def learning_algorithm(args):
name = args.algorithm
learning_rate = float(args.learning_rate)
momentum = args.momentum
clipping_threshold = args.clipping
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
if name == 'adam':
adam = Adam(learning_rate=learning_rate)
step_rule = CompositeRule([adam, clipping])
learning_rate = adam.learning_rate
elif name == 'rms_prop':
rms_prop = RMSProp(learning_rate=learning_rate)
step_rule = CompositeRule([clipping, rms_prop])
learning_rate = rms_prop.learning_rate
elif name == 'momentum':
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
step_rule = CompositeRule([clipping, sgd_momentum])
learning_rate = sgd_momentum.learning_rate
elif name == 'sgd':
sgd = Scale(learning_rate=learning_rate)
step_rule = CompositeRule([clipping, sgd])
learning_rate = sgd.learning_rate
else:
raise NotImplementedError
return step_rule, learning_rate
def test_momentum():
a = shared_floatx([3, 4])
cost = (a**2).sum()
steps, updates = Momentum(0.1, 0.5).compute_steps(
OrderedDict([(a, tensor.grad(cost, a))]))
f = theano.function([], [steps[a]], updates=updates)
assert_allclose(f()[0], [0.6, 0.8])
assert_allclose(f()[0], [0.9, 1.2])
assert_allclose(f()[0], [1.05, 1.4])
def learning_algorithm(args):
name = args.algorithm
learning_rate = float(args.learning_rate)
momentum = args.momentum
clipping_threshold = args.clipping
if name == 'adam':
adam = Adam(learning_rate=learning_rate)
step_rule = adam
elif name == 'rms_prop':
rms_prop = RMSProp(learning_rate=learning_rate, decay_rate=0.9)
step_rule = CompositeRule([StepClipping(1.), rms_prop])
else:
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
step_rule = sgd_momentum
return step_rule
def learning_algorithm(args):
name = args.algorithm
learning_rate = float(args.learning_rate)
momentum = args.momentum
clipping_threshold = args.clipping
if name == 'adam':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
adam = Adam(learning_rate=learning_rate)
step_rule = CompositeRule([adam, clipping])
elif name == 'rms_prop':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
rms_prop = RMSProp(learning_rate=learning_rate)
step_rule = CompositeRule([clipping, rms_prop])
else:
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
step_rule = CompositeRule([clipping, sgd_momentum])
return step_rule
def train_model(cost,
error_rate,
train_stream,
load_location=None,
save_location=None):
cost.name = "Cross_entropy"
error_rate.name = 'Error_rate'
# Define the model
model = Model(cost)
# Load the parameters from a dumped model
if load_location is not None:
logger.info('Loading parameters...')
model.set_param_values(load_parameter_values(load_location))
cg = ComputationGraph(cost)
step_rule = Momentum(learning_rate=0.1, momentum=0.9)
algorithm = GradientDescent(cost=cost,
step_rule=step_rule,
params=cg.parameters)
main_loop = MainLoop(
model=model,
data_stream=train_stream,
algorithm=algorithm,
extensions=[
# DataStreamMonitoring([cost], test_stream, prefix='test',
# after_epoch=False, every_n_epochs=10),
DataStreamMonitoring([cost],
train_stream,
prefix='train',
after_epoch=True),
Printing(after_epoch=True)
])
main_loop.run()
# Save the main loop
if save_location is not None:
logger.info('Saving the main loop...')
dump_manager = MainLoopDumpManager(save_location)
dump_manager.dump(main_loop)
logger.info('Saved')
def learning_algorithm(args):
name = args.algorithm
learning_rate = float(args.learning_rate)
momentum = args.momentum
clipping_threshold = args.clipping
if name == 'adam':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
adam = Adam(learning_rate=learning_rate)
# [adam, clipping] means 'step clipping'
# [clipping, adam] means 'gradient clipping'
step_rule = CompositeRule([adam, clipping])
elif name == 'rms_prop':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
rms_prop = RMSProp(learning_rate=learning_rate)
rm_non_finite = RemoveNotFinite()
step_rule = CompositeRule([clipping, rms_prop, rm_non_finite])
else:
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
rm_non_finite = RemoveNotFinite()
step_rule = CompositeRule([clipping, sgd_momentum, rm_non_finite])
return step_rule
def learning_algorithm(learning_rate,
momentum=0.0,
clipping_threshold=100,
algorithm='sgd'):
if algorithm == 'adam':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
adam = Adam(learning_rate=learning_rate)
# [adam, clipping] means 'step clipping'
# [clipping, adam] means 'gradient clipping'
step_rule = CompositeRule([adam, clipping])
elif algorithm == 'rms_prop':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
rms_prop = RMSProp(learning_rate=learning_rate)
rm_non_finite = RemoveNotFinite()
step_rule = CompositeRule([clipping, rms_prop, rm_non_finite])
elif algorithm == 'sgd':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
rm_non_finite = RemoveNotFinite()
step_rule = CompositeRule([clipping, sgd_momentum, rm_non_finite])
else:
raise NotImplementedError
return step_rule
def setup_algorithms(cost, cg, method, type="ff"):
"""Setup training algorithm.
Parameters
----------
cost : expression
cost expression
cg : ComputationGraph
Computation graph
method : string
training method: SGD, momentum SGD, AdaGrad, RMSprop
learning_rate : float
learning rate for learning method
Returns
-------
algorithm : GradientDescent
Gradient Descent algorithm based on different optimization method
"""
if method == "sgd":
step_rule = Scale(learning_rate=0.01)
elif method == "momentum":
step_rule = Momentum(learning_rate=0.01, momentum=0.95)
elif method == "adagrad":
step_rule = AdaGrad()
elif method == "rmsprop":
step_rule = RMSProp()
if type == "RNN":
step_rule = CompositeRule([StepClipping(1.0), step_rule])
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=step_rule)
return algorithm
def main(feature_maps=None, mlp_hiddens=None,
conv_sizes=None, pool_sizes=None, batch_size=None,
num_batches=None):
if feature_maps is None:
feature_maps = [32, 48, 64, 96, 96, 128]
if mlp_hiddens is None:
mlp_hiddens = [1000]
if conv_sizes is None:
conv_sizes = [9, 7, 5, 3, 2, 1]
if pool_sizes is None:
pool_sizes = [2, 2, 2, 2, 1, 1]
if batch_size is None:
batch_size = 64
conv_steps=[2, 1, 1, 1, 1, 1] #same as stride
image_size = (128, 128)
output_size = 2
learningRate = 0.001
drop_prob = 0.4
weight_noise = 0.75
num_epochs = 150
num_batches = None
host_plot='http://*****:*****@ %s' % (graph_name, datetime.datetime.now(), socket.gethostname()),
channels=[['train_error_rate', 'valid_error_rate'],
['train_total_gradient_norm']], after_epoch=True, server_url=host_plot))
PLOT_AVAILABLE = True
except ImportError:
PLOT_AVAILABLE = False
extensions.append(Checkpoint(save_to, after_epoch=True, after_training=True, save_separately=['log']))
logger.info("Building the model")
model = Model(cost)
########### Loading images #####################
main_loop = MainLoop(
algorithm,
stream_data_train,
model=model,
extensions=extensions)
main_loop.run()
cost_monitor = cost_train
else:
raise ValueError, conf.task
recognizer.initialize()
cg = ComputationGraph([cost_train, y_hat, x_m, y, y_m])
weights = VariableFilter(roles=[WEIGHT])(cg.variables)
cg = apply_noise(cg, weights, conf.weight_noise)
#************* training algorithm *************
model = Model(cost_train)
if conf.step_rule == 'AdaDelta':
step_rule = AdaDelta()
elif conf.step_rule == 'Momentum':
step_rule = Momentum(learning_rate=conf.learning_rate,
momentum=conf.momentum)
else:
raise ('step_rule not known: {}'.format(conf.step_rule))
step_rule = CompositeRule([step_rule, StepClipping(conf.step_clipping)])
algorithm = GradientDescent(cost=cost_train,
parameters=cg.parameters,
step_rule=step_rule)
#***************** main loop ****************
train_monitor = TrainingDataMonitoring([cost_monitor], prefix="train")
val_monitor = DataStreamMonitoring([cost_monitor], stream_val, prefix="val")
if conf.task == 'CTC':
PER_Monitor = PhonemeErrorRate
elif conf.task == 'framewise':
def train(algorithm, learning_rate, clipping, momentum, layer_size, epochs,
test_cost, experiment_path, initialization, init_width, weight_noise,
z_prob, z_prob_states, z_prob_cells, drop_prob_igates,
ogates_zoneout, batch_size, stoch_depth, share_mask, gaussian_drop,
rnn_type, num_layers, norm_cost_coeff, penalty, testing, seq_len,
decrease_lr_after_epoch, lr_decay, **kwargs):
print '.. PTB experiment'
print '.. arguments:', ' '.join(sys.argv)
t0 = time.time()
###########################################
#
# LOAD DATA
#
###########################################
def onehot(x, numclasses=None):
""" Convert integer encoding for class-labels (starting with 0 !)
to one-hot encoding.
The output is an array whose shape is the shape of the input array
plus an extra dimension, containing the 'one-hot'-encoded labels.
"""
if x.shape == ():
x = x[None]
if numclasses is None:
numclasses = x.max() + 1
result = numpy.zeros(list(x.shape) + [numclasses], dtype="int")
z = numpy.zeros(x.shape, dtype="int")
for c in range(numclasses):
z *= 0
z[numpy.where(x == c)] = 1
result[..., c] += z
return result.astype(theano.config.floatX)
alphabetsize = 10000
data = np.load('penntree_char_and_word.npz')
trainset = data['train_words']
validset = data['valid_words']
testset = data['test_words']
if testing:
trainset = trainset[:3000]
validset = validset[:3000]
if share_mask:
if not z_prob:
raise ValueError('z_prob must be provided when using share_mask')
if z_prob_cells or z_prob_states:
raise ValueError(
'z_prob_states and z_prob_cells must not be provided when using share_mask (use z_prob instead)'
)
z_prob_cells = z_prob
# we don't want to actually use these masks, so this is to debug
z_prob_states = None
else:
if z_prob:
raise ValueError('z_prob is only used with share_mask')
z_prob_cells = z_prob_cells or '1'
z_prob_states = z_prob_states or '1'
# rng = np.random.RandomState(seed)
###########################################
#
# MAKE STREAMS
#
###########################################
def prep_dataset(dataset):
dataset = dataset[:(len(dataset) - (len(dataset) %
(seq_len * batch_size)))]
dataset = dataset.reshape(batch_size, -1, seq_len).transpose((1, 0, 2))
stream = DataStream(
IndexableDataset(indexables=OrderedDict([('data', dataset)])),
iteration_scheme=SequentialExampleScheme(dataset.shape[0]))
stream = Transpose(stream, [(1, 0)])
stream = SampleDropsNPWord(stream, z_prob_states, z_prob_cells,
drop_prob_igates, layer_size, num_layers,
False, stoch_depth, share_mask,
gaussian_drop, alphabetsize)
stream.sources = ('data', ) * 3 + stream.sources + (
'zoneouts_states', 'zoneouts_cells', 'zoneouts_igates')
return (stream, )
train_stream, = prep_dataset(trainset)
valid_stream, = prep_dataset(validset)
test_stream, = prep_dataset(testset)
####################
data = train_stream.get_epoch_iterator(as_dict=True).next()
####################
###########################################
#
# BUILD MODEL
#
###########################################
print '.. building model'
x = T.tensor3('data')
y = x
zoneouts_states = T.tensor3('zoneouts_states')
zoneouts_cells = T.tensor3('zoneouts_cells')
zoneouts_igates = T.tensor3('zoneouts_igates')
x.tag.test_value = data['data']
zoneouts_states.tag.test_value = data['zoneouts_states']
zoneouts_cells.tag.test_value = data['zoneouts_cells']
zoneouts_igates.tag.test_value = data['zoneouts_igates']
if init_width and not initialization == 'uniform':
raise ValueError('Width is only for uniform init, whassup?')
