def test_progressbar_iter_per_epoch_indices():
iter_per_epoch = 100
progress_bar = ProgressBar()
main_loop = setup_mainloop(
None, iteration_scheme=SequentialExampleScheme(iter_per_epoch))
progress_bar.main_loop = main_loop
assert progress_bar.get_iter_per_epoch() == iter_per_epoch
def test_progressbar_iter_per_epoch_indices():
iter_per_epoch = 100
progress_bar = ProgressBar()
main_loop = setup_mainloop(
None, iteration_scheme=SequentialExampleScheme(iter_per_epoch))
progress_bar.main_loop = main_loop
assert progress_bar.get_iter_per_epoch() == iter_per_epoch
def test_progressbar_iter_per_epoch_batch_examples():
num_examples = 1000
batch_size = 10
progress_bar = ProgressBar()
main_loop = setup_mainloop(
None, iteration_scheme=ConstantScheme(batch_size, num_examples))
progress_bar.main_loop = main_loop
assert progress_bar.get_iter_per_epoch() == num_examples // batch_size
def test_progressbar_iter_per_epoch_batch_indices():
num_examples = 1000
batch_size = 10
progress_bar = ProgressBar()
main_loop = setup_mainloop(
None, iteration_scheme=SequentialScheme(num_examples, batch_size))
progress_bar.main_loop = main_loop
assert progress_bar.get_iter_per_epoch() == num_examples // batch_size
def test_progressbar_iter_per_epoch_batch_examples():
num_examples = 1000
batch_size = 10
progress_bar = ProgressBar()
main_loop = setup_mainloop(None,
iteration_scheme=ConstantScheme(
batch_size, num_examples))
progress_bar.main_loop = main_loop
assert progress_bar.get_iter_per_epoch() == num_examples // batch_size
def test_progressbar_iter_per_epoch_batch_indices():
num_examples = 1000
batch_size = 10
progress_bar = ProgressBar()
main_loop = setup_mainloop(None,
iteration_scheme=SequentialScheme(
num_examples, batch_size))
progress_bar.main_loop = main_loop
assert progress_bar.get_iter_per_epoch() == num_examples // batch_size
def create_main_loop():
model, bn_model, bn_updates = create_models()
ali, = bn_model.top_bricks
discriminator_loss, generator_loss = bn_model.outputs
step_rule = Adam(learning_rate=LEARNING_RATE, beta1=BETA1)
algorithm = ali_algorithm(discriminator_loss,
ali.discriminator_parameters, step_rule,
generator_loss, ali.generator_parameters,
step_rule)
algorithm.add_updates(bn_updates)
streams = create_gaussian_mixture_data_streams(
batch_size=BATCH_SIZE,
monitoring_batch_size=MONITORING_BATCH_SIZE,
means=MEANS,
variances=VARIANCES,
priors=PRIORS)
main_loop_stream, train_monitor_stream, valid_monitor_stream = streams
bn_monitored_variables = ([
v for v in bn_model.auxiliary_variables if 'norm' not in v.name
] + bn_model.outputs)
monitored_variables = (
[v
for v in model.auxiliary_variables if 'norm' not in v.name] +
model.outputs)
extensions = [
Timing(),
FinishAfter(after_n_epochs=NUM_EPOCHS),
DataStreamMonitoring(bn_monitored_variables,
train_monitor_stream,
prefix="train",
updates=bn_updates),
DataStreamMonitoring(monitored_variables,
valid_monitor_stream,
prefix="valid"),
Checkpoint(os.path.join(self._work_dir, "main_loop.tar"),
after_epoch=True,
after_training=True,
use_cpickle=True),
ProgressBar(),
Printing(),
#ModelLogger(folder=self._work_dir, after_epoch=True),
GraphLogger(num_modes=1,
num_samples=2500,
dimension=2,
r=0,
std=1,
folder=self._work_dir,
after_epoch=True,
after_training=True),
MetricLogger(means=MEANS,
variances=VARIANCES,
folder=self._work_dir,
after_epoch=True)
]
main_loop = MainLoop(model=bn_model,
data_stream=main_loop_stream,
algorithm=algorithm,
extensions=extensions)
return main_loop
def create_main_loop(save_path):
model, bn_model, bn_updates = create_models()
ali, = bn_model.top_bricks
discriminator_loss, generator_loss = bn_model.outputs
step_rule = Adam(learning_rate=LEARNING_RATE, beta1=BETA1)
algorithm = ali_algorithm(discriminator_loss, ali.discriminator_parameters,
step_rule, generator_loss,
ali.generator_parameters, step_rule)
algorithm.add_updates(bn_updates)
streams = create_celeba_data_streams(BATCH_SIZE, MONITORING_BATCH_SIZE)
main_loop_stream, train_monitor_stream, valid_monitor_stream = streams
bn_monitored_variables = (
[v for v in bn_model.auxiliary_variables if 'norm' not in v.name] +
bn_model.outputs)
monitored_variables = (
[v for v in model.auxiliary_variables if 'norm' not in v.name] +
model.outputs)
extensions = [
Timing(),
FinishAfter(after_n_epochs=NUM_EPOCHS),
DataStreamMonitoring(
bn_monitored_variables, train_monitor_stream, prefix="train",
updates=bn_updates),
DataStreamMonitoring(
monitored_variables, valid_monitor_stream, prefix="valid"),
Checkpoint(save_path, after_epoch=True, after_training=True,
use_cpickle=True),
ProgressBar(),
Printing(),
]
main_loop = MainLoop(model=bn_model, data_stream=main_loop_stream,
algorithm=algorithm, extensions=extensions)
return main_loop
def prepare_opti(cost, test, *args):
model = Model(cost)
logger.info("Model created")
algorithm = GradientDescent(cost=cost,
parameters=model.parameters,
step_rule=Adam(learning_rate=0.0015),
on_unused_sources='ignore')
to_monitor = [algorithm.cost]
if args:
to_monitor.extend(args)
extensions = [
FinishAfter(after_n_epochs=nb_epoch),
FinishIfNoImprovementAfter(notification_name='loglikelihood_nat',
epochs=patience),
TrainingDataMonitoring(to_monitor, prefix="train", after_epoch=True),
DataStreamMonitoring(to_monitor, test_stream, prefix="test"),
Printing(),
ProgressBar(),
ApplyMask(before_first_epoch=True, after_batch=True),
Checkpoint(check, every_n_epochs=save_every),
SaveModel(name=path + '/' + 'pixelcnn_{}'.format(dataset),
every_n_epochs=save_every),
GenerateSamples(every_n_epochs=save_every),
#Checkpoint(path+'/'+'exp.log', save_separately=['log'],every_n_epochs=save_every),
]
if resume:
logger.info("Restoring from previous checkpoint")
extensions = [Load(path + '/' + check)]
return model, algorithm, extensions
def prepare_opti(cost, test):
model = Model(cost)
algorithm = GradientDescent(
cost=cost,
parameters=model.parameters,
step_rule=RMSProp(),
on_unused_sources='ignore'
)
extensions = [
FinishAfter(after_n_epochs=nb_epoch),
FinishIfNoImprovementAfter(notification_name='test_cross_entropy', epochs=patience),
TrainingDataMonitoring(
[algorithm.cost],
prefix="train",
after_epoch=True),
DataStreamMonitoring(
[algorithm.cost],
test_stream,
prefix="test"),
Printing(),
ProgressBar(),
#Checkpoint(path, after_epoch=True)
]
if resume:
print "Restoring from previous breakpoint"
extensions.extend([
Load(path)
])
return model, algorithm, extensions
def construct_main_loop(name, task_name, patch_shape, batch_size,
n_spatial_dims, n_patches, n_epochs, learning_rate,
hyperparameters, **kwargs):
name = "%s_%s" % (name, task_name)
hyperparameters["name"] = name
task = get_task(**hyperparameters)
hyperparameters["n_channels"] = task.n_channels
x_uncentered, y = task.get_variables()
x = task.preprocess(x_uncentered)
# this is a theano variable; it may depend on the batch
hyperparameters["image_shape"] = x.shape[-n_spatial_dims:]
ram = construct_model(task=task, **hyperparameters)
ram.initialize()
hs = ram.compute(x, n_patches)
cost = ram.emitter.cost(hs, y, n_patches)
cost.name = "cost"
print "setting up main loop..."
