len(f_layers) == 1:
layers.append(f_layers[0][0])
else:
layers.append(MergeMultistream(layers=f_layers, merge="stack"))
layers.append(Reshape((-1, sentence_length)))
layers += [
DeepBiLSTM(lstm_hidden_size,
init,
activation=Logistic(),
gate_activation=Tanh(),
reset_cells=True,
depth=num_lstm_layers),
Dropout(keep=dropout),
Affine(num_labels, init, bias=init, activation=Softmax())
]
self._model = Model(layers=layers)
def fit(self, dataset, optimizer, cost, callbacks, epochs=10):
"""
fit a model
Args:
dataset: train/test set of CONLL2000 dataset
optimizer: optimizer (Neon based)
cost: cost function (Neon based)
callbacks: callbacks (Neon based)
epochs (int, optional): number of epochs to train
"""
self._model.fit(dataset,
optimizer=optimizer,
num_epochs=epochs,
cost=cost,
callbacks=callbacks)
def predict(self, dataset):
"""
predict output of given dataset
Args:
dataset: Neon based iterator
Returns:
prediction on given dataset
"""
return self._model.get_outputs(dataset)
def save(self, path):
"""
Save model weights to path
Args:
path (str): path to weights file
"""
self._model.save_params(path)
def get_model(self):
"""
Get model
Returns:
Neon model object
"""
return self._model
validation=False,
remove_history=False,
minimal_set=False,
next_N=3)
valid = HDF5Iterator(filenames,
ndata=(16 * 2014),
validation=True,
remove_history=False,
minimal_set=False,
next_N=1)
out1, out2, out3 = model.layers.get_terminal()
cost = Multicost(costs=[GeneralizedCost(costfunc=CrossEntropyMulti(usebits=True)),
GeneralizedCost(costfunc=CrossEntropyMulti(usebits=True)),
GeneralizedCost(costfunc=CrossEntropyMulti(usebits=True))])
schedule = ExpSchedule(decay=(1.0 / 50)) # halve the learning rate every 50 epochs
opt_gdm = GradientDescentMomentum(learning_rate=0.01,
momentum_coef=0.9,
stochastic_round=args.rounding,
gradient_clip_value=1,
gradient_clip_norm=5,
wdecay=0.0001,
schedule=schedule)
callbacks = Callbacks(model, eval_set=valid, metric=TopKMisclassification(5), **args.callback_args)
callbacks.add_save_best_state_callback(os.path.join(args.workspace_dir, "best_state_h5resnet.pkl"))
model.fit(train, optimizer=opt_gdm, num_epochs=num_epochs, cost=cost, callbacks=callbacks)
model.save_params(os.path.join(args.workspace_dir, "final_state_h5resnet.pkl"))
]
model = Model(layers=layers)
cost = GeneralizedCost(costfunc=CrossEntropyMulti())
# define optimizer
opt_w = GradientDescentMomentum(learning_rate=0.01,
momentum_coef=0.9,
wdecay=0.0005)
opt_b = GradientDescentMomentum(learning_rate=0.01, momentum_coef=0.9)
opt = MultiOptimizer({'default': opt_w, 'Bias': opt_b}, name='multiopt')
# configure callbacks
callbacks = Callbacks(model,
eval_set=valid_set,
metric=Misclassification(),
**args.callback_args)
callbacks.add_callback(
TrainByStageCallback(model, valid_set, Misclassification(),
max_patience=5))
logger.info('Training ...')