if initialization == 'glorot':
weights_init = NormalizedInitialization()
elif initialization == 'uniform':
weights_init = Uniform(width=init_width)
elif initialization == 'ortho':
weights_init = OrthogonalInitialization()
else:
raise ValueError('No such initialization')
if rnn_type.lower() == 'lstm':
in_to_hids = [
Linear(layer_size if l > 0 else alphabetsize,
layer_size * 4,
name='in_to_hid%d' % l,
weights_init=weights_init,
biases_init=Constant(0.0)) for l in range(num_layers)
]
recurrent_layers = [
DropLSTM(dim=layer_size,
weights_init=weights_init,
activation=Tanh(),
model_type=6,
name='rnn%d' % l,
ogates_zoneout=ogates_zoneout) for l in range(num_layers)
]
elif rnn_type.lower() == 'gru':
in_to_hids = [
Linear(layer_size if l > 0 else alphabetsize,
layer_size * 3,
name='in_to_hid%d' % l,
weights_init=weights_init,
biases_init=Constant(0.0)) for l in range(num_layers)
]
recurrent_layers = [
DropGRU(dim=layer_size,
weights_init=weights_init,
activation=Tanh(),
name='rnn%d' % l) for l in range(num_layers)
]
elif rnn_type.lower() == 'srnn': # FIXME!!! make ReLU
in_to_hids = [
Linear(layer_size if l > 0 else alphabetsize,
layer_size,
name='in_to_hid%d' % l,
weights_init=weights_init,
biases_init=Constant(0.0)) for l in range(num_layers)
]
recurrent_layers = [
DropSimpleRecurrent(dim=layer_size,
weights_init=weights_init,
activation=Rectifier(),
name='rnn%d' % l) for l in range(num_layers)
]
else:
raise NotImplementedError
hid_to_out = Linear(layer_size,
alphabetsize,
name='hid_to_out',
weights_init=weights_init,
biases_init=Constant(0.0))
for layer in in_to_hids:
layer.initialize()
for layer in recurrent_layers:
layer.initialize()
hid_to_out.initialize()
layer_input = x #in_to_hid.apply(x)
init_updates = OrderedDict()
for l, (in_to_hid, layer) in enumerate(zip(in_to_hids, recurrent_layers)):
rnn_embedding = in_to_hid.apply(layer_input)
if rnn_type.lower() == 'lstm':
states_init = theano.shared(
np.zeros((batch_size, layer_size), dtype=floatX))
cells_init = theano.shared(
np.zeros((batch_size, layer_size), dtype=floatX))
states_init.name, cells_init.name = "states_init", "cells_init"
states, cells = layer.apply(
rnn_embedding,
zoneouts_states[:, :, l * layer_size:(l + 1) * layer_size],
zoneouts_cells[:, :, l * layer_size:(l + 1) * layer_size],
zoneouts_igates[:, :, l * layer_size:(l + 1) * layer_size],
states_init, cells_init)
init_updates.update([(states_init, states[-1]),
(cells_init, cells[-1])])
elif rnn_type.lower() in ['gru', 'srnn']:
# untested!
states_init = theano.shared(
np.zeros((batch_size, layer_size), dtype=floatX))
states_init.name = "states_init"
states = layer.apply(rnn_embedding, zoneouts_states,
zoneouts_igates, states_init)
init_updates.update([(states_init, states[-1])])
else:
raise NotImplementedError
layer_input = states
y_hat_pre_softmax = hid_to_out.apply(T.join(0, [states_init], states[:-1]))
shape_ = y_hat_pre_softmax.shape
y_hat = Softmax().apply(y_hat_pre_softmax.reshape((-1, alphabetsize)))
####################
###########################################
#
# SET UP COSTS AND MONITORS
#
###########################################
cost = CategoricalCrossEntropy().apply(y.reshape((-1, alphabetsize)),
y_hat).copy('cost')
bpc = (cost / np.log(2.0)).copy(name='bpr')
perp = T.exp(cost).copy(name='perp')
cost_train = cost.copy(name='train_cost')
cg_train = ComputationGraph([cost_train])
###########################################
#
# NORM STABILIZER
#
###########################################
norm_cost = 0.
def _magnitude(x, axis=-1):
return T.sqrt(
T.maximum(T.sqr(x).sum(axis=axis),
numpy.finfo(x.dtype).tiny))
if penalty == 'cells':
assert VariableFilter(roles=[MEMORY_CELL])(cg_train.variables)
for cell in VariableFilter(roles=[MEMORY_CELL])(cg_train.variables):
norms = _magnitude(cell)
norm_cost += T.mean(
T.sum((norms[1:] - norms[:-1])**2, axis=0) / (seq_len - 1))
elif penalty == 'hids':
for l in range(num_layers):
assert 'rnn%d_apply_states' % l in [
o.name
for o in VariableFilter(roles=[OUTPUT])(cg_train.variables)
]
for output in VariableFilter(roles=[OUTPUT])(cg_train.variables):
for l in range(num_layers):
if output.name == 'rnn%d_apply_states' % l:
norms = _magnitude(output)
norm_cost += T.mean(
T.sum((norms[1:] - norms[:-1])**2, axis=0) /
(seq_len - 1))
norm_cost.name = 'norm_cost'
#cost_valid = cost_train
cost_train += norm_cost_coeff * norm_cost
cost_train = cost_train.copy(
'cost_train') #should this be cost_train.outputs[0]? no.
cg_train = ComputationGraph([cost_train])
###########################################
#
# WEIGHT NOISE
#
###########################################
if weight_noise > 0:
weights = VariableFilter(roles=[WEIGHT])(cg_train.variables)
cg_train = apply_noise(cg_train, weights, weight_noise)
cost_train = cg_train.outputs[0].copy(name='cost_train')
model = Model(cost_train)
learning_rate = float(learning_rate)
clipping = StepClipping(threshold=np.cast[floatX](clipping))
if algorithm == 'adam':
adam = Adam(learning_rate=learning_rate)
learning_rate = adam.learning_rate
step_rule = CompositeRule([adam, clipping])
elif algorithm == 'rms_prop':
rms_prop = RMSProp(learning_rate=learning_rate)
learning_rate = rms_prop.learning_rate
step_rule = CompositeRule([clipping, rms_prop])
elif algorithm == 'momentum':
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
learning_rate = sgd_momentum.learning_rate
step_rule = CompositeRule([clipping, sgd_momentum])
elif algorithm == 'sgd':
sgd = Scale(learning_rate=learning_rate)
learning_rate = sgd.learning_rate
step_rule = CompositeRule([clipping, sgd])
else:
raise NotImplementedError
algorithm = GradientDescent(step_rule=step_rule,
cost=cost_train,
parameters=cg_train.parameters)
# theano_func_kwargs={"mode": theano.compile.MonitorMode(post_func=detect_nan)})
algorithm.add_updates(init_updates)
def cond_number(x):
_, _, sing_vals = T.nlinalg.svd(x, True, True)
sing_mags = abs(sing_vals)
return T.max(sing_mags) / T.min(sing_mags)
def rms(x):
return (x * x).mean().sqrt()
whysplode_cond = []
whysplode_rms = []
for i, p in enumerate(init_updates):
v = p.get_value()
if p.get_value().shape == 2:
whysplode_cond.append(
cond_number(p).copy(
'ini%d:%s_cond(%s)' %
(i, p.name, "x".join(map(str,
p.get_value().shape)))))
whysplode_rms.append(
rms(p).copy('ini%d:%s_rms(%s)' %
(i, p.name, "x".join(map(str,
p.get_value().shape)))))
for i, p in enumerate(cg_train.parameters):
v = p.get_value()
if p.get_value().shape == 2:
whysplode_cond.append(
cond_number(p).copy(
'ini%d:%s_cond(%s)' %
(i, p.name, "x".join(map(str,
p.get_value().shape)))))
whysplode_rms.append(
rms(p).copy('ini%d:%s_rms(%s)' %
(i, p.name, "x".join(map(str,
p.get_value().shape)))))
observed_vars = [
cost_train, cost, bpc, perp, learning_rate,
aggregation.mean(
algorithm.total_gradient_norm).copy("gradient_norm_mean")
] # + whysplode_rms
parameters = model.get_parameter_dict()
for name, param in parameters.iteritems():
observed_vars.append(param.norm(2).copy(name=name + "_norm"))
observed_vars.append(
algorithm.gradients[param].norm(2).copy(name=name + "_grad_norm"))
train_monitor = TrainingDataMonitoring(variables=observed_vars,
prefix="train",
after_epoch=True)
dev_inits = [p.clone() for p in init_updates]
cg_dev = ComputationGraph([cost, bpc, perp] +
init_updates.values()).replace(
zip(init_updates.keys(), dev_inits))
dev_cost, dev_bpc, dev_perp = cg_dev.outputs[:3]
dev_init_updates = OrderedDict(zip(dev_inits, cg_dev.outputs[3:]))
dev_monitor = DataStreamMonitoring(variables=[dev_cost, dev_bpc, dev_perp],
data_stream=valid_stream,
prefix="dev",
updates=dev_init_updates)
# noone does this
if 'load_path' in kwargs:
with open(kwargs['load_path']) as f:
loaded = np.load(f)
model = Model(cost_train)
params_dicts = model.get_parameter_dict()
params_names = params_dicts.keys()
for param_name in params_names:
param = params_dicts[param_name]
# '/f_6_.W' --> 'f_6_.W'
slash_index = param_name.find('/')
param_name = param_name[slash_index + 1:]
if param.get_value().shape == loaded[param_name].shape:
print 'Found: ' + param_name
param.set_value(loaded[param_name])
else:
print 'Not found: ' + param_name
extensions = []
extensions.extend(
[FinishAfter(after_n_epochs=epochs), train_monitor, dev_monitor])
if test_cost:
test_inits = [p.clone() for p in init_updates]
cg_test = ComputationGraph([cost, bpc, perp] +
init_updates.values()).replace(
zip(init_updates.keys(), test_inits))
test_cost, test_bpc, test_perp = cg_test.outputs[:3]
test_init_updates = OrderedDict(zip(test_inits, cg_test.outputs[3:]))
test_monitor = DataStreamMonitoring(
variables=[test_cost, test_bpc, test_perp],
data_stream=test_stream,
prefix="test",
updates=test_init_updates)
extensions.extend([test_monitor])
if not os.path.exists(experiment_path):
os.makedirs(experiment_path)
log_path = os.path.join(experiment_path, 'log.txt')
fh = logging.FileHandler(filename=log_path)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
extensions.append(
SaveParams('dev_cost', model, experiment_path, every_n_epochs=1))
extensions.append(SaveLog(every_n_epochs=1))
extensions.append(ProgressBar())
extensions.append(Printing())
class RollsExtension(TrainingExtension):
""" rolls the cell and state activations between epochs so that first batch gets correct initial activations """
def __init__(self, shvars):
self.shvars = shvars
def before_epoch(self):
for v in self.shvars:
v.set_value(np.roll(v.get_value(), 1, 0))
extensions.append(
RollsExtension(init_updates.keys() + dev_init_updates.keys() +
(test_init_updates.keys() if test_cost else [])))
class LearningRateSchedule(TrainingExtension):
""" Lets you set a number to divide learning rate by each epoch + when to start doing that """
def __init__(self):
self.epoch_number = 0
def after_epoch(self):
self.epoch_number += 1
if self.epoch_number > decrease_lr_after_epoch:
learning_rate.set_value(learning_rate.get_value() / lr_decay)
if bool(lr_decay) != bool(decrease_lr_after_epoch):
raise ValueError(
'Need to define both lr_decay and decrease_lr_after_epoch')
if lr_decay and decrease_lr_after_epoch:
extensions.append(LearningRateSchedule())
main_loop = MainLoop(model=model,
data_stream=train_stream,
algorithm=algorithm,
extensions=extensions)
t1 = time.time()
print "Building time: %f" % (t1 - t0)
main_loop.run()
print "Execution time: %f" % (time.time() - t1)
from blocks.roles import WEIGHT
from blocks.graph import ComputationGraph, apply_noise, apply_dropout
from datastream import RandomTransposeIt
step_rule_name = 'adam'
learning_rate = 0.1
momentum = 0.9
if step_rule_name == 'adadelta':
step_rule = AdaDelta()
elif step_rule_name == 'rmsprop':
step_rule = RMSProp()
elif step_rule_name == 'momentum':
step_rule_name = "mom%s,%s" % (repr(learning_rate), repr(momentum))
step_rule = Momentum(learning_rate=learning_rate, momentum=momentum)
elif step_rule_name == 'adam':
step_rule = Adam()
else:
raise ValueError("No such step rule: " + step_rule_name)
ibatchsize = None
iter_scheme = RandomTransposeIt(ibatchsize, False, None, False)
valid_iter_scheme = RandomTransposeIt(ibatchsize, False, None, False)
w_noise_std = 0.05
r_dropout = 0.0
s_dropout = 0.0
i_dropout = 0.0
a_dropout = 0.0
def main(save_to,
num_epochs,
regularization=0.001,
subset=None,
num_batches=None,
batch_size=None,
histogram=None,
resume=False):
output_size = 10
convnet = create_all_conv_net()
x = tensor.tensor4('features')
y = tensor.lmatrix('targets')
