def run(discriminative_regularization=True):
streams = create_celeba_streams(training_batch_size=100,
monitoring_batch_size=500,
include_targets=False)
main_loop_stream, train_monitor_stream, valid_monitor_stream = streams[:3]
# Compute parameter updates for the batch normalization population
# statistics. They are updated following an exponential moving average.
rval = create_training_computation_graphs(discriminative_regularization)
cg, bn_cg, variance_parameters = rval
pop_updates = list(
set(get_batch_normalization_updates(bn_cg, allow_duplicates=True)))
decay_rate = 0.05
extra_updates = [(p, m * decay_rate + p * (1 - decay_rate))
for p, m in pop_updates]
model = Model(bn_cg.outputs[0])
selector = Selector(
find_bricks(
model.top_bricks,
lambda brick: brick.name in ('encoder_convnet', 'encoder_mlp',
'decoder_convnet', 'decoder_mlp')))
parameters = list(selector.get_parameters().values()) + variance_parameters
# Prepare algorithm
step_rule = Adam()
algorithm = GradientDescent(cost=bn_cg.outputs[0],
parameters=parameters,
step_rule=step_rule)
algorithm.add_updates(extra_updates)
# Prepare monitoring
monitored_quantities_list = []
for graph in [bn_cg, cg]:
cost, kl_term, reconstruction_term = graph.outputs
cost.name = 'nll_upper_bound'
avg_kl_term = kl_term.mean(axis=0)
avg_kl_term.name = 'avg_kl_term'
avg_reconstruction_term = -reconstruction_term.mean(axis=0)
avg_reconstruction_term.name = 'avg_reconstruction_term'
monitored_quantities_list.append(
[cost, avg_kl_term, avg_reconstruction_term])
train_monitoring = DataStreamMonitoring(
monitored_quantities_list[0], train_monitor_stream, prefix="train",
updates=extra_updates, after_epoch=False, before_first_epoch=False,
every_n_epochs=5)
valid_monitoring = DataStreamMonitoring(
monitored_quantities_list[1], valid_monitor_stream, prefix="valid",
after_epoch=False, before_first_epoch=False, every_n_epochs=5)
# Prepare checkpoint
save_path = 'celeba_vae_{}regularization.zip'.format(
'' if discriminative_regularization else 'no_')
checkpoint = Checkpoint(save_path, every_n_epochs=5, use_cpickle=True)
extensions = [Timing(), FinishAfter(after_n_epochs=75), train_monitoring,
valid_monitoring, checkpoint, Printing(), ProgressBar()]
main_loop = MainLoop(data_stream=main_loop_stream,
algorithm=algorithm, extensions=extensions)
main_loop.run()
def train_rnnrbm(train, rnnrbm, epochs=1000, test=None, bokeh=True,
load_path=None):
cdk = theano.shared(10)
lr = theano.shared(float32(0.004))
cost, v_sample = rnnrbm.cost(examples=x, mask=x_mask, k=cdk)
error_rate = MismulitclassificationRate().apply(x, v_sample[-1], x_mask)
error_rate.name = "error on note as a whole"
mistake_rate = MismulitmistakeRate().apply(x, v_sample[-1], x_mask)
mistake_rate.name = "single error within note"
cost.name = 'rbm_cost'
model = Model(cost)
cg = ComputationGraph([cost])
step_rule = CompositeRule(
[RemoveNotFinite(), StepClipping(30.0), Adam(learning_rate=lr), StepClipping(6.0),
RemoveNotFinite()]) # Scale(0.01)
gradients = dict(equizip(cg.parameters, T.grad(cost, cg.parameters, consider_constant=[v_sample])))
algorithm = GradientDescent(step_rule=step_rule, gradients=gradients, cost=cost,
params=cg.parameters)
algorithm.add_updates(cg.updates)
extensions = [
SharedVariableModifier(parameter=cdk,
function=lambda n, v: rnnrbm_cdk[n] if rnnrbm_cdk.get(n) else v),
SharedVariableModifier(parameter=lr,
function=lambda n, v: float32(0.78 * v) if n % (200 * 5) == 0 else v),