model.fit(train_set,
optimizer=opt,
num_epochs=250,
cost=cost,
callbacks=callbacks)
print('Accuracy = %.2f%%' %
(100. - model.eval(valid_set, metric=Misclassification()) * 100))
model.save_params('./models/mnist/mnist_cnn.pkl')
mlp = Model(layers=layers)
callbacks = Callbacks(mlp, eval_set=test) # Track cost function
# Train model
mlp.fit(train,
optimizer=opt_gdm,
num_epochs=num_epochs,
cost=cost,
callbacks=callbacks)
# Check performance
print 'Misclassification error = %.1f%%' % (
mlp.eval(test, metric=Misclassification()) * 100)
# Save trained model
mlp.save_params(param_file_name)
# Sanity check 1
from PIL import Image
import numpy as np
from neon.data.dataiterator import ArrayIterator
W = img_size
H = img_size
L = W * H * 3
size = H, W
x_new = np.zeros((128, L), dtype=np.float32)
def load_sample(test_file_name):
image = Image.open(test_file_name)
lunaModel = Model(layers=vgg_layers)
if args.model_file:
import os
assert os.path.exists(args.model_file), '%s not found' % args.model_file
lunaModel.load_params(args.model_file)
# configure callbacks
#callbacks = Callbacks(lunaModel, eval_set=valid_set, **args.callback_args)
callbacks = Callbacks(lunaModel, eval_set=valid_set, metric=Misclassification(), **args.callback_args)
if args.deconv:
callbacks.add_deconv_callback(train_set, valid_set)
lunaModel.fit(train_set, optimizer=opt, num_epochs=num_epochs,
cost=cost, callbacks=callbacks)
lunaModel.save_params('LUNA16_VGG_model_no_batch.prm')
# neon_logger.display('Finished training. Calculating error on the validation set...')
# neon_logger.display('Misclassification error (validation) = {:.2f}%'.format(lunaModel.eval(valid_set, metric=Misclassification())[0] * 100))
# neon_logger.display('Precision/recall (validation) = {}'.format(lunaModel.eval(valid_set, metric=PrecisionRecall(num_classes=2))))
# neon_logger.display('Calculating metrics on the test set. This could take a while...')
# neon_logger.display('Misclassification error (test) = {:.2f}%'.format(lunaModel.eval(test_set, metric=Misclassification())[0] * 100))
# neon_logger.display('Precision/recall (test) = {}'.format(lunaModel.eval(test_set, metric=PrecisionRecall(num_classes=2))))
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)
train = HDF5Iterator(filenames,
[h['X'] for h in h5s],
[h['y'] for h in h5s],
ndata=(256 * 1024),
validation=False,
remove_history=True)
valid = HDF5Iterator(filenames,
[h['X'] for h in h5s],
[h['y'] for h in h5s],
ndata=1024,
validation=True,
remove_history=True)
cost = GeneralizedCost(costfunc=CrossEntropyBinary())
opt_gdm = GradientDescentMomentum(learning_rate=0.01,
momentum_coef=0.9,
stochastic_round=args.rounding)
callbacks = Callbacks(model, eval_set=valid, metric=TopKMisclassification(5), **args.callback_args)
old_params = get_model_params(args.server_address)
num_iterations = 1
while True:
update_model(model, old_params)
model.fit(train, optimizer=opt_gdm, num_epochs=1, cost=cost, callbacks=callbacks)
model.save_params(os.path.join(args.workspace_dir, "iter_{}.pkl".format(num_iterations)))
deltas = compute_deltas(old_params, model)
old_params = put_deltas(args.server_address, deltas)
num_iterations += 1
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
targets = preq.asnumpyarray()
# 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.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._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)
lunaModel = Model(layers=vgg_layers)
if args.model_file:
import os
assert os.path.exists(args.model_file), '%s not found' % args.model_file
lunaModel.load_params(args.model_file)
# configure callbacks
#callbacks = Callbacks(lunaModel, eval_set=valid_set, **args.callback_args)
callbacks = Callbacks(lunaModel, eval_set=valid_set, metric=Misclassification(), **args.callback_args)
if args.deconv:
callbacks.add_deconv_callback(train_set, valid_set)
lunaModel.fit(train_set, optimizer=opt, num_epochs=num_epochs,
cost=cost, callbacks=callbacks)
lunaModel.save_params('LUNA16_VGG_model.prm')
neon_logger.display('Finished training. Calculating error on the validation set...')