# Normalize input and apply the convnet
probs = convnet.apply(x)
test_cost = (CategoricalCrossEntropy().apply(y.flatten(),
probs).copy(name='cost'))
test_components = (ComponentwiseCrossEntropy().apply(
y.flatten(), probs).copy(name='components'))
test_error_rate = (MisclassificationRate().apply(
y.flatten(), probs).copy(name='error_rate'))
test_confusion = (ConfusionMatrix().apply(y.flatten(),
probs).copy(name='confusion'))
test_confusion.tag.aggregation_scheme = Sum(test_confusion)
test_cg = ComputationGraph([test_cost, test_error_rate, test_components])
# Apply dropout to all layer outputs except final softmax
dropout_vars = VariableFilter(
roles=[OUTPUT],
bricks=[Convolutional],
theano_name_regex="^conv_[25]_apply_output$")(test_cg.variables)
drop_cg = apply_dropout(test_cg, dropout_vars, 0.5)
# Apply 0.2 dropout to the pre-averaging layer
# dropout_vars_2 = VariableFilter(
# roles=[OUTPUT], bricks=[Convolutional],
# theano_name_regex="^conv_8_apply_output$")(drop_cg.variables)
# train_cg = apply_dropout(drop_cg, dropout_vars_2, 0.2)
# Apply 0.2 dropout to the input, as in the paper
# train_cg = apply_dropout(drop_cg, [x], 0.2)
train_cg = drop_cg
# train_cg = test_cg
train_cost, train_error_rate, train_components = train_cg.outputs
# Apply regularization to the cost
biases = VariableFilter(roles=[BIAS])(train_cg.parameters)
weights = VariableFilter(roles=[WEIGHT])(train_cg.variables)
l2_norm = sum([(W**2).sum() for W in weights])
l2_norm.name = 'l2_norm'
l2_regularization = regularization * l2_norm
l2_regularization.name = 'l2_regularization'
test_cost = test_cost + l2_regularization
test_cost.name = 'cost_with_regularization'
# Training version of cost
train_cost_without_regularization = train_cost
train_cost_without_regularization.name = 'cost_without_regularization'
train_cost = train_cost + regularization * l2_norm
train_cost.name = 'cost_with_regularization'
cifar10_train = CIFAR10(("train", ))
#cifar10_train_stream = RandomPadCropFlip(
# NormalizeBatchLevels(DataStream.default_stream(
# cifar10_train, iteration_scheme=ShuffledScheme(
# cifar10_train.num_examples, batch_size)),
# which_sources=('features',)),
# (32, 32), pad=5, which_sources=('features',))
cifar10_train_stream = NormalizeBatchLevels(DataStream.default_stream(
cifar10_train,
iteration_scheme=ShuffledScheme(cifar10_train.num_examples,
batch_size)),
which_sources=('features', ))
test_batch_size = 1000
cifar10_test = CIFAR10(("test", ))
cifar10_test_stream = NormalizeBatchLevels(DataStream.default_stream(
cifar10_test,
iteration_scheme=ShuffledScheme(cifar10_test.num_examples,
test_batch_size)),
which_sources=('features', ))
momentum = Momentum(0.002, 0.9)
# Create a step rule that doubles the learning rate of biases, like Caffe.
# scale_bias = Restrict(Scale(2), biases)
# step_rule = CompositeRule([scale_bias, momentum])
# step_rule = CompositeRule([StepClipping(100), momentum])
step_rule = momentum
# Train with simple SGD
algorithm = GradientDescent(cost=train_cost,
parameters=train_cg.parameters,
step_rule=step_rule)
# `Timing` extension reports time for reading data, aggregating a batch
# and monitoring;
# `ProgressBar` displays a nice progress bar during training.
extensions = [
Timing(),
FinishAfter(after_n_epochs=num_epochs, after_n_batches=num_batches),
EpochSchedule(momentum.learning_rate, [(1, 0.005), (3, 0.01),
(5, 0.02), (200, 0.002),
(250, 0.0002), (300, 0.00002)]),
DataStreamMonitoring([test_cost, test_error_rate, test_confusion],
cifar10_test_stream,
prefix="test"),
TrainingDataMonitoring([
train_cost, train_error_rate, train_cost_without_regularization,
l2_regularization, momentum.learning_rate,
aggregation.mean(algorithm.total_gradient_norm)
],
prefix="train",
every_n_batches=10),
# after_epoch=True),
Plot('Training performance for ' + save_to,
channels=[
[
'train_cost_with_regularization',
'train_cost_without_regularization',
'train_l2_regularization'
],
['train_error_rate'],
['train_total_gradient_norm'],
],
every_n_batches=10),
# after_batch=True),
Plot('Test performance for ' + save_to,
channels=[[
'train_error_rate',
'test_error_rate',
]],
after_epoch=True),
Checkpoint(save_to),
ProgressBar(),
Printing()
]
if histogram:
attribution = AttributionExtension(components=train_components,
parameters=cg.parameters,
components_size=output_size,
after_batch=True)
extensions.insert(0, attribution)
if resume:
extensions.append(Load(save_to, True, True))
model = Model(train_cost)
main_loop = MainLoop(algorithm,
cifar10_train_stream,
model=model,
extensions=extensions)
main_loop.run()
if histogram:
save_attributions(attribution, filename=histogram)
with open('execution-log.json', 'w') as outfile:
json.dump(main_loop.log, outfile, cls=NumpyEncoder)
def main(args):
"""Run experiment. """
lr_tag = float_tag(args.learning_rate)
x_dim, train_stream, valid_stream, test_stream = datasets.get_streams(
args.data, args.batch_size)
#------------------------------------------------------------
# Setup model
deterministic_act = Tanh
deterministic_size = 1.
if args.method == 'vae':
sizes_tag = args.layer_spec.replace(",", "-")
layer_sizes = [int(i) for i in args.layer_spec.split(",")]
layer_sizes, z_dim = layer_sizes[:-1], layer_sizes[-1]
name = "%s-%s-%s-lr%s-spl%d-%s" % \
(args.data, args.method, args.name, lr_tag, args.n_samples, sizes_tag)
if args.activation == "tanh":
hidden_act = Tanh()
elif args.activation == "logistic":
hidden_act = Logistic()
elif args.activation == "relu":
hidden_act = Rectifier()
else:
raise "Unknown hidden nonlinearity %s" % args.hidden_act
model = VAE(x_dim=x_dim,
hidden_layers=layer_sizes,
hidden_act=hidden_act,
z_dim=z_dim,
batch_norm=args.batch_normalization)
model.initialize()
elif args.method == 'dvae':
sizes_tag = args.layer_spec.replace(",", "-")
layer_sizes = [int(i) for i in args.layer_spec.split(",")]
layer_sizes, z_dim = layer_sizes[:-1], layer_sizes[-1]
name = "%s-%s-%s-lr%s-spl%d-%s" % \
(args.data, args.method, args.name, lr_tag, args.n_samples, sizes_tag)
if args.activation == "tanh":
hidden_act = Tanh()
elif args.activation == "logistic":
hidden_act = Logistic()
elif args.activation == "relu":
hidden_act = Rectifier()
else:
raise "Unknown hidden nonlinearity %s" % args.hidden_act
model = DVAE(x_dim=x_dim,
hidden_layers=layer_sizes,
hidden_act=hidden_act,
z_dim=z_dim,
batch_norm=args.batch_normalization)
model.initialize()
elif args.method == 'rws':
sizes_tag = args.layer_spec.replace(",", "-")
qbase = "" if not args.no_qbaseline else "noqb-"
name = "%s-%s-%s-%slr%s-dl%d-spl%d-%s" % \
(args.data, args.method, args.name, qbase, lr_tag, args.deterministic_layers, args.n_samples, sizes_tag)
p_layers, q_layers = create_layers(args.layer_spec, x_dim,
args.deterministic_layers,
deterministic_act,
deterministic_size)
model = ReweightedWakeSleep(
p_layers,
q_layers,
qbaseline=(not args.no_qbaseline),
)
model.initialize()
elif args.method == 'bihm-rws':
sizes_tag = args.layer_spec.replace(",", "-")
name = "%s-%s-%s-lr%s-dl%d-spl%d-%s" % \
(args.data, args.method, args.name, lr_tag, args.deterministic_layers, args.n_samples, sizes_tag)
p_layers, q_layers = create_layers(args.layer_spec, x_dim,
args.deterministic_layers,
deterministic_act,
deterministic_size)
model = BiHM(
p_layers,
q_layers,
l1reg=args.l1reg,
l2reg=args.l2reg,
)
model.initialize()
elif args.method == 'continue':
import cPickle as pickle
from os.path import basename, splitext
with open(args.model_file, 'rb') as f:
m = pickle.load(f)
if isinstance(m, MainLoop):
m = m.model
model = m.get_top_bricks()[0]
while len(model.parents) > 0:
model = model.parents[0]
assert isinstance(model, (BiHM, ReweightedWakeSleep, VAE))
mname, _, _ = basename(args.model_file).rpartition("_model.pkl")
name = "%s-cont-%s-lr%s-spl%s" % (mname, args.name, lr_tag,
args.n_samples)
else:
raise ValueError("Unknown training method '%s'" % args.method)
#------------------------------------------------------------
x = tensor.matrix('features')
#------------------------------------------------------------
# Testset monitoring
train_monitors = []
valid_monitors = []
test_monitors = []
for s in [1, 10, 100, 1000]:
log_p, log_ph = model.log_likelihood(x, s)
log_p = -log_p.mean()
log_ph = -log_ph.mean()
log_p.name = "log_p_%d" % s
log_ph.name = "log_ph_%d" % s
#train_monitors += [log_p, log_ph]
#valid_monitors += [log_p, log_ph]
test_monitors += [log_p, log_ph]
#------------------------------------------------------------
# Z estimation
#for s in [100000]:
# z2 = tensor.exp(model.estimate_log_z2(s)) / s
# z2.name = "z2_%d" % s
#
# valid_monitors += [z2]
# test_monitors += [z2]
#------------------------------------------------------------
# Gradient and training monitoring
if args.method in ['vae', 'dvae']:
log_p_bound, gradients = model.get_gradients(x, args.n_samples)
log_p_bound = -log_p_bound.mean()
log_p_bound.name = "log_p_bound"
cost = log_p_bound
train_monitors += [
log_p_bound,
named(model.kl_term.mean(), 'kl_term'),
named(model.recons_term.mean(), 'recons_term')
]
valid_monitors += [
log_p_bound,
named(model.kl_term.mean(), 'kl_term'),
named(model.recons_term.mean(), 'recons_term')
]
test_monitors += [
log_p_bound,
named(model.kl_term.mean(), 'kl_term'),
named(model.recons_term.mean(), 'recons_term')
]
else:
log_p, log_ph, gradients = model.get_gradients(x, args.n_samples)
log_p = -log_p.mean()
log_ph = -log_ph.mean()
log_p.name = "log_p"
log_ph.name = "log_ph"
cost = log_ph
train_monitors += [log_p, log_ph]
valid_monitors += [log_p, log_ph]
#------------------------------------------------------------
# Detailed monitoring
"""
n_layers = len(p_layers)
log_px, w, log_p, log_q, samples = model.log_likelihood(x, n_samples)
exp_samples = []
for l in xrange(n_layers):
e = (w.dimshuffle(0, 1, 'x')*samples[l]).sum(axis=1)
e.name = "inference_h%d" % l
e.tag.aggregation_scheme = aggregation.TakeLast(e)
exp_samples.append(e)
s1 = samples[1]
sh1 = s1.shape
s1_ = s1.reshape([sh1[0]*sh1[1], sh1[2]])
s0, _ = model.p_layers[0].sample_expected(s1_)
s0 = s0.reshape([sh1[0], sh1[1], s0.shape[1]])
s0 = (w.dimshuffle(0, 1, 'x')*s0).sum(axis=1)
s0.name = "inference_h0^"
s0.tag.aggregation_scheme = aggregation.TakeLast(s0)
exp_samples.append(s0)
# Draw P-samples
p_samples, _, _ = model.sample_p(100)
#weights = model.importance_weights(samples)
#weights = weights / weights.sum()
for i, s in enumerate(p_samples):
s.name = "psamples_h%d" % i
s.tag.aggregation_scheme = aggregation.TakeLast(s)
#
samples = model.sample(100, oversample=100)
for i, s in enumerate(samples):
s.name = "samples_h%d" % i
s.tag.aggregation_scheme = aggregation.TakeLast(s)
"""
cg = ComputationGraph([cost])
#------------------------------------------------------------
if args.step_rule == "momentum":
step_rule = Momentum(args.learning_rate, 0.95)
elif args.step_rule == "rmsprop":
step_rule = RMSProp(args.learning_rate)
elif args.step_rule == "adam":
step_rule = Adam(args.learning_rate)
else:
raise "Unknown step_rule %s" % args.step_rule
#parameters = cg.parameters[:4] + cg.parameters[5:]
parameters = cg.parameters
algorithm = GradientDescent(
cost=cost,
parameters=parameters,
gradients=gradients,
step_rule=CompositeRule([
#StepClipping(25),
step_rule,
#RemoveNotFinite(1.0),
]))
#------------------------------------------------------------
train_monitors += [
aggregation.mean(algorithm.total_gradient_norm),
aggregation.mean(algorithm.total_step_norm)
]