graph = ComputationGraph(cost)
uselessflunky = Model(cost)
algorithm = GradientDescent(cost=cost,
parameters=graph.parameters,
step_rule=Adam(learning_rate=learning_rate))
monitors = construct_monitors(x=x,
x_uncentered=x_uncentered,
y=y,
hs=hs,
cost=cost,
algorithm=algorithm,
task=task,
model=uselessflunky,
ram=ram,
graph=graph,
**hyperparameters)
main_loop = MainLoop(
data_stream=task.get_stream("train"),
algorithm=algorithm,
extensions=(
monitors + [
FinishAfter(after_n_epochs=n_epochs),
DumpMinimum(name + '_best', channel_name='valid_error_rate'),
Dump(name + '_dump', every_n_epochs=10),
#Checkpoint(name+'_checkpoint.pkl', every_n_epochs=10, on_interrupt=False),
ProgressBar(),
Timing(),
Printing(),
PrintingTo(name + "_log")
]),
model=uselessflunky)
return main_loop
def train(cost, error_rate, batch_size=100, num_epochs=150):
# Setting Loggesetr
timestr = time.strftime("%Y_%m_%d_at_%H_%M")
save_path = 'results/memory_' + timestr
log_path = os.path.join(save_path, 'log.txt')
os.makedirs(save_path)
fh = logging.FileHandler(filename=log_path)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
# Training
blocks_model = Model(cost)
all_params = blocks_model.parameters
print "Number of found parameters:" + str(len(all_params))
print all_params
training_algorithm = GradientDescent(cost=cost,
parameters=all_params,
step_rule=Adam(learning_rate=0.001))
# training_algorithm = GradientDescent(
# cost=cost, params=all_params,
# step_rule=Scale(learning_rate=model.default_lr))
monitored_variables = [cost, error_rate]
# the rest is for validation
# train_data_stream, valid_data_stream = get_mnist_streams(
# 50000, batch_size)
train_data_stream, valid_data_stream = get_mnist_video_streams(batch_size)
train_monitoring = TrainingDataMonitoring(variables=monitored_variables,
prefix="train",
after_epoch=True)
valid_monitoring = DataStreamMonitoring(variables=monitored_variables,
data_stream=valid_data_stream,
prefix="valid",
after_epoch=True)
main_loop = MainLoop(
algorithm=training_algorithm,
data_stream=train_data_stream,
model=blocks_model,
extensions=[
train_monitoring, valid_monitoring,
FinishAfter(after_n_epochs=num_epochs),
SaveParams('valid_misclassificationrate_apply_error_rate',
blocks_model, save_path),
SaveLog(save_path, after_epoch=True),
ProgressBar(),
Printing()
])
main_loop.run()
def create_main_loop(save_path):
model, bn_model, bn_updates = create_models()
ali, = bn_model.top_bricks
discriminator_loss, generator_loss = bn_model.outputs
step_rule = Adam(learning_rate=LEARNING_RATE, beta1=BETA1)
algorithm = ali_algorithm(discriminator_loss, ali.discriminator_parameters,
step_rule, generator_loss,
ali.generator_parameters, step_rule)
algorithm.add_updates(bn_updates)
streams = create_cifar10_data_streams(BATCH_SIZE, MONITORING_BATCH_SIZE)
main_loop_stream, train_monitor_stream, valid_monitor_stream = streams
for d in main_loop_stream.get_epoch_iterator(as_dict=True):
print(d.keys)
print(d['features'].shape, d['features'].dtype)
break
main_loop_stream = ShapesDataset(num_examples=600, img_size=32, min_diameter=3, seed=1234).create_stream(batch_size=BATCH_SIZE, is_train=True)
for d in main_loop_stream.get_epoch_iterator(as_dict=True):
print(d.keys)
print(d['features'].shape, d['features'].dtype)
break
train_monitor_stream = ShapesDataset(num_examples=100, img_size=32, min_diameter=3, seed=1234).create_stream(batch_size=BATCH_SIZE, is_train=False)
valid_monitor_stream = ShapesDataset(num_examples=100, img_size=32, min_diameter=3, seed=5678).create_stream(batch_size=BATCH_SIZE, is_train=False)
bn_monitored_variables = (
[v for v in bn_model.auxiliary_variables if 'norm' not in v.name] +
bn_model.outputs)
monitored_variables = (
[v for v in model.auxiliary_variables if 'norm' not in v.name] +
model.outputs)
extensions = [
Timing(),
FinishAfter(after_n_epochs=NUM_EPOCHS),
DataStreamMonitoring(
bn_monitored_variables, train_monitor_stream, prefix="train",
updates=bn_updates),
DataStreamMonitoring(
monitored_variables, valid_monitor_stream, prefix="valid"),
Checkpoint(save_path, after_epoch=True, after_training=True,
use_cpickle=True),
ProgressBar(),
Printing(),
]
main_loop = MainLoop(model=bn_model, data_stream=main_loop_stream,
algorithm=algorithm, extensions=extensions)
return main_loop
def run():
streams = create_celeba_streams(training_batch_size=100,
monitoring_batch_size=500,
include_targets=True)
main_loop_stream = streams[0]
train_monitor_stream = streams[1]
valid_monitor_stream = streams[2]
cg, bn_dropout_cg = create_training_computation_graphs()