neon_logger.display('Misclassification error (validation) = {:.2f}%'.format(lunaModel.eval(valid_set, metric=Misclassification())[0] * 100))
neon_logger.display('Precision/recall (validation) = {}'.format(lunaModel.eval(valid_set, metric=PrecisionRecall(num_classes=2))))
neon_logger.display('Calculating metrics on the test set. This could take a while...')
neon_logger.display('Misclassification error (test) = {:.2f}%'.format(lunaModel.eval(test_set, metric=Misclassification())[0] * 100))
neon_logger.display('Precision/recall (test) = {}'.format(lunaModel.eval(test_set, metric=PrecisionRecall(num_classes=2))))
def train_mlp():
"""
Train data and save scaling and network weights and biases to file
to be used by forward prop phase on test data
"""
parser = NeonArgparser(__doc__)
args = parser.parse_args()
logger = logging.getLogger()
logger.setLevel(args.log_thresh)
# hyperparameters
num_epochs = args.epochs
#preprocessor
std_scale = preprocessing.StandardScaler(with_mean=True,with_std=True)
#std_scale = feature_scaler(type='Standardizer',with_mean=True,with_std=True)
#number of non one-hot encoded features, including ground truth
num_feat = 4
# load up the mnist data set
# split into train and tests sets
#load data from csv-files and rescale
#training
traindf = pd.DataFrame.from_csv('data/train.csv')
ncols = traindf.shape[1]
#tmpmat=std_scale.fit_transform(traindf.as_matrix())
#print std_scale.scale_
#print std_scale.mean_
tmpmat = traindf.as_matrix()
#print tmpmat[:,1:num_feat]
tmpmat[:,:num_feat] = std_scale.fit_transform(tmpmat[:,:num_feat])
X_train = tmpmat[:,1:]
y_train = np.reshape(tmpmat[:,0],(tmpmat[:,0].shape[0],1))
#validation
validdf = pd.DataFrame.from_csv('data/validate.csv')
ncols = validdf.shape[1]
tmpmat = validdf.as_matrix()
tmpmat[:,:num_feat] = std_scale.transform(tmpmat[:,:num_feat])
X_valid = tmpmat[:,1:]
y_valid = np.reshape(tmpmat[:,0],(tmpmat[:,0].shape[0],1))
#test
testdf = pd.DataFrame.from_csv('data/test.csv')
ncols = testdf.shape[1]
tmpmat = testdf.as_matrix()
tmpmat[:,:num_feat] = std_scale.transform(tmpmat[:,:num_feat])
X_test = tmpmat[:,1:]
y_test = np.reshape(tmpmat[:,0],(tmpmat[:,0].shape[0],1))
# setup a training set iterator
train_set = CustomDataIterator(X_train, lshape=(X_train.shape[1]), y_c=y_train)
# setup a validation data set iterator
valid_set = CustomDataIterator(X_valid, lshape=(X_valid.shape[1]), y_c=y_valid)
# setup a validation data set iterator
test_set = CustomDataIterator(X_test, lshape=(X_test.shape[1]), y_c=y_test)
# setup weight initialization function
init_norm = Xavier()
# setup model layers
layers = [Affine(nout=X_train.shape[1], init=init_norm, activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=X_train.shape[1]/2, init=init_norm, activation=Rectlin()),
Linear(nout=1, init=init_norm)]
# setup cost function as CrossEntropy
cost = GeneralizedCost(costfunc=SmoothL1Loss())
# setup optimizer
#schedule
#schedule = ExpSchedule(decay=0.3)
#optimizer = GradientDescentMomentum(0.0001, momentum_coef=0.9, stochastic_round=args.rounding, schedule=schedule)
optimizer = Adam(learning_rate=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1.e-8)
# initialize model object
mlp = Model(layers=layers)
# configure callbacks
if args.callback_args['eval_freq'] is None:
args.callback_args['eval_freq'] = 1
# configure callbacks
callbacks = Callbacks(mlp, eval_set=valid_set, **args.callback_args)
callbacks.add_early_stop_callback(stop_func)
callbacks.add_save_best_state_callback(os.path.join(args.data_dir, "early_stop-best_state.pkl"))
# run fit
mlp.fit(train_set, optimizer=optimizer, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
#evaluate model
print('Evaluation Error = %.4f'%(mlp.eval(valid_set, metric=SmoothL1Metric())))
print('Test set error = %.4f'%(mlp.eval(test_set, metric=SmoothL1Metric())))
# Saving the model
print 'Saving model parameters!'