#------------------------------------------------------------
# Live plotting?
plotting_extensions = []
if args.live_plotting:
plotting_extensions = [
PlotManager(
name,
[
Plotter(channels=[[
"valid_%s" % cost.name, "valid_log_p"
], ["train_total_gradient_norm", "train_total_step_norm"]],
titles=[
"validation cost",
"norm of training gradient and step"
]),
DisplayImage(
[
WeightDisplay(model.p_layers[0].mlp.
linear_transformations[0].W,
n_weights=100,
image_shape=(28, 28))
]
#ImageDataStreamDisplay(test_stream, image_shape=(28,28))]
)
])
]
main_loop = MainLoop(
model=Model(cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=[
Timing(),
ProgressBar(),
TrainingDataMonitoring(
train_monitors, prefix="train", after_epoch=True),
DataStreamMonitoring(
valid_monitors, data_stream=valid_stream, prefix="valid"),
DataStreamMonitoring(test_monitors,
data_stream=test_stream,
prefix="test",
after_epoch=False,
after_training=True,
every_n_epochs=10),
#SharedVariableModifier(
# algorithm.step_rule.components[0].learning_rate,
# half_lr_func,
# before_training=False,
# after_epoch=False,
# after_batch=False,
# every_n_epochs=half_lr),
TrackTheBest('valid_%s' % cost.name),
Checkpoint(name + ".pkl", save_separately=['log', 'model']),
FinishIfNoImprovementAfter('valid_%s_best_so_far' % cost.name,
epochs=args.patience),
FinishAfter(after_n_epochs=args.max_epochs),
Printing()
] + plotting_extensions)
main_loop.run()
def test_momentum_broadcastable():
verify_broadcastable_handling(Momentum(0.5))
def initialaze_algorithm(config, save_path, bokeh_name, params, bokeh_server,
bokeh, use_load_ext, load_log, fast_start,
recognizer, data, model, cg, regularized_cg,
cost, train_cost, parameters,
max_norm_rules, observables,
batch_size, batch_cost, weights_entropy,
labels_mask, labels, gradients=None):
primary_observables = observables
secondary_observables = []
validation_observables = []
root_path, extension = os.path.splitext(save_path)
train_conf = config['training']
# Define the training algorithm.
clipping = StepClipping(train_conf['gradient_threshold'])
clipping.threshold.name = "gradient_norm_threshold"
rule_names = train_conf.get('rules', ['momentum'])
core_rules = []
if 'momentum' in rule_names:
logger.info("Using scaling and momentum for training")
core_rules.append(Momentum(train_conf['scale'], train_conf['momentum']))
if 'adadelta' in rule_names:
logger.info("Using AdaDelta for training")
core_rules.append(AdaDelta(train_conf['decay_rate'], train_conf['epsilon']))
if 'adam' in rule_names:
assert len(rule_names) == 1
logger.info("Using Adam for training")
core_rules.append(
Adam(learning_rate=train_conf.get('scale', 0.002),
beta1=train_conf.get('beta1', 0.1),
beta2=train_conf.get('beta2', 0.001),
epsilon=train_conf.get('epsilon', 1e-8),
decay_factor=train_conf.get('decay_rate', (1 - 1e-8))))
burn_in = []
if train_conf.get('burn_in_steps', 0):
burn_in.append(
BurnIn(num_steps=train_conf['burn_in_steps']))
algorithm = GradientDescent(
cost=train_cost,
parameters=parameters.values(),
gradients=gradients,
step_rule=CompositeRule(
[clipping] + core_rules + max_norm_rules +
# Parameters are not changed at all
# when nans are encountered.
[RemoveNotFinite(0.0)] + burn_in),
on_unused_sources='warn')
#theano_func_kwargs={'mode':NanGuardMode(nan_is_error=True)})
logger.debug("Scan Ops in the gradients")
gradient_cg = ComputationGraph(algorithm.gradients.values())
for op in ComputationGraph(gradient_cg).scans:
logger.debug(op)
# More variables for debugging: some of them can be added only
# after the `algorithm` object is created.
secondary_observables += list(regularized_cg.outputs)
if not 'train_cost' in [v.name for v in secondary_observables]:
secondary_observables += [train_cost]
secondary_observables += [
algorithm.total_step_norm, algorithm.total_gradient_norm,
clipping.threshold]
for name, param in parameters.items():
num_elements = numpy.product(param.get_value().shape)
norm = param.norm(2) / num_elements ** 0.5
grad_norm = algorithm.gradients[param].norm(2) / num_elements ** 0.5
step_norm = algorithm.steps[param].norm(2) / num_elements ** 0.5
stats = tensor.stack(norm, grad_norm, step_norm, step_norm / grad_norm)
stats.name = name + '_stats'
secondary_observables.append(stats)
primary_observables += [
train_cost,
algorithm.total_gradient_norm,
algorithm.total_step_norm, clipping.threshold]
validation_observables += [
rename(aggregation.mean(batch_cost, batch_size), cost.name),
rename(aggregation.sum_(batch_size), 'num_utterances')] + weights_entropy
def attach_aggregation_schemes(variables):
# Aggregation specification has to be factored out as a separate
# function as it has to be applied at the very last stage
# separately to training and validation observables.
result = []
for var in variables:
if var.name.startswith('weights_entropy'):
chld_id = recognizer.child_id_from_postfix(var.name)
result.append(rename(aggregation.mean(var, labels_mask[chld_id].sum()),
'weights_entropy_per_label'+
recognizer.children[chld_id].names_postfix))
elif var.name.endswith('_nll'):
chld_id = recognizer.child_id_from_postfix(var.name)
result.append(rename(aggregation.mean(var.sum(),
labels_mask[chld_id].sum()),
var.name+'_per_label'))
else:
result.append(var)
return result
mon_conf = config['monitoring']
# Build main loop.
logger.info("Initialize extensions")
extensions = []
if use_load_ext and params:
extensions.append(Load(params, load_iteration_state=True, load_log=True))
if load_log and params:
extensions.append(LoadLog(params))
extensions += [
Timing(after_batch=True),
CGStatistics(),
#CodeVersion(['lvsr']),
]
extensions.append(TrainingDataMonitoring(
primary_observables, after_batch=True))
average_monitoring = TrainingDataMonitoring(
attach_aggregation_schemes(secondary_observables),
prefix="average", every_n_batches=10)
extensions.append(average_monitoring)
validation = DataStreamMonitoring(
attach_aggregation_schemes(validation_observables),
data.get_stream("valid", shuffle=False, **data_params_valid), prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['validate_every_epochs'],
every_n_batches=mon_conf['validate_every_batches'],
after_training=False)
extensions.append(validation)
additional_patience_notifiers = []
uas = DependencyErrorRate(recognizer.children[0], data,
**config['monitoring']['search'])
las = AuxiliaryErrorRates(uas, name='LAS')
lab = AuxiliaryErrorRates(uas, name='LAB')
per_monitoring = DataStreamMonitoring(
[uas, las, lab], data.get_one_stream("valid", data.langs[0], batches=False, shuffle=False, **data_params_valid)[0],
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['search_every_epochs'],
every_n_batches=mon_conf['search_every_batches'],
after_training=False)
extensions.append(per_monitoring)
track_the_best_uas = TrackTheBest(
per_monitoring.record_name(uas)).set_conditions(
before_first_epoch=True, after_epoch=True)
track_the_best_las = TrackTheBest(
per_monitoring.record_name(las)).set_conditions(
before_first_epoch=True, after_epoch=True)
track_the_best_lab = TrackTheBest(
per_monitoring.record_name(lab)).set_conditions(
before_first_epoch=True, after_epoch=True)
extensions += [track_the_best_uas,
track_the_best_las,
track_the_best_lab,
]
per = uas
track_the_best_per = track_the_best_uas
additional_patience_notifiers = [track_the_best_lab,
track_the_best_las]
track_the_best_cost = TrackTheBest(
validation.record_name(cost)).set_conditions(
before_first_epoch=True, after_epoch=True)
extensions += [track_the_best_cost]
extensions.append(AdaptiveClipping(
algorithm.total_gradient_norm.name,
clipping, train_conf['gradient_threshold'],
decay_rate=0.998, burnin_period=500,
num_stds=train_conf.get('clip_stds', 1.0)))
extensions += [
SwitchOffLengthFilter(
data.length_filter,
after_n_batches=train_conf.get('stop_filtering')),
FinishAfter(after_n_batches=train_conf['num_batches'],
after_n_epochs=train_conf['num_epochs']),
# .add_condition(["after_batch"], _gradient_norm_is_none),
]
main_postfix = recognizer.children[0].names_postfix
channels = [
# Plot 1: training and validation costs
[average_monitoring.record_name(train_cost),
validation.record_name(cost)],
# Plot 2: gradient norm,
[average_monitoring.record_name(algorithm.total_gradient_norm),
average_monitoring.record_name(clipping.threshold)],
# Plot 3: phoneme error rate
[per_monitoring.record_name(per)],
# Plot 4: training and validation mean weight entropy
[average_monitoring._record_name('weights_entropy_per_label'+main_postfix),
validation._record_name('weights_entropy_per_label'+main_postfix)],
# Plot 5: training and validation monotonicity penalty
[average_monitoring._record_name('weights_penalty_per_recording'+main_postfix),
validation._record_name('weights_penalty_per_recording'+main_postfix)]]
if bokeh:
extensions += [
Plot(bokeh_name if bokeh_name
else os.path.basename(save_path),
channels,
every_n_batches=10,
server_url=bokeh_server),]
extensions += [
Checkpoint(save_path,
before_first_epoch=not fast_start, after_epoch=True,
every_n_batches=train_conf.get('save_every_n_batches'),
save_separately=["model", "log"],
use_cpickle=True)
.add_condition(
['after_epoch'],
OnLogRecord(track_the_best_per.notification_name),
(root_path + "_best" + extension,))
.add_condition(
['after_epoch'],
OnLogRecord(track_the_best_cost.notification_name),
(root_path + "_best_ll" + extension,)),
ProgressBar()]
extensions.append(EmbedIPython(use_main_loop_run_caller_env=True))
if train_conf.get('patience'):
patience_conf = train_conf['patience']
if not patience_conf.get('notification_names'):
# setdefault will not work for empty list
patience_conf['notification_names'] = [
track_the_best_per.notification_name,
track_the_best_cost.notification_name] + additional_patience_notifiers
extensions.append(Patience(**patience_conf))
if train_conf.get('min_performance_stops'):
extensions.append(EarlyTermination(
param_name=track_the_best_per.best_name,
min_performance_by_epoch=train_conf['min_performance_stops']))
extensions.append(Printing(every_n_batches=1,
attribute_filter=PrintingFilterList()))
return model, algorithm, data, extensions
def main(port_data, LR, Init1, Init2):
mlp_hiddens = [500]
filter_sizes = [(3, 3), (3, 3), (3, 3)]
feature_maps = [20, 50, 80]
pooling_sizes = [(2, 2), (2, 2), (2, 2)]
save_to = "DvC_parcour.pkl"
image_size = (128, 128)
output_size = 2
learningRate = LR
num_epochs = 10
num_batches = None
if socket.gethostname() == 'tim-X550JX':