# Compute parameter updates for the batch normalization population
# statistics. They are updated following an exponential moving average.
pop_updates = get_batch_normalization_updates(bn_dropout_cg)
decay_rate = 0.05
extra_updates = [(p, m * decay_rate + p * (1 - decay_rate))
for p, m in pop_updates]
# Prepare algorithm
step_rule = Adam()
algorithm = GradientDescent(cost=bn_dropout_cg.outputs[0],
parameters=bn_dropout_cg.parameters,
step_rule=step_rule)
algorithm.add_updates(extra_updates)
# Prepare monitoring
cost = bn_dropout_cg.outputs[0]
cost.name = 'cost'
train_monitoring = DataStreamMonitoring(
[cost], train_monitor_stream, prefix="train",
before_first_epoch=False, after_epoch=False, after_training=True,
updates=extra_updates)
cost, accuracy = cg.outputs
cost.name = 'cost'
accuracy.name = 'accuracy'
monitored_quantities = [cost, accuracy]
valid_monitoring = DataStreamMonitoring(
monitored_quantities, valid_monitor_stream, prefix="valid",
before_first_epoch=False, after_epoch=False, every_n_epochs=5)
# Prepare checkpoint
checkpoint = Checkpoint(
'celeba_classifier.zip', every_n_epochs=5, use_cpickle=True)
extensions = [Timing(), FinishAfter(after_n_epochs=50), train_monitoring,
valid_monitoring, checkpoint, Printing(), ProgressBar()]
main_loop = MainLoop(data_stream=main_loop_stream, algorithm=algorithm,
extensions=extensions)
main_loop.run()
def infer_population(data_stream, model, n_batches):
""" Sets the population parameters for a given model"""
# construct a main loop with algorithm
algorithm = BatchNormAccumulate(model)
main_loop = MainLoop(
algorithm=algorithm,
data_stream=data_stream,
model=model,
extensions=[FinishAfter(after_n_batches=n_batches),
ProgressBar()])
main_loop.run()
parameters = get_batchnorm_parameters(model)
batchnorm_bricks = set([get_brick(p) for p in parameters])
for b in batchnorm_bricks:
b.use_population = True
def prepare_opti(cost, test):
model = Model(cost)
algorithm = GradientDescent(cost=cost,
parameters=model.parameters,
step_rule=Adam(),
on_unused_sources='ignore')
extensions = [
FinishAfter(after_n_epochs=nb_epoch),
FinishIfNoImprovementAfter(notification_name='test_vae_cost',
epochs=patience),
TrainingDataMonitoring([algorithm.cost], after_epoch=True),
DataStreamMonitoring([algorithm.cost], test, prefix="test"),
Printing(),
ProgressBar(),
#SaveModel(name='vae', after_n_epochs=save_every)
]
return model, algorithm, extensions
def build_extensions_list(self):
epochs = self.config['epochs']
save_freq = self.config['save_freq']
original_save = self.config['original_save'] if self.config.has_key('original_save') \
else False
tracked_train = self.model.linksholder.filter_and_broadcast_request(
'train', 'TrackingLink')
tracked_valid = self.model.linksholder.filter_and_broadcast_request(
'valid', 'TrackingLink')
extensions = [
Timing(),
Experiment(self.filemanager.exp_name, self.filemanager.local_path,
self.filemanager.network_path,
self.filemanager.full_dump),
FinishAfter(after_n_epochs=epochs),
TrainingDataMonitoring(tracked_train,
prefix="train",
after_epoch=True),
DataStreamMonitoring(tracked_valid,
self.streams['valid'],
prefix="valid",
after_epoch=True),
#Ipdb(after_batch=True),
SaveExperiment(self.model.parameters,
original_save=original_save,
every_n_epochs=save_freq),
]
extensions += [
#LoadExperiment(
# parameters,
# full_load=False),
ProgressBar(),
Printing(),
]
self.extensions = extensions
def main(save_to, num_epochs,
regularization=0.0003, subset=None, num_batches=None,
histogram=None, resume=False):
batch_size = 500
output_size = 10
convnet = create_lenet_5()
layers = convnet.layers
x = tensor.tensor4('features')
y = tensor.lmatrix('targets')
# Normalize input and apply the convnet
probs = convnet.apply(x)
cost = (CategoricalCrossEntropy().apply(y.flatten(), probs)
.copy(name='cost'))
components = (ComponentwiseCrossEntropy().apply(y.flatten(), probs)
.copy(name='components'))
error_rate = (MisclassificationRate().apply(y.flatten(), probs)
.copy(name='error_rate'))
confusion = (ConfusionMatrix().apply(y.flatten(), probs)
.copy(name='confusion'))
confusion.tag.aggregation_scheme = Sum(confusion)
cg = ComputationGraph([cost, error_rate, components])
# Apply regularization to the cost
weights = VariableFilter(roles=[WEIGHT])(cg.variables)
l2_norm = sum([(W ** 2).sum() for W in weights])
l2_norm.name = 'l2_norm'
cost = cost + regularization * l2_norm
cost.name = 'cost_with_regularization'
if subset:
start = 30000 - subset // 2
mnist_train = MNIST(("train",), subset=slice(start, start+subset))
else:
mnist_train = MNIST(("train",))
mnist_train_stream = DataStream.default_stream(
mnist_train, iteration_scheme=ShuffledScheme(
mnist_train.num_examples, batch_size))
mnist_test = MNIST(("test",))
mnist_test_stream = DataStream.default_stream(
mnist_test,
iteration_scheme=ShuffledScheme(
mnist_test.num_examples, batch_size))
# Train with simple SGD
algorithm = GradientDescent(
cost=cost, parameters=cg.parameters,
step_rule=AdaDelta(decay_rate=0.99))
# `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),
DataStreamMonitoring(
[cost, error_rate, confusion],
mnist_test_stream,
prefix="test"),
TrainingDataMonitoring(
[cost, error_rate, l2_norm,
aggregation.mean(algorithm.total_gradient_norm)],
prefix="train",
after_epoch=True),
Checkpoint(save_to),
ProgressBar(),
Printing()]
if histogram:
attribution = AttributionExtension(
components=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(cost)
main_loop = MainLoop(
algorithm,
mnist_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(save_to, model, train, test, num_epochs, input_size = (150,150), learning_rate=0.01,
batch_size=50, num_batches=None, flatten_stream=False):
"""
save_to : where to save trained model
model : model given in input must be already initialised (works with convnet and mlp)
input_size : the shape of the reshaped image in input (before flattening is applied if flatten_stream is True)
"""
if flatten_stream :
x = tensor.matrix('image_features')
else :
x = tensor.tensor4('image_features')
y = tensor.lmatrix('targets')
#Data augmentation
#insert data augmentation here
#Generating stream
train_stream = DataStream.default_stream(
train,
iteration_scheme=ShuffledScheme(train.num_examples, batch_size)
)
test_stream = DataStream.default_stream(
test,
iteration_scheme=ShuffledScheme(test.num_examples, batch_size)
)
#Reshaping procedure
#Add a crop option in scikitresize so that the image is not deformed
#Resize to desired square shape
train_stream = ScikitResize(train_stream, input_size, which_sources=('image_features',))
test_stream = ScikitResize(test_stream, input_size, which_sources=('image_features',))
#Flattening the stream
if flatten_stream is True:
train_stream = Flatten(train_stream, which_sources=('image_features',))
test_stream = Flatten(test_stream, which_sources=('image_features',))
# Apply input to model
probs = model.apply(x)
#Defining cost and various indices to watch
#print(probs)
#cost = SquaredError().apply(y.flatten(),probs)
cost = CategoricalCrossEntropy().apply(y.flatten(), probs).copy(name='cost')
error_rate = MisclassificationRate().apply(y.flatten(), probs).copy(
name='error_rate')
#Building Computation Graph
cg = ComputationGraph([cost, error_rate])
# Train with simple SGD
algorithm = GradientDescent(
cost=cost, parameters=cg.parameters,
step_rule=Scale(learning_rate=learning_rate))
#Defining extensions
extensions = [Timing(),
FinishAfter(after_n_epochs=num_epochs,
after_n_batches=num_batches),
TrainingDataMonitoring([cost, error_rate,aggregation.mean(algorithm.total_gradient_norm)], prefix="train", every_n_batches=5),
DataStreamMonitoring([cost, error_rate],test_stream,prefix="test", every_n_batches=25),
Checkpoint(save_to),
ProgressBar(),
Printing(every_n_batches=5)]