FinishAfter(after_n_epochs=epochs),
TrainingDataMonitoring(
[cost, error_rate, mistake_rate, ], # hidden_states, debug_val, param_nans,
# aggregation.mean(algorithm.total_gradient_norm)], #+ params,
prefix="train",
after_epoch=False, every_n_batches=40),
Timing(),
Printing(),
ProgressBar()]
if test is not None:
extensions.append(DataStreamMonitoring(
[cost, error_rate, mistake_rate],
data_stream=test,
updates=cg.updates,
prefix="test", after_epoch=False, every_n_batches=40))
if bokeh:
extensions.append(Plot(
'Training RNN-RBM',
channels=[
['train_error on note as a whole', 'train_single error within note',
'test_error on note as a whole',
'test_single error within note'],
['train_final_cost'],
# ['train_total_gradient_norm'],
]))
main_loop = MainLoop(algorithm=algorithm,
data_stream=train,
model=model,
extensions=extensions
)
return main_loop
def test_gradient_descent_finds_inputs_additional_updates():
W = shared_floatx(numpy.array([[1, 2], [3, 4]]))
n = shared_floatx(1)
m = tensor.scalar('m')
algorithm = GradientDescent(gradients=OrderedDict([(W, W + 1)]))
algorithm.add_updates([(n, n + m)])
algorithm.initialize()
assert m in algorithm.inputs
def test_gradient_descent_finds_inputs_additional_updates():
W = shared_floatx(numpy.array([[1, 2], [3, 4]]))
n = shared_floatx(1)
m = tensor.scalar('m')
algorithm = GradientDescent(gradients=OrderedDict([(W, W + 1)]))
algorithm.add_updates([(n, n + m)])
algorithm.initialize()
assert m in algorithm.inputs
def create_main_loop(dataset, nvis, nhid, num_epochs, debug_level=0, lrate=1e-3):
seed = 188229
n_inference_steps = 6
num_examples = dataset.num_examples
batch_size = num_examples
train_loop_stream = Flatten(
DataStream.default_stream(
dataset=dataset,
iteration_scheme=SequentialScheme(dataset.num_examples, batch_size) # Repeat(
# , n_inference_steps)
# ShuffledScheme(dataset.num_examples, batch_size), n_inference_steps))
),
which_sources=("features",),
)
model_brick = FivEM(
nvis=nvis,
nhid=nhid,
epsilon=0.01,
batch_size=batch_size,
weights_init=IsotropicGaussian(0.1),
biases_init=Constant(0),
noise_scaling=1,
debug=debug_level,
lateral_x=False,
lateral_h=False,
n_inference_steps=n_inference_steps,
)
model_brick.initialize()
x = tensor.matrix("features")
cost = model_brick.cost(x)
computation_graph = ComputationGraph([cost])
model = Model(cost)
# step_rule = Adam(learning_rate=2e-5, beta1=0.1, beta2=0.001, epsilon=1e-8,
# decay_factor=(1 - 1e-8))
step_rule = Momentum(learning_rate=lrate, momentum=0.95)
# step_rule = AdaDelta()
# step_rule = RMSProp(learning_rate=0.01)
# step_rule = AdaGrad(learning_rate=1e-4)
algorithm = GradientDescent(cost=cost, params=computation_graph.parameters, step_rule=step_rule)
algorithm.add_updates(computation_graph.updates)
extensions = [
Timing(),
FinishAfter(after_n_epochs=num_epochs),
TrainingDataMonitoring([cost] + computation_graph.auxiliary_variables, after_batch=False, after_epoch=True),
# every_n_epochs=1),
Printing(after_epoch=True, after_batch=False), # every_n_epochs=1,
# Checkpoint(path="./fivem.zip",every_n_epochs=10,after_training=True)
]
main_loop = MainLoop(model=model, data_stream=train_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 train_language_model(new_training_job, config, save_path, params,
fast_start, fuel_server, seed):
c = config
if seed:
fuel.config.default_seed = seed
blocks.config.config.default_seed = seed
data, lm, retrieval = initialize_data_and_model(config)