host_plot = 'http://*****:*****@ %s' %
('CNN ', datetime.datetime.now(), socket.gethostname()),
channels=[['train_error_rate', 'valid_error_rate'],
['train_total_gradient_norm']],
after_epoch=True,
server_url=host_plot))
model = Model(cost)
main_loop = MainLoop(algorithm,
stream_data_train,
model=model,
extensions=extensions)
main_loop.run()
params = cg.parameters
model = Model([cost])
print "model parameters:"
print model.get_parameter_dict()
if "adagrad" in config:
print "using adagrad"
thisRule=AdaGrad(learning_rate=learning_rate)
elif "adadelta" in config:
print "using adadelta"
thisRule=AdaDelta()
elif "momentum" in config:
print "using momentum"
mWeight = float(config["momentum"])
thisRule=Momentum(learning_rate=learning_rate, momentum=mWeight)
else:
print "using traditional SGD"
thisRule=Scale(learning_rate=learning_rate)
if "gradientClipping" in config:
threshold = float(config["gradientClipping"])
print "using gradient clipping with threshold ", threshold
thisRule=CompositeRule([StepClipping(threshold), thisRule])
#step_rule=CompositeRule([StepClipping(config['step_clipping']),
# eval(config['step_rule'])()])
algorithm = GradientDescent(cost=cost,
parameters=params,
step_rule=thisRule,
def main(n_hiddens=1500,
n_out=200,
noise_std=0.5,
learning_rate=2e-1,
momentum=0.9,
gamma=2.0):
######################
# Model
######################
import theano
from theano.sandbox.rng_mrg import MRG_RandomStreams
import numpy
import theano.tensor as tensor
from lvq.lvq import AaronLVQ, SupervisedNG, initialize_prototypes
from blocks.bricks import Linear, LinearMaxout
from blocks.bricks import Rectifier, Softmax, Logistic
from blocks.bricks.cost import MisclassificationRate
from blocks.config import config
from blocks.initialization import Uniform, Constant
from blocks.model import Model
from lvq.batch_norm import MLP
#from blocks.bricks import MLP
seed = config.default_seed
rng = MRG_RandomStreams(seed)
x = tensor.tensor4('features')
flat_x = tensor.flatten(x, outdim=2)
flat_x_noise = flat_x + rng.normal(size=flat_x.shape, std=noise_std)
y = tensor.imatrix('targets')
flat_y = tensor.flatten(y, outdim=1)
act = Rectifier()
mlp = MLP(dims=[784, n_hiddens, n_hiddens, n_out],
activations=[act, act, None])
mlp.weights_init = Uniform(0.0, 0.001)
mlp.biases_init = Constant(0.0)
mlp.initialize()
train_out = mlp.apply(flat_x_noise)
test_out = mlp.inference(flat_x)
lamb = theano.shared(numpy.float32(10.0))
lvq = SupervisedNG(10,
n_out,
nonlin=True,
gamma=gamma,
lamb=lamb,
name='lvq')
test_loss, test_misclass = lvq.apply(test_out, flat_y, 'test')
loss, misclass = lvq.apply(train_out, flat_y, 'train')
model = Model(loss)
######################
# Data
######################
#from mnist import MNIST
from fuel.datasets import MNIST
from fuel.streams import DataStream
from fuel.schemes import ShuffledScheme, SequentialScheme
from fuel.transformers import ScaleAndShift, ForceFloatX
mnist_train = MNIST([
'train'
]) #MNIST('train', drop_input=False, sources=('features', 'targets'))
mnist_test = MNIST([
'test'
]) # MNIST('test', drop_input=False, sources=('features', 'targets'))
batch_size = 100 #Batch size for training
batch_size_mon = 2000 # Batch size for monitoring and batch normalization
n_batches = int(
numpy.ceil(float(mnist_train.num_examples) / batch_size_mon))
num_protos = 10
ind = range(mnist_train.num_examples)
train_ind = ind[:50000]
val_ind = ind[50000:]
def preprocessing(data_stream):
return ForceFloatX(ScaleAndShift(data_stream,
1 / 255.0,
0.0,
which_sources=('features', )),
which_sources=('features', ))
train_stream_mon = preprocessing(
DataStream(mnist_train,
iteration_scheme=ShuffledScheme(train_ind, batch_size_mon)))
train_stream_bn = preprocessing(
DataStream(mnist_train,
iteration_scheme=ShuffledScheme(train_ind, batch_size_mon)))
train_stream = preprocessing(
DataStream(mnist_train,
iteration_scheme=ShuffledScheme(train_ind, batch_size)))
valid_stream = preprocessing(
DataStream(mnist_train,
iteration_scheme=ShuffledScheme(val_ind, batch_size)))
test_stream = preprocessing(
DataStream(mnist_test,
iteration_scheme=ShuffledScheme(mnist_test.num_examples,
batch_size_mon)))
initialize_prototypes(lvq, x, train_out, train_stream_mon)
######################
# Training
######################
from blocks.main_loop import MainLoop
from blocks.extensions import FinishAfter, Timing, Printing
from blocks.extensions.monitoring import DataStreamMonitoring, TrainingDataMonitoring
from blocks.algorithms import GradientDescent, Momentum, RMSProp, Adam
from blocks.extensions.saveload import Checkpoint
from lvq.extensions import EarlyStopping, LRDecay, MomentumSwitchOff, NCScheduler
from lvq.batch_norm import BatchNormExtension
lr = theano.shared(numpy.float32(1e-3))
step_rule = Momentum(
learning_rate, 0.9
) #Adam(learning_rate=lr) #RMSProp(learning_rate=1e-5, max_scaling=1e4)
num_epochs = 100
earlystop = EarlyStopping('valid_lvq_apply_misclass', 100,
'./exp/alvq.pkl')
main_loop = MainLoop(
model=model,
data_stream=train_stream,
algorithm=GradientDescent(cost=model.outputs[0],
parameters=model.parameters,
step_rule=step_rule),
extensions=[
FinishAfter(after_n_epochs=num_epochs),
BatchNormExtension(model, train_stream_bn, n_batches),
LRDecay(step_rule.learning_rate, [20, 40, 60, 80]),
MomentumSwitchOff(step_rule.momentum, num_epochs - 20),
NCScheduler(lamb, 30., 0, num_epochs),
DataStreamMonitoring(variables=[test_loss, test_misclass],
data_stream=train_stream_mon,
prefix='train'),
DataStreamMonitoring(variables=[test_loss, test_misclass],
data_stream=valid_stream,
prefix='valid'),
DataStreamMonitoring(variables=[test_loss, test_misclass],
data_stream=test_stream,
prefix='test'),
Timing(), earlystop,
Printing(after_epoch=True)
])
main_loop.run()
return main_loop.status.get('best_test_lvq_apply_misclass', None)
algorithm = GradientDescent(cost=cost,
parameters=params,
step_rule=AdaGrad(learning_rate=learning_rate),
on_unused_sources='warn')
elif "adadelta" in config:
print "using adadelta"
algorithm = GradientDescent(cost=cost,
parameters=params,
step_rule=AdaDelta(),
on_unused_sources='warn')
elif "momentum" in config:
print "using momentum"
mWeight = float(config["momentum"])
algorithm = GradientDescent(cost=cost,
parameters=params,
step_rule=Momentum(learning_rate=learning_rate,
momentum=mWeight),
on_unused_sources='warn')
else:
print "using traditional SGD"
algorithm = GradientDescent(cost=cost,
parameters=params,
step_rule=Scale(learning_rate=learning_rate),
on_unused_sources='warn')
extensions = []
if "saveModel" in config:
print "saving model after training"
extensions.append(Checkpoint(path=networkfile, after_n_epochs=n_epochs))
extensions.append(FinishAfter(after_n_epochs=n_epochs))
extensions.append(
F1Extension(layer3=layer3,
y=y,
def main(save_to, cost_name, learning_rate, momentum, num_epochs):
mlp = MLP([None], [784, 10],
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0))
mlp.initialize()
x = tensor.matrix('features')
y = tensor.lmatrix('targets')
scores = mlp.apply(x)
batch_size = y.shape[0]
indices = tensor.arange(y.shape[0])
target_scores = tensor.set_subtensor(
tensor.zeros((batch_size, 10))[indices, y.flatten()], 1)
score_diff = scores - target_scores
# Logistic Regression
if cost_name == 'lr':
cost = Softmax().categorical_cross_entropy(y.flatten(), scores).mean()
# MSE
elif cost_name == 'mse':
cost = (score_diff**2).mean()
# Perceptron
elif cost_name == 'perceptron':
cost = (scores.max(axis=1) - scores[indices, y.flatten()]).mean()
# TLE
elif cost_name == 'minmin':
cost = abs(score_diff[indices, y.flatten()]).mean()
cost += abs(score_diff[indices, scores.argmax(axis=1)]).mean()
# TLEcut
elif cost_name == 'minmin_cut':
# Score of the groundtruth should be greater or equal than its target score
cost = tensor.maximum(0, -score_diff[indices, y.flatten()]).mean()
# Score of the prediction should be less or equal than its actual score
cost += tensor.maximum(0, score_diff[indices,
scores.argmax(axis=1)]).mean()
# TLE2
elif cost_name == 'minmin2':
cost = ((score_diff[tensor.arange(y.shape[0]), y.flatten()])**2).mean()
cost += ((score_diff[tensor.arange(y.shape[0]),
scores.argmax(axis=1)])**2).mean()
# Direct loss minimization
elif cost_name == 'direct':
epsilon = 0.1
cost = (-scores[indices,
(scores + epsilon * target_scores).argmax(axis=1)] +
scores[indices, scores.argmax(axis=1)]).mean()
cost /= epsilon
elif cost_name == 'svm':
cost = (scores[indices, (scores - 1 * target_scores).argmax(axis=1)] -
scores[indices, y.flatten()]).mean()
else:
raise ValueError("Unknown cost " + cost)
error_rate = MisclassificationRate().apply(y.flatten(), scores)
error_rate.name = 'error_rate'
cg = ComputationGraph([cost])
cost.name = 'cost'
mnist_train = MNIST(("train", ))
mnist_test = MNIST(("test", ))
if learning_rate == None:
learning_rate = 0.0001
if momentum == None:
momentum = 0.0
rule = Momentum(learning_rate=learning_rate, momentum=momentum)
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=rule)
extensions = [
Timing(),
FinishAfter(after_n_epochs=num_epochs),
DataStreamMonitoring([cost, error_rate],
Flatten(DataStream.default_stream(
mnist_test,
iteration_scheme=SequentialScheme(
mnist_test.num_examples, 500)),
which_sources=('features', )),
prefix="test"),
# CallbackExtension(
# lambda: rule.learning_rate.set_value(rule.learning_rate.get_value() * 0.9),
# after_epoch=True),
TrainingDataMonitoring([
cost, error_rate,
aggregation.mean(algorithm.total_gradient_norm), rule.learning_rate
],
prefix="train",
after_epoch=True),
Checkpoint(save_to),
Printing()
]
if BLOCKS_EXTRAS_AVAILABLE:
extensions.append(
Plot('MNIST example',
channels=[['test_cost', 'test_error_rate'],
['train_total_gradient_norm']]))
main_loop = MainLoop(algorithm,
Flatten(DataStream.default_stream(
mnist_train,
iteration_scheme=SequentialScheme(
mnist_train.num_examples, 50)),
which_sources=('features', )),
model=Model(cost),
extensions=extensions)
main_loop.run()
df = pandas.DataFrame.from_dict(main_loop.log, orient='index')
res = {
'cost': cost_name,
'learning_rate': learning_rate,
'momentum': momentum,
'train_cost': df.train_cost.iloc[-1],
'test_cost': df.test_cost.iloc[-1],
'best_test_cost': df.test_cost.min(),
'train_error': df.train_error_rate.iloc[-1],
'test_error': df.test_error_rate.iloc[-1],
'best_test_error': df.test_error_rate.min()
}
res = {
k: float(v) if isinstance(v, numpy.ndarray) else v
for k, v in res.items()
}
json.dump(res, sys.stdout)
sys.stdout.flush()