mlp.save_params("model/homeapp_model.prm")
# Reloading saved model
# This should go in run.py
mlp=Model("model/homeapp_model.prm")
print('Test set error = %.4f'%(mlp.eval(test_set, metric=SmoothL1Metric())))
# save the preprocessor vectors:
np.savez("model/homeapp_preproc", mean=std_scale.mean_, std=std_scale.scale_)
return 1
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)
def test_model_serialize(backend_default, data):
dataset = MNIST(path=data)
(X_train, y_train), (X_test, y_test), nclass = dataset.load_data()
train_set = ArrayIterator(
[X_train, X_train], y_train, nclass=nclass, lshape=(1, 28, 28))
init_norm = Gaussian(loc=0.0, scale=0.01)
# initialize model
path1 = Sequential([Conv((5, 5, 16), init=init_norm, bias=Constant(0), activation=Rectlin()),
Pooling(2),
Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin())])
path2 = Sequential([Affine(nout=100, init=init_norm, bias=Constant(0), activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=20, init=init_norm, bias=init_norm, activation=Rectlin())])
layers = [MergeMultistream(layers=[path1, path2], merge="stack"),
Affine(nout=20, init=init_norm, batch_norm=True, activation=Rectlin()),
Affine(nout=10, init=init_norm, activation=Logistic(shortcut=True))]
tmp_save = 'test_model_serialize_tmp_save.pickle'
mlp = Model(layers=layers)
mlp.optimizer = GradientDescentMomentum(learning_rate=0.1, momentum_coef=0.9)
mlp.cost = GeneralizedCost(costfunc=CrossEntropyBinary())
mlp.initialize(train_set, cost=mlp.cost)
n_test = 3
num_epochs = 3
# Train model for num_epochs and n_test batches
for epoch in range(num_epochs):
for i, (x, t) in enumerate(train_set):
x = mlp.fprop(x)
delta = mlp.cost.get_errors(x, t)
mlp.bprop(delta)
mlp.optimizer.optimize(mlp.layers_to_optimize, epoch=epoch)
if i > n_test:
break
# Get expected outputs of n_test batches and states of all layers
outputs_exp = []
pdicts_exp = [l.get_params_serialize() for l in mlp.layers_to_optimize]
for i, (x, t) in enumerate(train_set):
outputs_exp.append(mlp.fprop(x, inference=True))
if i > n_test:
break
# Serialize model
mlp.save_params(tmp_save, keep_states=True)
# Load model
mlp = Model(tmp_save)
mlp.initialize(train_set)
outputs = []
pdicts = [l.get_params_serialize() for l in mlp.layers_to_optimize]
for i, (x, t) in enumerate(train_set):
outputs.append(mlp.fprop(x, inference=True))
if i > n_test:
break
# Check outputs, states, and params are the same
for output, output_exp in zip(outputs, outputs_exp):
assert allclose_with_out(output.get(), output_exp.get())
for pd, pd_exp in zip(pdicts, pdicts_exp):
for s, s_e in zip(pd['states'], pd_exp['states']):
if isinstance(s, list): # this is the batch norm case
for _s, _s_e in zip(s, s_e):
assert allclose_with_out(_s, _s_e)
else:
assert allclose_with_out(s, s_e)
for p, p_e in zip(pd['params'], pd_exp['params']):
assert type(p) == type(p_e)
if isinstance(p, list): # this is the batch norm case
for _p, _p_e in zip(p, p_e):
assert allclose_with_out(_p, _p_e)
elif isinstance(p, np.ndarray):
assert allclose_with_out(p, p_e)
else:
assert p == p_e
os.remove(tmp_save)
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)