# `Timing` extension reports time for reading data, aggregating a batch
# and monitoring;
# `ProgressBar` displays a nice progress bar during training.
model = Model(cost)
main_loop = MainLoop(
algorithm,
train_stream,
model=model,
extensions=extensions)
main_loop.run()
def rf_lstm_experiment(data_name, exp_network, in_dim, out_dim, num_layers,
start_neurons, num_neurons, batch_size, num_epochs):
"""LSTM Experiment."""
# load dataset
train_set = IterableDataset(
ds.transform_sequence(data_name, "train", batch_size))
test_set = IterableDataset(
ds.transform_sequence(data_name, "test", batch_size))
stream_train = DataStream(dataset=train_set)
stream_test = DataStream(dataset=test_set)
methods = ['sgd', 'momentum', 'adagrad', 'rmsprop']
for n_layers in xrange(1, num_layers + 1):
for n_neurons in xrange(start_neurons, num_neurons + 5, 5):
for method in methods:
X = T.tensor3("features")
y = T.matrix("targets")
x_to_h = Linear(in_dim,
n_neurons * 4,
name='x_to_h',
weights_init=IsotropicGaussian(),
biases_init=Constant(0.0))
lstm = LSTM(n_neurons,
name='lstm',
weights_init=IsotropicGaussian(),
biases_init=Constant(0.0))
h_to_o = nc.setup_ff_network(n_neurons, out_dim, n_layers - 1,
n_neurons)
X_trans = x_to_h.apply(X)
h, c = lstm.apply(X_trans)
y_hat = h_to_o.apply(h[-1])
cost, cg = nc.create_cg_and_cost(y, y_hat, "none")
lstm.initialize()
x_to_h.initialize()
h_to_o.initialize()
algorithm = nc.setup_algorithms(cost, cg, method, type="RNN")
test_monitor = DataStreamMonitoring(variables=[cost],
data_stream=stream_test,
prefix="test")
train_monitor = TrainingDataMonitoring(variables=[cost],
prefix="train",
after_epoch=True)
main_loop = MainLoop(
algorithm=algorithm,
data_stream=stream_train,
extensions=[
test_monitor, train_monitor,
FinishAfter(after_n_epochs=num_epochs),
Printing(),
ProgressBar()
])
main_loop.run()
# Saving results
exp_id = ds.create_exp_id(exp_network, n_layers, n_neurons,
batch_size, num_epochs, method,
"none")
# prepare related functions
predict = theano.function([X], y_hat)
# prepare related data
train_features, train_targets = ds.get_iter_data(train_set)
test_features, test_targets = ds.get_iter_data(test_set)
# Prediction of result
train_predicted = gen_prediction(predict, train_features)
test_predicted = gen_prediction(predict, test_features)
# Get cost
cost = ds.get_cost_data(test_monitor, train_set.num_examples,
num_epochs)
# logging
ds.save_experiment(train_targets, train_predicted,
test_targets, test_predicted, cost,
exp_network, n_layers, n_neurons,
batch_size, num_epochs, method, "none",
exp_id, "../results/")
def main(save_to):
batch_size = 365
feature_maps = [6, 16]
mlp_hiddens = [120, 84]
conv_sizes = [5, 5]
pool_sizes = [2, 2]
image_size = (28, 28)
output_size = 10
# The above are from LeCun's paper. The blocks example had:
# feature_maps = [20, 50]
# mlp_hiddens = [500]
# Use ReLUs everywhere and softmax for the final prediction
conv_activations = [Rectifier() for _ in feature_maps]
mlp_activations = [Rectifier() for _ in mlp_hiddens] + [Softmax()]
convnet = LeNet(conv_activations, 1, image_size,
filter_sizes=zip(conv_sizes, conv_sizes),
feature_maps=feature_maps,
pooling_sizes=zip(pool_sizes, pool_sizes),
top_mlp_activations=mlp_activations,
top_mlp_dims=mlp_hiddens + [output_size],
border_mode='valid',
weights_init=Uniform(width=.2),
biases_init=Constant(0))
# We push initialization config to set different initialization schemes
# for convolutional layers.
convnet.push_initialization_config()
convnet.layers[0].weights_init = Uniform(width=.2)
convnet.layers[1].weights_init = Uniform(width=.09)
convnet.top_mlp.linear_transformations[0].weights_init = Uniform(width=.08)
convnet.top_mlp.linear_transformations[1].weights_init = Uniform(width=.11)
convnet.initialize()
logging.info("Input dim: {} {} {}".format(
*convnet.children[0].get_dim('input_')))
for i, layer in enumerate(convnet.layers):
if isinstance(layer, Activation):
logging.info("Layer {} ({})".format(
i, layer.__class__.__name__))
else:
logging.info("Layer {} ({}) dim: {} {} {}".format(
i, layer.__class__.__name__, *layer.get_dim('output')))
x = tensor.tensor4('features')
# Normalize input and apply the convnet
probs = convnet.apply(x)
cg = ComputationGraph([probs])
outs = VariableFilter(
roles=[OUTPUT], bricks=[Convolutional, Linear])(cg.variables)
# Create an interior activation model
model = Model([probs] + outs)
# Load it with trained parameters
params = load_parameters(open(save_to, 'rb'))
model.set_parameter_values(params)
algorithm = MaximumActivationSearch(outputs=outs)
# Use the mnist test set, unshuffled
mnist_test = MNIST(("test",), sources=['features'])
mnist_test_stream = DataStream.default_stream(
mnist_test,
iteration_scheme=SequentialScheme(
mnist_test.num_examples, batch_size))
extensions = [Timing(),
FinishAfter(after_n_epochs=1),
DataStreamMonitoring(
[],
mnist_test_stream,
prefix="test"),
Checkpoint("maxact.tar"),
ProgressBar(),
Printing()]
main_loop = MainLoop(
algorithm,
mnist_test_stream,
model=model,
extensions=extensions)
main_loop.run()
examples = mnist_test.get_example_stream()
example = examples.get_data(0)[0]
layers = convnet.layers
for output, record in algorithm.maximum_activations.items():
layer = get_brick(output)
activations, indices, snapshots = (
r.get_value() if r else None for r in record[1:])
filmstrip = Filmstrip(
example.shape[-2:], (indices.shape[1], indices.shape[0]),
background='blue')
if layer in layers:
fieldmap = layerarray_fieldmap(layers[0:layers.index(layer) + 1])
for unit in range(indices.shape[1]):
for index in range(100):
mask = make_mask(example.shape[-2:], fieldmap, numpy.clip(
snapshots[index, unit, :, :], 0, numpy.inf))
imagenum = indices[index, unit, 0]
filmstrip.set_image((unit, index),
examples.get_data(imagenum)[0], mask)
else:
for unit in range(indices.shape[1]):
for index in range(100):
imagenum = indices[index, unit]
filmstrip.set_image((unit, index),
examples.get_data(imagenum)[0])
filmstrip.save(layer.name + '_maxact.jpg')
def construct_main_loop(name, task_name, patch_shape, batch_size,
n_spatial_dims, n_patches, max_epochs, patience_epochs,
learning_rate, gradient_limiter, hyperparameters,
**kwargs):
task = tasks.get_task(**hyperparameters)
hyperparameters["n_channels"] = task.n_channels
extensions = []
# let theta noise decay as training progresses
for key in "location_std scale_std".split():
hyperparameters[key] = theano.shared(hyperparameters[key], name=key)
extensions.append(
util.ExponentialDecay(hyperparameters[key],
hyperparameters["%s_decay" % key],
after_batch=True))
print "constructing graphs..."