out_soft1 = Flattener(name='Flatt1').apply(conv_sequence1.apply(out8))
predict1 = NDimensionalSoftmax(name='Soft1').apply(out_soft1)
cost = CategoricalCrossEntropy(name='Cross1').apply(y.flatten(), predict1).copy(name='cost')
error = MisclassificationRate().apply(y.flatten(), predict1)
#Little trick to plot the error rate in two different plots (We can't use two time the same data in the plot for a unknow reason)
error_rate = error.copy(name='error_rate')
error_rate2 = error.copy(name='error_rate2')
cg = ComputationGraph([cost, error_rate])
########### GET THE DATA #####################
stream_train = ServerDataStream(('image_features','targets'), False, port=5512, hwm=40)
stream_valid = ServerDataStream(('image_features','targets'), False, port=5513, hwm=40)
########### DEFINE THE ALGORITHM #############
algorithm = GradientDescent(cost=cost, parameters=cg.parameters, step_rule=Momentum(learning_rate=0.01, momentum=0.9))
extensions = [Timing(),
FinishAfter(after_n_epochs=num_epochs),
TrainingDataMonitoring([cost, error_rate,
aggregation.mean(algorithm.total_gradient_norm)],
prefix="train",
after_epoch=True),
DataStreamMonitoring([cost, error_rate, error_rate2], stream_valid, prefix="valid", after_epoch=True),
Checkpoint("catsVsDogs_fire2_new.pkl", after_epoch=True, save_separately=['log']),
ProgressBar(),
Printing()]
#Adding a live plot with the bokeh server
extensions.append(Plot(
'CatsVsDogs_Fire2',
channels=[['train_error_rate', 'valid_error_rate'],
def main():
nclasses = 27
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--seed", type=int, default=1)
parser.add_argument("--length", type=int, default=180)
parser.add_argument("--num-epochs", type=int, default=100)
parser.add_argument("--batch-size", type=int, default=64)
parser.add_argument("--learning-rate", type=float, default=1e-3)
parser.add_argument("--epsilon", type=float, default=1e-5)
parser.add_argument("--num-hidden", type=int, default=1000)
parser.add_argument("--baseline", action="store_true")
parser.add_argument("--initialization",
choices="identity glorot orthogonal uniform".split(),
default="identity")
parser.add_argument("--initial-gamma", type=float, default=1e-1)
parser.add_argument("--initial-beta", type=float, default=0)
parser.add_argument("--cluster", action="store_true")
parser.add_argument("--activation",
choices=list(activations.keys()),
default="tanh")
parser.add_argument("--optimizer",
choices="sgdmomentum adam rmsprop",
default="rmsprop")
parser.add_argument("--continue-from")
parser.add_argument("--evaluate")
parser.add_argument("--dump-hiddens")
args = parser.parse_args()
np.random.seed(args.seed)
blocks.config.config.default_seed = args.seed
if args.continue_from:
from blocks.serialization import load
main_loop = load(args.continue_from)
main_loop.run()
sys.exit(0)
graphs, extensions, updates = construct_graphs(args, nclasses)
### optimization algorithm definition
if args.optimizer == "adam":
optimizer = Adam(learning_rate=args.learning_rate)
elif args.optimizer == "rmsprop":
optimizer = RMSProp(learning_rate=args.learning_rate, decay_rate=0.9)
elif args.optimizer == "sgdmomentum":
optimizer = Momentum(learning_rate=args.learning_rate, momentum=0.99)
step_rule = CompositeRule([
StepClipping(1.),
optimizer,
])
algorithm = GradientDescent(cost=graphs["training"].outputs[0],
parameters=graphs["training"].parameters,
step_rule=step_rule)
algorithm.add_updates(updates["training"])
model = Model(graphs["training"].outputs[0])
extensions = extensions["training"] + extensions["inference"]
# step monitor
step_channels = []
step_channels.extend([
algorithm.steps[param].norm(2).copy(name="step_norm:%s" % name)
for name, param in model.get_parameter_dict().items()
])
step_channels.append(
algorithm.total_step_norm.copy(name="total_step_norm"))
step_channels.append(
algorithm.total_gradient_norm.copy(name="total_gradient_norm"))
step_channels.extend(graphs["training"].outputs)
logger.warning("constructing training data monitor")
extensions.append(
TrainingDataMonitoring(step_channels,
prefix="iteration",
after_batch=True))
# parameter monitor
extensions.append(
DataStreamMonitoring([
param.norm(2).copy(name="parameter.norm:%s" % name)
for name, param in model.get_parameter_dict().items()
],
data_stream=None,
after_epoch=True))
validation_interval = 500
# performance monitor
for situation in "training inference".split():
if situation == "inference" and not args.evaluate:
# save time when we don't need the inference graph
continue
for which_set in "train valid test".split():
logger.warning("constructing %s %s monitor" %
(which_set, situation))
channels = list(graphs[situation].outputs)
extensions.append(
DataStreamMonitoring(channels,
prefix="%s_%s" % (which_set, situation),
every_n_batches=validation_interval,
data_stream=get_stream(
which_set=which_set,
batch_size=args.batch_size,
num_examples=10000,
length=args.length)))
extensions.extend([
TrackTheBest("valid_training_error_rate",
"best_valid_training_error_rate"),
DumpBest("best_valid_training_error_rate", "best.zip"),
FinishAfter(after_n_epochs=args.num_epochs),
#FinishIfNoImprovementAfter("best_valid_error_rate", epochs=50),
Checkpoint("checkpoint.zip",
on_interrupt=False,
every_n_epochs=1,
use_cpickle=True),
DumpLog("log.pkl", after_epoch=True)
])
if not args.cluster:
extensions.append(ProgressBar())
extensions.extend([
Timing(),
Printing(every_n_batches=validation_interval),
PrintingTo("log"),
])
main_loop = MainLoop(data_stream=get_stream(which_set="train",
batch_size=args.batch_size,
length=args.length,
augment=True),
algorithm=algorithm,
extensions=extensions,
model=model)
if args.dump_hiddens:
dump_hiddens(args, main_loop)
return
if args.evaluate:
evaluate(args, main_loop)
return
main_loop.run()
def initialize_all(config, save_path, bokeh_name, params, bokeh_server, bokeh,
test_tag, use_load_ext, load_log, fast_start):
root_path, extension = os.path.splitext(save_path)
data = Data(**config['data'])
train_conf = config['training']
recognizer = create_model(config, data, test_tag)
# Separate attention_params to be handled differently
# when regularization is applied
attention = recognizer.generator.transition.attention
attention_params = Selector(attention).get_parameters().values()
logger.info(
"Initialization schemes for all bricks.\n"
"Works well only in my branch with __repr__ added to all them,\n"
"there is an issue #463 in Blocks to do that properly.")
def show_init_scheme(cur):
result = dict()
for attr in dir(cur):
if attr.endswith('_init'):
result[attr] = getattr(cur, attr)
for child in cur.children:
result[child.name] = show_init_scheme(child)
return result
logger.info(pprint.pformat(show_init_scheme(recognizer)))
prediction, prediction_mask = add_exploration(recognizer, data, train_conf)
#
# Observables:
#
primary_observables = [] # monitored each batch
secondary_observables = [] # monitored every 10 batches
validation_observables = [] # monitored on the validation set
cg = recognizer.get_cost_graph(batch=True,
prediction=prediction,
prediction_mask=prediction_mask)
labels, = VariableFilter(applications=[recognizer.cost], name='labels')(cg)
labels_mask, = VariableFilter(applications=[recognizer.cost],
name='labels_mask')(cg)
gain_matrix = VariableFilter(
theano_name=RewardRegressionEmitter.GAIN_MATRIX)(cg)
if len(gain_matrix):
gain_matrix, = gain_matrix
primary_observables.append(rename(gain_matrix.min(), 'min_gain'))
primary_observables.append(rename(gain_matrix.max(), 'max_gain'))
batch_cost = cg.outputs[0].sum()
batch_size = rename(recognizer.labels.shape[1], "batch_size")
# Assumes constant batch size. `aggregation.mean` is not used because
# of Blocks #514.
cost = batch_cost / batch_size
cost.name = "sequence_total_cost"
logger.info("Cost graph is built")
# Fetch variables useful for debugging.
# It is important not to use any aggregation schemes here,
# as it's currently impossible to spread the effect of
# regularization on their variables, see Blocks #514.
cost_cg = ComputationGraph(cost)
r = recognizer
energies, = VariableFilter(applications=[r.generator.readout.readout],
name="output_0")(cost_cg)
bottom_output = VariableFilter(
# We need name_regex instead of name because LookupTable calls itsoutput output_0
applications=[r.bottom.apply],
name_regex="output")(cost_cg)[-1]
attended, = VariableFilter(applications=[r.generator.transition.apply],
name="attended")(cost_cg)
attended_mask, = VariableFilter(applications=[
r.generator.transition.apply
],
name="attended_mask")(cost_cg)
weights, = VariableFilter(applications=[r.generator.evaluate],
name="weights")(cost_cg)
from blocks.roles import AUXILIARY
l2_cost, = VariableFilter(roles=[AUXILIARY],
theano_name='l2_cost_aux')(cost_cg)
cost_forward, = VariableFilter(roles=[AUXILIARY],
theano_name='costs_forward_aux')(cost_cg)
max_recording_length = rename(bottom_output.shape[0],
"max_recording_length")
# To exclude subsampling related bugs
max_attended_mask_length = rename(attended_mask.shape[0],
"max_attended_mask_length")
max_attended_length = rename(attended.shape[0], "max_attended_length")
max_num_phonemes = rename(labels.shape[0], "max_num_phonemes")
min_energy = rename(energies.min(), "min_energy")
max_energy = rename(energies.max(), "max_energy")
mean_attended = rename(abs(attended).mean(), "mean_attended")
mean_bottom_output = rename(
abs(bottom_output).mean(), "mean_bottom_output")
weights_penalty = rename(monotonicity_penalty(weights, labels_mask),
"weights_penalty")
weights_entropy = rename(entropy(weights, labels_mask), "weights_entropy")
mask_density = rename(labels_mask.mean(), "mask_density")
cg = ComputationGraph([
cost, weights_penalty, weights_entropy, min_energy, max_energy,
mean_attended, mean_bottom_output, batch_size, max_num_phonemes,
mask_density
])
# Regularization. It is applied explicitly to all variables
# of interest, it could not be applied to the cost only as it
# would not have effect on auxiliary variables, see Blocks #514.
reg_config = config.get('regularization', dict())
regularized_cg = cg
if reg_config.get('dropout'):
logger.info('apply dropout')
regularized_cg = apply_dropout(cg, [bottom_output], 0.5)
if reg_config.get('noise'):
logger.info('apply noise')
noise_subjects = [
p for p in cg.parameters if p not in attention_params
]
regularized_cg = apply_noise(cg, noise_subjects, reg_config['noise'])
train_cost = regularized_cg.outputs[0]
if reg_config.get("penalty_coof", .0) > 0:
# big warning!!!
# here we assume that:
# regularized_weights_penalty = regularized_cg.outputs[1]
train_cost = (train_cost + reg_config.get("penalty_coof", .0) *
regularized_cg.outputs[1] / batch_size)
if reg_config.get("decay", .0) > 0:
train_cost = (
train_cost + reg_config.get("decay", .0) *
l2_norm(VariableFilter(roles=[WEIGHT])(cg.parameters))**2)
train_cost = rename(train_cost, 'train_cost')
gradients = None
if reg_config.get('adaptive_noise'):
logger.info('apply adaptive noise')
if ((reg_config.get("penalty_coof", .0) > 0)
or (reg_config.get("decay", .0) > 0)):
logger.error('using adaptive noise with alignment weight panalty '
'or weight decay is probably stupid')
train_cost, regularized_cg, gradients, noise_brick = apply_adaptive_noise(
cg,
cg.outputs[0],
variables=cg.parameters,
num_examples=data.get_dataset('train').num_examples,
parameters=Model(
regularized_cg.outputs[0]).get_parameter_dict().values(),
**reg_config.get('adaptive_noise'))
train_cost.name = 'train_cost'
adapt_noise_cg = ComputationGraph(train_cost)
model_prior_mean = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_prior_mean')(adapt_noise_cg)[0],
'model_prior_mean')
model_cost = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_cost')(adapt_noise_cg)[0], 'model_cost')
model_prior_variance = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_prior_variance')(adapt_noise_cg)[0],
'model_prior_variance')
regularized_cg = ComputationGraph(
[train_cost, model_cost] + regularized_cg.outputs +
[model_prior_mean, model_prior_variance])
primary_observables += [
regularized_cg.outputs[1], # model cost
regularized_cg.outputs[2], # task cost
regularized_cg.outputs[-2], # model prior mean
regularized_cg.outputs[-1]
] # model prior variance
model = Model(train_cost)
if params:
logger.info("Load parameters from " + params)
# please note: we cannot use recognizer.load_params
# as it builds a new computation graph that dies not have
# shapred variables added by adaptive weight noise
with open(params, 'r') as src:
param_values = load_parameters(src)
model.set_parameter_values(param_values)
parameters = model.get_parameter_dict()
logger.info("Parameters:\n" +
pprint.pformat([(key, parameters[key].get_value().shape)
for key in sorted(parameters.keys())],
width=120))
# Define the training algorithm.
clipping = StepClipping(train_conf['gradient_threshold'])
clipping.threshold.name = "gradient_norm_threshold"
rule_names = train_conf.get('rules', ['momentum'])
core_rules = []
if 'momentum' in rule_names:
logger.info("Using scaling and momentum for training")
core_rules.append(Momentum(train_conf['scale'],
train_conf['momentum']))
if 'adadelta' in rule_names:
logger.info("Using AdaDelta for training")
core_rules.append(
AdaDelta(train_conf['decay_rate'], train_conf['epsilon']))
max_norm_rules = []
if reg_config.get('max_norm', False) > 0:
logger.info("Apply MaxNorm")
maxnorm_subjects = VariableFilter(roles=[WEIGHT])(cg.parameters)
if reg_config.get('max_norm_exclude_lookup', False):
maxnorm_subjects = [
v for v in maxnorm_subjects
if not isinstance(get_brick(v), LookupTable)
]
logger.info("Parameters covered by MaxNorm:\n" + pprint.pformat(
[name for name, p in parameters.items() if p in maxnorm_subjects]))
logger.info("Parameters NOT covered by MaxNorm:\n" + pprint.pformat([
name for name, p in parameters.items() if not p in maxnorm_subjects
]))
max_norm_rules = [
Restrict(VariableClipping(reg_config['max_norm'], axis=0),
maxnorm_subjects)
]
burn_in = []
if train_conf.get('burn_in_steps', 0):
burn_in.append(BurnIn(num_steps=train_conf['burn_in_steps']))
algorithm = GradientDescent(
cost=train_cost,
parameters=parameters.values(),
gradients=gradients,
step_rule=CompositeRule(
[clipping] + core_rules + max_norm_rules +
# Parameters are not changed at all
# when nans are encountered.
[RemoveNotFinite(0.0)] + burn_in),
on_unused_sources='warn')
logger.debug("Scan Ops in the gradients")
gradient_cg = ComputationGraph(algorithm.gradients.values())
for op in ComputationGraph(gradient_cg).scans:
logger.debug(op)
# More variables for debugging: some of them can be added only
# after the `algorithm` object is created.
secondary_observables += list(regularized_cg.outputs)
if not 'train_cost' in [v.name for v in secondary_observables]:
secondary_observables += [train_cost]
secondary_observables += [
algorithm.total_step_norm, algorithm.total_gradient_norm,
clipping.threshold
]
for name, param in parameters.items():
num_elements = numpy.product(param.get_value().shape)
norm = param.norm(2) / num_elements**0.5
grad_norm = algorithm.gradients[param].norm(2) / num_elements**0.5
step_norm = algorithm.steps[param].norm(2) / num_elements**0.5
stats = tensor.stack(norm, grad_norm, step_norm, step_norm / grad_norm)
stats.name = name + '_stats'
secondary_observables.append(stats)
primary_observables += [
train_cost, algorithm.total_gradient_norm, algorithm.total_step_norm,
clipping.threshold, max_recording_length, max_attended_length,
max_attended_mask_length
]
validation_observables += [
rename(aggregation.mean(batch_cost, batch_size), cost.name),
rename(aggregation.sum_(batch_size), 'num_utterances'),
weights_entropy, weights_penalty
]
def attach_aggregation_schemes(variables):
# Aggregation specification has to be factored out as a separate
# function as it has to be applied at the very last stage
# separately to training and validation observables.
result = []
for var in variables:
if var.name == 'weights_penalty':
result.append(
rename(aggregation.mean(var, batch_size),
'weights_penalty_per_recording'))
elif var.name == 'weights_entropy':
result.append(
rename(aggregation.mean(var, labels_mask.sum()),
'weights_entropy_per_label'))
else:
result.append(var)
return result
mon_conf = config['monitoring']
# Build main loop.
logger.info("Initialize extensions")
extensions = []
if use_load_ext and params:
extensions.append(
Load(params, load_iteration_state=True, load_log=True))
if load_log and params:
extensions.append(LoadLog(params))
extensions += [
Timing(after_batch=True),
CGStatistics(),
#CodeVersion(['lvsr']),
]
extensions.append(
TrainingDataMonitoring(primary_observables + [l2_cost, cost_forward],
after_batch=True))
average_monitoring = TrainingDataMonitoring(
attach_aggregation_schemes(secondary_observables),
prefix="average",
every_n_batches=10)
extensions.append(average_monitoring)
validation = DataStreamMonitoring(
attach_aggregation_schemes(validation_observables +
[l2_cost, cost_forward]),
data.get_stream("valid", shuffle=False),
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['validate_every_epochs'],
every_n_batches=mon_conf['validate_every_batches'],
after_training=False)
extensions.append(validation)
per = PhonemeErrorRate(recognizer, data, **config['monitoring']['search'])
per_monitoring = DataStreamMonitoring(
[per],
data.get_stream("valid", batches=False, shuffle=False),
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['search_every_epochs'],
every_n_batches=mon_conf['search_every_batches'],
after_training=False)
extensions.append(per_monitoring)
track_the_best_per = TrackTheBest(
per_monitoring.record_name(per)).set_conditions(
before_first_epoch=True, after_epoch=True)
track_the_best_cost = TrackTheBest(
validation.record_name(cost)).set_conditions(before_first_epoch=True,
after_epoch=True)
extensions += [track_the_best_cost, track_the_best_per]
extensions.append(
AdaptiveClipping(algorithm.total_gradient_norm.name,
clipping,
train_conf['gradient_threshold'],
decay_rate=0.998,
burnin_period=500))
extensions += [
SwitchOffLengthFilter(
data.length_filter,
after_n_batches=train_conf.get('stop_filtering')),
FinishAfter(after_n_batches=train_conf.get('num_batches'),
after_n_epochs=train_conf.get('num_epochs')).add_condition(
["after_batch"], _gradient_norm_is_none),
]
channels = [
# Plot 1: training and validation costs
[
average_monitoring.record_name(train_cost),
validation.record_name(cost)
],
# Plot 2: gradient norm,
[
average_monitoring.record_name(algorithm.total_gradient_norm),
average_monitoring.record_name(clipping.threshold)
],
# Plot 3: phoneme error rate
[per_monitoring.record_name(per)],
# Plot 4: training and validation mean weight entropy
[
average_monitoring._record_name('weights_entropy_per_label'),
validation._record_name('weights_entropy_per_label')
],
# Plot 5: training and validation monotonicity penalty
[
average_monitoring._record_name('weights_penalty_per_recording'),
validation._record_name('weights_penalty_per_recording')
]
]
if bokeh:
extensions += [
Plot(bokeh_name if bokeh_name else os.path.basename(save_path),
channels,
every_n_batches=10,
server_url=bokeh_server),
]
extensions += [
Checkpoint(save_path,
before_first_epoch=not fast_start,
after_epoch=True,
every_n_batches=train_conf.get('save_every_n_batches'),
save_separately=["model", "log"],
use_cpickle=True).add_condition(
['after_epoch'],
OnLogRecord(track_the_best_per.notification_name),
(root_path + "_best" + extension, )).add_condition(
['after_epoch'],
OnLogRecord(track_the_best_cost.notification_name),
(root_path + "_best_ll" + extension, )),
ProgressBar()
]
extensions.append(EmbedIPython(use_main_loop_run_caller_env=True))
if config['net']['criterion']['name'].startswith('mse'):
extensions.append(
LogInputsGains(labels, cg, recognizer.generator.readout.emitter,
data))
if train_conf.get('patience'):
patience_conf = train_conf['patience']
if not patience_conf.get('notification_names'):
# setdefault will not work for empty list
patience_conf['notification_names'] = [
track_the_best_per.notification_name,
track_the_best_cost.notification_name
]
extensions.append(Patience(**patience_conf))
extensions.append(
Printing(every_n_batches=1, attribute_filter=PrintingFilterList()))
return model, algorithm, data, extensions
def main(save_to, num_epochs,
weight_decay=0.0001, noise_pressure=0, subset=None, num_batches=None,
batch_size=None, histogram=None, resume=False):
output_size = 10
prior_noise_level = -10
noise_step_rule = Scale(1e-6)
noise_rate = theano.shared(numpy.asarray(1e-5, dtype=theano.config.floatX))
convnet = create_res_net(out_noise=True, tied_noise=True, tied_sigma=True,
noise_rate=noise_rate,
prior_noise_level=prior_noise_level)
x = tensor.tensor4('features')
y = tensor.lmatrix('targets')
# Normalize input and apply the convnet
test_probs = convnet.apply(x)
test_cost = (CategoricalCrossEntropy().apply(y.flatten(), test_probs)
.copy(name='cost'))
test_error_rate = (MisclassificationRate().apply(y.flatten(), test_probs)
.copy(name='error_rate'))
test_confusion = (ConfusionMatrix().apply(y.flatten(), test_probs)
.copy(name='confusion'))
test_confusion.tag.aggregation_scheme = Sum(test_confusion)
test_cg = ComputationGraph([test_cost, test_error_rate])
# Apply dropout to all layer outputs except final softmax
# dropout_vars = VariableFilter(
# roles=[OUTPUT], bricks=[Convolutional],
# theano_name_regex="^conv_[25]_apply_output$")(test_cg.variables)
# drop_cg = apply_dropout(test_cg, dropout_vars, 0.5)
# Apply 0.2 dropout to the pre-averaging layer
# dropout_vars_2 = VariableFilter(
# roles=[OUTPUT], bricks=[Convolutional],
# theano_name_regex="^conv_8_apply_output$")(test_cg.variables)
# train_cg = apply_dropout(test_cg, dropout_vars_2, 0.2)
# Apply 0.2 dropout to the input, as in the paper
# train_cg = apply_dropout(test_cg, [x], 0.2)
# train_cg = drop_cg
# train_cg = apply_batch_normalization(test_cg)
# train_cost, train_error_rate, train_components = train_cg.outputs
with batch_normalization(convnet):
with training_noise(convnet):
train_probs = convnet.apply(x)
train_cost = (CategoricalCrossEntropy().apply(y.flatten(), train_probs)
.copy(name='cost'))
train_components = (ComponentwiseCrossEntropy().apply(y.flatten(),
train_probs).copy(name='components'))
train_error_rate = (MisclassificationRate().apply(y.flatten(),
train_probs).copy(name='error_rate'))
train_cg = ComputationGraph([train_cost,
train_error_rate, train_components])
population_updates = get_batch_normalization_updates(train_cg)
bn_alpha = 0.9
extra_updates = [(p, p * bn_alpha + m * (1 - bn_alpha))
for p, m in population_updates]
# for annealing
nit_penalty = theano.shared(numpy.asarray(noise_pressure, dtype=theano.config.floatX))
nit_penalty.name = 'nit_penalty'
# Compute noise rates for training graph
train_logsigma = VariableFilter(roles=[LOG_SIGMA])(train_cg.variables)
train_mean_log_sigma = tensor.concatenate([n.flatten() for n in train_logsigma]).mean()
train_mean_log_sigma.name = 'mean_log_sigma'
train_nits = VariableFilter(roles=[NITS])(train_cg.auxiliary_variables)
train_nit_rate = tensor.concatenate([n.flatten() for n in train_nits]).mean()
train_nit_rate.name = 'nit_rate'
train_nit_regularization = nit_penalty * train_nit_rate
train_nit_regularization.name = 'nit_regularization'
# Apply regularization to the cost
trainable_parameters = VariableFilter(roles=[WEIGHT, BIAS])(
train_cg.parameters)
mask_parameters = [p for p in trainable_parameters
if get_brick(p).name == 'mask']
noise_parameters = VariableFilter(roles=[NOISE])(train_cg.parameters)
biases = VariableFilter(roles=[BIAS])(train_cg.parameters)
weights = VariableFilter(roles=[WEIGHT])(train_cg.variables)
nonmask_weights = [p for p in weights if get_brick(p).name != 'mask']
l2_norm = sum([(W ** 2).sum() for W in nonmask_weights])
l2_norm.name = 'l2_norm'
l2_regularization = weight_decay * l2_norm
l2_regularization.name = 'l2_regularization'
# testversion
test_cost = test_cost + l2_regularization
test_cost.name = 'cost_with_regularization'
# Training version of cost
train_cost_without_regularization = train_cost
train_cost_without_regularization.name = 'cost_without_regularization'
train_cost = train_cost + l2_regularization + train_nit_regularization
train_cost.name = 'cost_with_regularization'
cifar10_train = CIFAR10(("train",))
cifar10_train_stream = RandomPadCropFlip(
NormalizeBatchLevels(DataStream.default_stream(
cifar10_train, iteration_scheme=ShuffledScheme(
cifar10_train.num_examples, batch_size)),
which_sources=('features',)),
(32, 32), pad=4, which_sources=('features',))
test_batch_size = 128
cifar10_test = CIFAR10(("test",))
cifar10_test_stream = NormalizeBatchLevels(DataStream.default_stream(
cifar10_test,
iteration_scheme=ShuffledScheme(
cifar10_test.num_examples, test_batch_size)),
which_sources=('features',))
momentum = Momentum(0.01, 0.9)