graphs, outputs, updates = construct_graphs(task=task, **hyperparameters)
print "setting up main loop..."
from blocks.model import Model
model = Model(outputs["train"]["cost"])
from blocks.algorithms import GradientDescent, CompositeRule, StepClipping, Adam, RMSProp
from extensions import Compressor
if gradient_limiter == "clip":
limiter = StepClipping(1.)
elif gradient_limiter == "compress":
limiter = Compressor()
else:
raise ValueError()
algorithm = GradientDescent(
cost=outputs["train"]["cost"],
parameters=graphs["train"].parameters,
step_rule=CompositeRule([limiter,
Adam(learning_rate=learning_rate)]))
algorithm.add_updates(updates["train"])
extensions.extend(
construct_monitors(algorithm=algorithm,
task=task,
model=model,
graphs=graphs,
outputs=outputs,
updates=updates,
**hyperparameters))
from blocks.extensions import FinishAfter, Printing, ProgressBar, Timing
from blocks.extensions.stopping import FinishIfNoImprovementAfter
from blocks.extensions.training import TrackTheBest
from blocks.extensions.saveload import Checkpoint
from dump import DumpBest, LightCheckpoint, PrintingTo, DumpGraph, DumpLog
extensions.extend([
TrackTheBest("valid_error_rate", "best_valid_error_rate"),
FinishIfNoImprovementAfter("best_valid_error_rate",
epochs=patience_epochs),
FinishAfter(after_n_epochs=max_epochs),
DumpBest("best_valid_error_rate", name + "_best.zip"),
Checkpoint(hyperparameters["checkpoint_save_path"],
on_interrupt=False,
every_n_epochs=10,
use_cpickle=True),
DumpLog("log.pkl", after_epoch=True),
ProgressBar(),
Timing(),
Printing(),
PrintingTo(name + "_log"),
DumpGraph(name + "_grad_graph")
])
from blocks.main_loop import MainLoop
main_loop = MainLoop(data_stream=task.get_stream("train"),
algorithm=algorithm,
extensions=extensions,
model=model)
from tabulate import tabulate
print "parameter sizes:"
print tabulate(
(key, "x".join(map(str,
value.get_value().shape)), value.get_value().size)
for key, value in main_loop.model.get_parameter_dict().items())
return main_loop
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 _train_model(self, train_stream, valid_stream, load_from, save_to,
*args, **kwargs):
# Build model
self.model = self.config.Model(self.config,
self.dataset) # with word2id
cg = Model(self.model.cg_generator)
algorithm = GradientDescent(cost=self.model.cg_generator,
step_rule=self.config.step_rule,
parameters=cg.parameters,
on_unused_sources='ignore')
if plot_avail:
extensions = [
FinishAfter(after_n_epochs=1),
TrainingDataMonitoring(
[v for l in self.model.monitor_train_vars for v in l],
prefix='train',
every_n_batches=self.config.print_freq),
Plot('Training Process',
channels=[
v.name for l in self.model.monitor_train_vars
for v in l
],
after_batch=True)
]
else:
extensions = [
TrainingDataMonitoring(
[v for l in self.model.monitor_train_vars for v in l],
prefix='train',
every_n_batches=self.config.print_freq)
]
saver_loader = self.model_save_loader(load_from=load_from,
save_to=save_to,
model=cg,
dataset=self.dataset)
saver_loader.do_load()
n_batches = numpy.ceil(self.n_samples /
self.config.batch_size).astype('int32')
n_valid_batches = numpy.ceil(n_batches *
self.config.valid_freq).astype('int32')
extensions += [
EvaluatorWithEarlyStop(
coverage=1.,
tolerate_time=self.config.tolerate_time,
variables=[
v for l in self.model.monitor_valid_vars for v in l
],
monitor_variable=self.model.stop_monitor_var,
data_stream=valid_stream,
saver=saver_loader,
prefix='valid',
every_n_batches=n_valid_batches)
]
extensions += [
Printing(every_n_batches=self.config.print_freq, after_epoch=True),
ProgressBar()
]
extensions += [EpochMonitor(1)]
main_loop = MainLoop(model=cg,
data_stream=train_stream,
algorithm=algorithm,
extensions=extensions)