# Create a step rule that doubles the learning rate of biases, like Caffe.
# scale_bias = Restrict(Scale(2), biases)
# step_rule = CompositeRule([scale_bias, momentum])
# Create a step rule that reduces the learning rate of noise
scale_mask = Restrict(noise_step_rule, mask_parameters)
step_rule = CompositeRule([scale_mask, momentum])
# from theano.compile.nanguardmode import NanGuardMode
# Train with simple SGD
algorithm = GradientDescent(
cost=train_cost, parameters=trainable_parameters,
step_rule=step_rule)
algorithm.add_updates(extra_updates)
#,
# theano_func_kwargs={
# 'mode': NanGuardMode(
# nan_is_error=True, inf_is_error=True, big_is_error=True)})
exp_name = save_to.replace('.%d', '')
# `Timing` extension reports time for reading data, aggregating a batch
# and monitoring;
# `ProgressBar` displays a nice progress bar during training.
extensions = [Timing(),
FinishAfter(after_n_epochs=num_epochs,
after_n_batches=num_batches),
EpochSchedule(momentum.learning_rate, [
(0, 0.01), # Warm up with 0.01 learning rate
(50, 0.1), # Then go back to 0.1
(100, 0.01),
(150, 0.001)
# (83, 0.01), # Follow the schedule in the paper
# (125, 0.001)
]),
EpochSchedule(noise_step_rule.learning_rate, [
(0, 1e-2),
(2, 1e-1),
(4, 1)
# (0, 1e-6),
# (2, 1e-5),
# (4, 1e-4)
]),
EpochSchedule(noise_rate, [
(0, 1e-2),
(2, 1e-1),
(4, 1)
# (0, 1e-6),
# (2, 1e-5),
# (4, 1e-4),
# (6, 3e-4),
# (8, 1e-3), # Causes nit rate to jump
# (10, 3e-3),
# (12, 1e-2),
# (15, 3e-2),
# (19, 1e-1),
# (24, 3e-1),
# (30, 1)
]),
NoiseExtension(
noise_parameters=noise_parameters),
NoisyDataStreamMonitoring(
[test_cost, test_error_rate, test_confusion],
cifar10_test_stream,
noise_parameters=noise_parameters,
prefix="test"),
TrainingDataMonitoring(
[train_cost, train_error_rate, train_nit_rate,
train_cost_without_regularization,
l2_regularization,
train_nit_regularization,
momentum.learning_rate,
train_mean_log_sigma,
aggregation.mean(algorithm.total_gradient_norm)],
prefix="train",
every_n_batches=17),
# after_epoch=True),
Plot('Training performance for ' + exp_name,
channels=[
['train_cost_with_regularization',
'train_cost_without_regularization',
'train_nit_regularization',
'train_l2_regularization'],
['train_error_rate'],
['train_total_gradient_norm'],
['train_mean_log_sigma'],
],
every_n_batches=17),
Plot('Test performance for ' + exp_name,
channels=[[
'train_error_rate',
'test_error_rate',
]],
after_epoch=True),
EpochCheckpoint(save_to, use_cpickle=True, after_epoch=True),
ProgressBar(),
Printing()]
if histogram:
attribution = AttributionExtension(
components=train_components,
parameters=cg.parameters,
components_size=output_size,
after_batch=True)
extensions.insert(0, attribution)
if resume:
extensions.append(Load(exp_name, True, True))
model = Model(train_cost)
main_loop = MainLoop(
algorithm,
cifar10_train_stream,
model=model,
extensions=extensions)
main_loop.run()
if histogram:
save_attributions(attribution, filename=histogram)
with open('execution-log.json', 'w') as outfile:
json.dump(main_loop.log, outfile, cls=NumpyEncoder)
def run(model_name, port_train, port_valid):
running_on_laptop = socket.gethostname() == 'yop'
X = tensor.tensor4('image_features', dtype='float32')
T = tensor.matrix('targets', dtype='float32')
image_border_size = (100, 100)
if running_on_laptop:
host_plot = 'http://*****:*****@ %s' %
(model_name, datetime.datetime.now(), socket.gethostname()),
channels=[['loss'], ['error', 'valid_error']],
after_epoch=True,
server_url=host_plot),
Printing(),
Checkpoint('/tmp/train_bn2')
]
main_loop = MainLoop(data_stream=train_stream,
algorithm=algorithm,
extensions=extensions,
model=model)
main_loop.run()
from blocks.serialization import load
main_loop = load(args.continue_from)
main_loop.run()
sys.exit(0)
graphs, extensions, updates = construct_graphs(args, nclasses)
### optimization algorithm definition
if args.optimizer == "adam":
optimizer = Adam(learning_rate=args.learning_rate)
# zzzz
optimizer.learning_rate = theano.shared(np.asarray(optimizer.learning_rate, dtype=theano.config.floatX))
elif args.optimizer == "rmsprop":
optimizer = RMSProp(learning_rate=args.learning_rate, decay_rate=0.9)
elif args.optimizer == "sgdmomentum":
optimizer = Momentum(learning_rate=args.learning_rate, momentum=0.99)
step_rule = CompositeRule([
StepClipping(1.),
optimizer,
])
algorithm = GradientDescent(cost=graphs["training"].outputs[0],
parameters=graphs["training"].parameters,
step_rule=step_rule)
algorithm.add_updates(updates["training"])
model = Model(graphs["training"].outputs[0])
extensions = extensions["training"] + extensions["inference"]
extensions.append(SharedVariableModifier(
optimizer.learning_rate,
functools.partial(learning_rate_decayer, args.learning_rate_decay)))
batch_size_mon)))
######################
# Training
######################
from blocks.main_loop import MainLoop
from blocks.extensions import FinishAfter, Timing, Printing
from blocks.extensions.monitoring import DataStreamMonitoring
from blocks.algorithms import GradientDescent, Momentum, Adam
from blocks.extensions.saveload import Checkpoint
from lvq.batch_norm import BatchNormExtension
from lvq.extensions import EarlyStopping, LRDecay, MomentumSwitchOff
learning_rate = theano.shared(numpy.float32(1e-4))
step_rule = Momentum(5e-2, 0.9)
main_loop = MainLoop(
model=model,
data_stream=train_stream,
algorithm=GradientDescent(cost=model.outputs[0],
parameters=model.parameters,
step_rule=step_rule),
extensions=[
FinishAfter(after_n_epochs=100),
BatchNormExtension(model, train_stream_bn, n_batches),
LRDecay(step_rule.learning_rate, [20, 40, 60, 80]),
#MomentumSwitchOff(step_rule.momentum, 140),
DataStreamMonitoring(variables=[test_loss, test_misclass],
data_stream=train_stream_mon,
prefix='train'),
n_begin_end_pts = 5 # how many points we consider at the beginning and end of the known trajectory
with open(os.path.join(data.path, 'arrival-clusters.pkl')) as f:
tgtcls = cPickle.load(f)
dim_embeddings = [
('origin_call', data.origin_call_train_size, 10),
('origin_stand', data.stands_size, 10),
('week_of_year', 52, 10),
('day_of_week', 7, 10),
('qhour_of_day', 24 * 4, 10),
('day_type', 3, 10),
('taxi_id', 448, 10),
]
dim_input = n_begin_end_pts * 2 * 2 + sum(x for (_, _, x) in dim_embeddings)
dim_hidden = [100]
dim_output = tgtcls.shape[0]
embed_weights_init = IsotropicGaussian(0.01)
mlp_weights_init = IsotropicGaussian(0.1)
mlp_biases_init = Constant(0.01)
step_rule = Momentum(learning_rate=0.01, momentum=0.9)
batch_size = 200
valid_set = 'cuts/test_times_0'
max_splits = 100
def main(num_epochs=1000):
x = tensor.matrix('features')
y = tensor.lmatrix('targets')
softmax_regressor = SoftmaxRegressor(input_dim=784, n_classes=10)
probs = softmax_regressor.get_probs(features=x)
params = softmax_regressor.get_params()
weights = softmax_regressor.get_weights()
cost = softmax_regressor.get_cost(probs=probs, targets=y).mean()
cost.name = 'cost'
misclassification = softmax_regressor.get_misclassification(
probs=probs, targets=y).mean()
misclassification.name = 'misclassification'
train_dataset = MNIST('train')
test_dataset = MNIST('test')
algorithm = GradientDescent(cost=cost,
params=params,
step_rule=Momentum(learning_rate=0.1,
momentum=0.1))
train_data_stream = ForceFloatX(
data_stream=DataStream(dataset=train_dataset,
iteration_scheme=ShuffledScheme(
examples=train_dataset.num_examples,
batch_size=100,
)))
test_data_stream = ForceFloatX(
data_stream=DataStream(dataset=test_dataset,
iteration_scheme=SequentialScheme(
examples=test_dataset.num_examples,
batch_size=1000,
)))
model = Model(cost)
extensions = []
extensions.append(Timing())
extensions.append(FinishAfter(after_n_epochs=num_epochs))
extensions.append(
DataStreamMonitoring([cost, misclassification],
test_data_stream,
prefix='test'))
extensions.append(
TrainingDataMonitoring([cost, misclassification],
prefix='train',
after_epoch=True))
plotters = []
plotters.append(
Plotter(channels=[[
'test_cost', 'test_misclassification', 'train_cost',
'train_misclassification'
]],
titles=['Costs']))
display_train = ImageDataStreamDisplay(
data_stream=copy.deepcopy(train_data_stream),
image_shape=(28, 28, 1),
axes=(0, 1, 'c'),
shift=-0.5,
rescale=2.,
)
weight_display = WeightDisplay(weights=weights,
transpose=(1, 0),
image_shape=(28, 28, 1),
axes=(0, 1, 'c'),
shift=-0.5,
rescale=2.,
grid_shape=(1, 10))
images_displayer = DisplayImage(
image_getters=[display_train, weight_display],
titles=['Training examples', 'Softmax weights'])
plotters.append(images_displayer)
extensions.append(
PlotManager('MNIST softmax examples',
plotters=plotters,
after_epoch=False,
every_n_epochs=10,
after_training=True))
extensions.append(Printing())
main_loop = MainLoop(model=model,
data_stream=train_data_stream,
algorithm=algorithm,
extensions=extensions)
main_loop.run()