# Run the model!
main_loop.run()
def main(nvis, nhid, encoding_lstm_dim, decoding_lstm_dim, T=1):
x = tensor.matrix('features')
# Construct and initialize model
encoding_mlp = MLP([Tanh()], [None, None])
decoding_mlp = MLP([Tanh()], [None, None])
encoding_lstm = LSTM(dim=encoding_lstm_dim)
decoding_lstm = LSTM(dim=decoding_lstm_dim)
draw = DRAW(nvis=nvis,
nhid=nhid,
T=T,
encoding_mlp=encoding_mlp,
decoding_mlp=decoding_mlp,
encoding_lstm=encoding_lstm,
decoding_lstm=decoding_lstm,
biases_init=Constant(0),
weights_init=Orthogonal())
draw.push_initialization_config()
encoding_lstm.weights_init = IsotropicGaussian(std=0.001)
decoding_lstm.weights_init = IsotropicGaussian(std=0.001)
draw.initialize()
# Compute cost
cost = -draw.log_likelihood_lower_bound(x).mean()
cost.name = 'nll_upper_bound'
model = Model(cost)
# Datasets and data streams
mnist_train = BinarizedMNIST('train')
train_loop_stream = ForceFloatX(
DataStream(dataset=mnist_train,
iteration_scheme=SequentialScheme(mnist_train.num_examples,
100)))
train_monitor_stream = ForceFloatX(
DataStream(dataset=mnist_train,
iteration_scheme=SequentialScheme(mnist_train.num_examples,
500)))
mnist_valid = BinarizedMNIST('valid')
valid_monitor_stream = ForceFloatX(
DataStream(dataset=mnist_valid,
iteration_scheme=SequentialScheme(mnist_valid.num_examples,
500)))
mnist_test = BinarizedMNIST('test')
test_monitor_stream = ForceFloatX(
DataStream(dataset=mnist_test,
iteration_scheme=SequentialScheme(mnist_test.num_examples,
500)))
# Get parameters and monitoring channels
computation_graph = ComputationGraph([cost])
params = VariableFilter(roles=[PARAMETER])(computation_graph.variables)
monitoring_channels = dict([
('avg_' + channel.tag.name, channel.mean())
for channel in VariableFilter(
name='.*term$')(computation_graph.auxiliary_variables)
])
for name, channel in monitoring_channels.items():
channel.name = name
monitored_quantities = monitoring_channels.values() + [cost]
# Training loop
step_rule = RMSProp(learning_rate=1e-3, decay_rate=0.95)
algorithm = GradientDescent(cost=cost, params=params, step_rule=step_rule)
algorithm.add_updates(computation_graph.updates)
main_loop = MainLoop(
model=model,
data_stream=train_loop_stream,
algorithm=algorithm,
extensions=[
Timing(),
SerializeMainLoop('vae.pkl', save_separately=['model']),
FinishAfter(after_n_epochs=200),
DataStreamMonitoring(monitored_quantities,
train_monitor_stream,
prefix="train",
updates=computation_graph.updates),
DataStreamMonitoring(monitored_quantities,
valid_monitor_stream,
prefix="valid",
updates=computation_graph.updates),
DataStreamMonitoring(monitored_quantities,
test_monitor_stream,
prefix="test",
updates=computation_graph.updates),
ProgressBar(),
Printing()
])
main_loop.run()
def main(name, epochs, batch_size, learning_rate, attention, n_iter, enc_dim,
dec_dim, z_dim):
if name is None:
tag = "watt" if attention else "woatt"
name = "%s-t%d-enc%d-dec%d-z%d" % (tag, n_iter, enc_dim, dec_dim,
z_dim)
print("\nRunning experiment %s" % name)
print(" learning rate: %5.3f" % learning_rate)
print(" attention: %s" % attention)
print(" n_iterations: %d" % n_iter)
print(" encoder dimension: %d" % enc_dim)
print(" z dimension: %d" % z_dim)
print(" decoder dimension: %d" % dec_dim)
print()
#------------------------------------------------------------------------
x_dim = 28 * 28
img_height, img_width = (28, 28)
rnninits = {
'weights_init': Orthogonal(),
#'weights_init': IsotropicGaussian(0.001),
'biases_init': Constant(0.),
}
inits = {
'weights_init': Orthogonal(),
#'weights_init': IsotropicGaussian(0.01),
'biases_init': Constant(0.),
}
prior_mu = T.zeros([z_dim])
prior_log_sigma = T.zeros([z_dim])
if attention:
read_N = 4
write_N = 6
read_dim = 2 * read_N**2
reader = AttentionReader(x_dim=x_dim,
dec_dim=dec_dim,
width=img_width,
height=img_height,
N=read_N,
**inits)
writer = AttentionWriter(input_dim=dec_dim,
output_dim=x_dim,
width=img_width,
height=img_height,
N=read_N,
**inits)
else:
read_dim = 2 * x_dim
reader = Reader(x_dim=x_dim, dec_dim=dec_dim, **inits)
writer = Writer(input_dim=dec_dim, output_dim=x_dim, **inits)
encoder = LSTM(dim=enc_dim, name="RNN_enc", **rnninits)
decoder = LSTM(dim=dec_dim, name="RNN_dec", **rnninits)
encoder_mlp = MLP([Tanh()], [(read_dim + dec_dim), 4 * enc_dim],
name="MLP_enc",
**inits)
decoder_mlp = MLP([Tanh()], [z_dim, 4 * dec_dim], name="MLP_dec", **inits)
q_sampler = Qsampler(input_dim=enc_dim, output_dim=z_dim, **inits)
for brick in [
reader, writer, encoder, decoder, encoder_mlp, decoder_mlp,
q_sampler
]:
brick.allocate()
brick.initialize()
#------------------------------------------------------------------------
x = tensor.matrix('features')
# This is one iteration
def one_iteration(c, h_enc, c_enc, z_mean, z_log_sigma, z, h_dec, c_dec,
x):
x_hat = x - T.nnet.sigmoid(c)
r = reader.apply(x, x_hat, h_dec)
i_enc = encoder_mlp.apply(T.concatenate([r, h_dec], axis=1))
h_enc, c_enc = encoder.apply(states=h_enc,
cells=c_enc,
inputs=i_enc,
iterate=False)
z_mean, z_log_sigma, z = q_sampler.apply(h_enc)
i_dec = decoder_mlp.apply(z)
h_dec, c_dec = decoder.apply(states=h_dec,
cells=c_dec,
inputs=i_dec,
iterate=False)
c = c + writer.apply(h_dec)
return c, h_enc, c_enc, z_mean, z_log_sigma, z, h_dec, c_dec
outputs_info = [
T.zeros([batch_size, x_dim]), # c
T.zeros([batch_size, enc_dim]), # h_enc
T.zeros([batch_size, enc_dim]), # c_enc
T.zeros([batch_size, z_dim]), # z_mean
T.zeros([batch_size, z_dim]), # z_log_sigma
T.zeros([batch_size, z_dim]), # z
T.zeros([batch_size, dec_dim]), # h_dec
T.zeros([batch_size, dec_dim]), # c_dec
]
outputs, scan_updates = theano.scan(fn=one_iteration,
sequences=[],
outputs_info=outputs_info,
non_sequences=[x],
n_steps=n_iter)
c, h_enc, c_enc, z_mean, z_log_sigma, z, h_dec, c_dec = outputs
kl_terms = (prior_log_sigma - z_log_sigma + 0.5 *
(tensor.exp(2 * z_log_sigma) +
(z_mean - prior_mu)**2) / tensor.exp(2 * prior_log_sigma) -
0.5).sum(axis=-1)
x_recons = T.nnet.sigmoid(c[-1, :, :])
recons_term = BinaryCrossEntropy().apply(x, x_recons)
recons_term.name = "recons_term"
cost = recons_term + kl_terms.sum(axis=0).mean()
cost.name = "nll_bound"
#------------------------------------------------------------
cg = ComputationGraph([cost])
params = VariableFilter(roles=[PARAMETER])(cg.variables)
algorithm = GradientDescent(
cost=cost,
params=params,
step_rule=CompositeRule([
#StepClipping(3.),
Adam(learning_rate),
])
#step_rule=RMSProp(learning_rate),
#step_rule=Momentum(learning_rate=learning_rate, momentum=0.95)
)
algorithm.add_updates(scan_updates)
#------------------------------------------------------------------------
# Setup monitors
monitors = [cost]
for t in range(n_iter):
kl_term_t = kl_terms[t, :].mean()
kl_term_t.name = "kl_term_%d" % t
x_recons_t = T.nnet.sigmoid(c[t, :, :])
recons_term_t = BinaryCrossEntropy().apply(x, x_recons_t)
recons_term_t = recons_term_t.mean()
recons_term_t.name = "recons_term_%d" % t
monitors += [kl_term_t, recons_term_t]
train_monitors = monitors[:]
train_monitors += [aggregation.mean(algorithm.total_gradient_norm)]
train_monitors += [aggregation.mean(algorithm.total_step_norm)]
# Live plotting...
plot_channels = [["train_nll_bound", "test_nll_bound"],
["train_kl_term_%d" % t for t in range(n_iter)],
["train_recons_term_%d" % t for t in range(n_iter)],
["train_total_gradient_norm", "train_total_step_norm"]]
#------------------------------------------------------------
mnist_train = BinarizedMNIST("train", sources=['features'])
mnist_test = BinarizedMNIST("test", sources=['features'])
#mnist_train = MNIST("train", binary=True, sources=['features'])
#mnist_test = MNIST("test", binary=True, sources=['features'])
main_loop = MainLoop(
model=None,
data_stream=ForceFloatX(
DataStream(mnist_train,
iteration_scheme=SequentialScheme(
mnist_train.num_examples, batch_size))),
algorithm=algorithm,
extensions=[
Timing(),
FinishAfter(after_n_epochs=epochs),
DataStreamMonitoring(
monitors,
ForceFloatX(
DataStream(mnist_test,
iteration_scheme=SequentialScheme(
mnist_test.num_examples, batch_size))),
updates=scan_updates,
prefix="test"),
TrainingDataMonitoring(train_monitors,
prefix="train",
after_every_epoch=True),
SerializeMainLoop(name + ".pkl"),
Plot(name, channels=plot_channels),
ProgressBar(),
Printing()
])
main_loop.run()
save_str += 'elephant_' + str(args.drop_prob) + '-'
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(save_str + "checkpoint.zip",
on_interrupt=False,
every_n_epochs=1,
use_cpickle=True),
DumpLog(save_str + "log.pkl", after_epoch=True)
])
if not args.cluster:
extensions.append(ProgressBar())
extensions.extend([
Timing(),
Printing(),
PrintingTo("log"),
])
train_stream = get_stream(which_set="train",
for_evaluation=False,
batch_size=args.batch_size,
drop_prob=args.drop_prob,
hidden_dim=args.num_hidden)
main_loop = MainLoop(data_stream=train_stream,
algorithm=algorithm,
extensions=extensions,
model=model)
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 test_progressbar():
main_loop = setup_mainloop(ProgressBar())
# We are happy if it does not crash or raise any exceptions
main_loop.run()
def create_main_loop(save_to,
num_epochs,
unit_order=None,
batch_size=500,
num_batches=None):
image_size = (28, 28)
output_size = 10
convnet = create_lenet_5()
x = tensor.tensor4('features')
y = tensor.lmatrix('targets')
# Normalize input and apply the convnet
probs = convnet.apply(x)
case_costs = CasewiseCrossEntropy().apply(y.flatten(), probs)
cost = case_costs.mean().copy(name='cost')
# cost = (CategoricalCrossEntropy().apply(y.flatten(), probs)
# .copy(name='cost'))
error_rate = (MisclassificationRate().apply(y.flatten(),
probs).copy(name='error_rate'))
cg = ComputationGraph([cost, error_rate])
# Apply regularization to the cost
weights = VariableFilter(roles=[WEIGHT])(cg.variables)
cost = cost + sum([0.0003 * (W**2).sum() for W in weights])
cost.name = 'cost_with_regularization'
mnist_train = MNIST(("train", ))
mnist_train_stream = DataStream.default_stream(
mnist_train,
iteration_scheme=ShuffledScheme(mnist_train.num_examples, batch_size))
mnist_test = MNIST(("test", ))
mnist_test_stream = DataStream.default_stream(
mnist_test,
iteration_scheme=ShuffledScheme(mnist_test.num_examples, batch_size))
# Generate pics for biases
biases = VariableFilter(roles=[BIAS])(cg.parameters)
# Train with simple SGD
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=AdaDelta())
# Find layer outputs to probe
outs = OrderedDict(
reversed(
list((get_brick(out).name, out)
for out in VariableFilter(roles=[OUTPUT],
bricks=[Convolutional, Linear])(
cg.variables))))
actpic_extension = ActpicExtension(actpic_variables=outs,
case_labels=y,
pics=x,
label_count=output_size,
rectify=-1,
data_stream=mnist_test_stream,
after_batch=True)
synpic_extension = SynpicExtension(synpic_parameters=biases,
case_costs=case_costs,
case_labels=y,
pics=x,
batch_size=batch_size,
pic_size=image_size,
label_count=output_size,
after_batch=True)
# Impose an orderint for the SaveImages extension
if unit_order is not None:
with open(unit_order, 'rb') as handle:
histograms = pickle.load(handle)
unit_order = compute_unit_order(histograms)
# `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),
actpic_extension, synpic_extension,
SaveImages(picsources=[synpic_extension, actpic_extension],
title="LeNet-5: batch {i}, " +
"cost {cost_with_regularization:.2f}, " +
"trainerr {error_rate:.3f}",
data=[cost, error_rate],
graph='error_rate',
graph_len=500,
unit_order=unit_order,
after_batch=True),
DataStreamMonitoring([cost, error_rate],
mnist_test_stream,
prefix="test"),
TrainingDataMonitoring([
cost, error_rate,
aggregation.mean(algorithm.total_gradient_norm)
],
prefix="train",
after_epoch=True),
Checkpoint(save_to),
ProgressBar(),
Printing()
]
model = Model(cost)
main_loop = MainLoop(algorithm,
mnist_train_stream,
model=model,
extensions=extensions)
return main_loop
def main(name, dataset, epochs, batch_size, learning_rate, attention, n_iter,
enc_dim, dec_dim, z_dim, oldmodel):
image_size, data_train, data_valid, data_test = datasets.get_data(dataset)
train_stream = Flatten(
DataStream(data_train,
iteration_scheme=SequentialScheme(data_train.num_examples,
batch_size)))
valid_stream = Flatten(
DataStream(data_valid,
iteration_scheme=SequentialScheme(data_valid.num_examples,
batch_size)))
test_stream = Flatten(
DataStream(data_test,
iteration_scheme=SequentialScheme(data_test.num_examples,
batch_size)))
if name is None:
name = dataset
img_height, img_width = image_size
x_dim = img_height * img_width
rnninits = {
#'weights_init': Orthogonal(),
'weights_init': IsotropicGaussian(0.01),
'biases_init': Constant(0.),
}
inits = {
#'weights_init': Orthogonal(),
'weights_init': IsotropicGaussian(0.01),
'biases_init': Constant(0.),
}
if attention != "":
read_N, write_N = attention.split(',')
read_N = int(read_N)
write_N = int(write_N)
read_dim = 2 * read_N**2
reader = AttentionReader(x_dim=x_dim,
dec_dim=dec_dim,
width=img_width,
height=img_height,
N=read_N,
**inits)
writer = AttentionWriter(input_dim=dec_dim,
output_dim=x_dim,
width=img_width,
height=img_height,
N=write_N,
**inits)
attention_tag = "r%d-w%d" % (read_N, write_N)
else:
read_dim = 2 * x_dim
reader = Reader(x_dim=x_dim, dec_dim=dec_dim, **inits)
writer = Writer(input_dim=dec_dim, output_dim=x_dim, **inits)
attention_tag = "full"
#----------------------------------------------------------------------
# Learning rate
def lr_tag(value):
""" Convert a float into a short tag-usable string representation. E.g.:
0.1 -> 11
0.01 -> 12
0.001 -> 13
0.005 -> 53
"""
exp = np.floor(np.log10(value))
leading = ("%e" % value)[0]
return "%s%d" % (leading, -exp)
lr_str = lr_tag(learning_rate)
subdir = time.strftime("%Y%m%d-%H%M%S") + "-" + name
longname = "%s-%s-t%d-enc%d-dec%d-z%d-lr%s" % (
dataset, attention_tag, n_iter, enc_dim, dec_dim, z_dim, lr_str)
pickle_file = subdir + "/" + longname + ".pkl"
print("\nRunning experiment %s" % longname)
print(" dataset: %s" % dataset)
print(" subdirectory: %s" % subdir)
print(" learning rate: %g" % learning_rate)
print(" attention: %s" % attention)
print(" n_iterations: %d" % n_iter)
print(" encoder dimension: %d" % enc_dim)
print(" z dimension: %d" % z_dim)
print(" decoder dimension: %d" % dec_dim)
print(" batch size: %d" % batch_size)
print(" epochs: %d" % epochs)
print()
#----------------------------------------------------------------------
encoder_rnn = LSTM(dim=enc_dim, name="RNN_enc", **rnninits)
decoder_rnn = LSTM(dim=dec_dim, name="RNN_dec", **rnninits)
encoder_mlp = MLP([Identity()], [(read_dim + dec_dim), 4 * enc_dim],
name="MLP_enc",
**inits)
decoder_mlp = MLP([Identity()], [z_dim, 4 * dec_dim],
name="MLP_dec",
**inits)
q_sampler = Qsampler(input_dim=enc_dim, output_dim=z_dim, **inits)
draw = DrawModel(n_iter,
reader=reader,
encoder_mlp=encoder_mlp,
encoder_rnn=encoder_rnn,
sampler=q_sampler,
decoder_mlp=decoder_mlp,
decoder_rnn=decoder_rnn,
writer=writer)
draw.initialize()
#------------------------------------------------------------------------
x = tensor.matrix('features')
#x_recons = 1. + x
x_recons, kl_terms = draw.reconstruct(x)
#x_recons, _, _, _, _ = draw.silly(x, n_steps=10, batch_size=100)
#x_recons = x_recons[-1,:,:]
#samples = draw.sample(100)
#x_recons = samples[-1, :, :]
#x_recons = samples[-1, :, :]
recons_term = BinaryCrossEntropy().apply(x, x_recons)
recons_term.name = "recons_term"
cost = recons_term + kl_terms.sum(axis=0).mean()
cost.name = "nll_bound"
#------------------------------------------------------------
cg = ComputationGraph([cost])
params = VariableFilter(roles=[PARAMETER])(cg.variables)
algorithm = GradientDescent(
cost=cost,
params=params,
step_rule=CompositeRule([
StepClipping(10.),
Adam(learning_rate),
])
#step_rule=RMSProp(learning_rate),
#step_rule=Momentum(learning_rate=learning_rate, momentum=0.95)
)
#algorithm.add_updates(scan_updates)
#------------------------------------------------------------------------
# Setup monitors
monitors = [cost]
for t in range(n_iter):
kl_term_t = kl_terms[t, :].mean()
kl_term_t.name = "kl_term_%d" % t
#x_recons_t = T.nnet.sigmoid(c[t,:,:])
#recons_term_t = BinaryCrossEntropy().apply(x, x_recons_t)
#recons_term_t = recons_term_t.mean()
#recons_term_t.name = "recons_term_%d" % t
monitors += [kl_term_t]
train_monitors = monitors[:]
train_monitors += [aggregation.mean(algorithm.total_gradient_norm)]
train_monitors += [aggregation.mean(algorithm.total_step_norm)]
# Live plotting...
plot_channels = [
["train_nll_bound", "test_nll_bound"],
["train_kl_term_%d" % t for t in range(n_iter)],
#["train_recons_term_%d" % t for t in range(n_iter)],
["train_total_gradient_norm", "train_total_step_norm"]
]
#------------------------------------------------------------
if not os.path.exists(subdir):
os.makedirs(subdir)
main_loop = MainLoop(
model=Model(cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=[
Timing(),
FinishAfter(after_n_epochs=epochs),
TrainingDataMonitoring(train_monitors,
prefix="train",
after_epoch=True),
# DataStreamMonitoring(
# monitors,
# valid_stream,
## updates=scan_updates,
# prefix="valid"),
DataStreamMonitoring(
monitors,
test_stream,
# updates=scan_updates,
prefix="test"),
Checkpoint(name,
before_training=False,
after_epoch=True,
save_separately=['log', 'model']),
#Checkpoint(image_size=image_size, save_subdir=subdir, path=pickle_file, before_training=False, after_epoch=True, save_separately=['log', 'model']),
Plot(name, channels=plot_channels),
ProgressBar(),
Printing()
])
if oldmodel is not None:
print("Initializing parameters with old model %s" % oldmodel)
with open(oldmodel, "rb") as f:
oldmodel = pickle.load(f)
main_loop.model.set_param_values(oldmodel.get_param_values())
del oldmodel
main_loop.run()
DataStreamMonitoring([v for l in m.monitor_vars_valid for v in l],
valid_stream,
prefix='valid'),
]
if plot_avail:
plot_channels = [
['train_' + v.name for v in lt] + ['valid_' + v.name for v in lv]
for lt, lv in zip(m.monitor_vars, m.monitor_vars_valid)
]
extensions += [
Plot(
document='deepmind_qa_' + model_name,
channels=plot_channels,
# server_url='http://localhost:5006/', # If you need, change this
every_n_batches=config.print_freq)
]
extensions += [
Printing(after_epoch=True),
# EvaluateModel(path="", model=test_model, data_stream=valid_stream, vocab_size = ds.vocab_size, vocab = ds.vocab, eval_mode='batch', quiet=True, after_epoch=True),
ProgressBar()
]
main_loop = MainLoop(model=model,
data_stream=train_stream,
algorithm=algorithm,
extensions=extensions)
# Run the model !
main_loop.run()
main_loop.profile.report()
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)