# full main loop can be saved...
main_loop_path = os.path.join(save_path, 'main_loop.tar')
# or only state (log + params) which can be useful not to pickle embeddings
state_path = os.path.join(save_path, 'training_state.tar')
stream_path = os.path.join(save_path, 'stream.pkl')
best_tar_path = os.path.join(save_path, "best_model.tar")
words = tensor.ltensor3('words')
words_mask = tensor.matrix('words_mask')
if theano.config.compute_test_value != 'off':
test_value_data = next(
data.get_stream('train', batch_size=4,
max_length=5).get_epoch_iterator())
words.tag.test_value = test_value_data[0]
words_mask.tag.test_value = test_value_data[1]
costs, updates = lm.apply(words, words_mask)
cost = rename(costs.mean(), 'mean_cost')
cg = Model(cost)
if params:
logger.debug("Load parameters from {}".format(params))
with open(params) as src:
cg.set_parameter_values(load_parameters(src))
length = rename(words.shape[1], 'length')
perplexity, = VariableFilter(name='perplexity')(cg)
perplexities = VariableFilter(name_regex='perplexity.*')(cg)
monitored_vars = [length, cost] + perplexities
if c['dict_path']:
num_definitions, = VariableFilter(name='num_definitions')(cg)
monitored_vars.extend([num_definitions])
parameters = cg.get_parameter_dict()
trained_parameters = parameters.values()
saved_parameters = parameters.values()
if c['embedding_path']:
logger.debug("Exclude word embeddings from the trained parameters")
trained_parameters = [
p for p in trained_parameters
if not p == lm.get_def_embeddings_params()
]
saved_parameters = [
p for p in saved_parameters
if not p == lm.get_def_embeddings_params()
]
if c['cache_size'] != 0:
logger.debug("Enable fake recursivity for looking up embeddings")
trained_parameters = [
p for p in trained_parameters if not p == lm.get_cache_params()
]
logger.info("Cost parameters" + "\n" + pprint.pformat([
" ".join(
(key, str(parameters[key].get_value().shape),
'trained' if parameters[key] in trained_parameters else 'frozen'))
for key in sorted(parameters.keys())
],
width=120))
rules = []
if c['grad_clip_threshold']:
rules.append(StepClipping(c['grad_clip_threshold']))
rules.append(Adam(learning_rate=c['learning_rate'], beta1=c['momentum']))
algorithm = GradientDescent(cost=cost,
parameters=trained_parameters,
step_rule=CompositeRule(rules))
if c['cache_size'] != 0:
algorithm.add_updates(updates)
train_monitored_vars = list(monitored_vars)
if c['grad_clip_threshold']:
train_monitored_vars.append(algorithm.total_gradient_norm)
word_emb_RMS, = VariableFilter(name='word_emb_RMS')(cg)
main_rnn_in_RMS, = VariableFilter(name='main_rnn_in_RMS')(cg)
train_monitored_vars.extend([word_emb_RMS, main_rnn_in_RMS])
if c['monitor_parameters']:
train_monitored_vars.extend(parameter_stats(parameters, algorithm))
# We use a completely random seed on purpose. With Fuel server
# it's currently not possible to restore the state of the training
# stream. That's why it's probably better to just have it stateless.
stream_seed = numpy.random.randint(0, 10000000) if fuel_server else None
training_stream = data.get_stream('train',
batch_size=c['batch_size'],
max_length=c['max_length'],
seed=stream_seed)
valid_stream = data.get_stream('valid',
batch_size=c['batch_size_valid'],
max_length=c['max_length'],
seed=stream_seed)
original_training_stream = training_stream
if fuel_server:
# the port will be configured by the StartFuelServer extension
training_stream = ServerDataStream(
sources=training_stream.sources,
produces_examples=training_stream.produces_examples)
validation = DataStreamMonitoring(monitored_vars,
valid_stream,
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
on_resumption=True,
every_n_batches=c['mon_freq_valid'])
track_the_best = TrackTheBest(validation.record_name(perplexity),
choose_best=min).set_conditions(
on_resumption=True,
after_epoch=True,
every_n_batches=c['mon_freq_valid'])
# don't save them the entire main loop to avoid pickling everything
if c['fast_checkpoint']:
load = (LoadNoUnpickling(state_path,
load_iteration_state=True,
load_log=True).set_conditions(
before_training=not new_training_job))
cp_args = {
'save_main_loop': False,
'save_separately': ['log', 'iteration_state'],
'parameters': saved_parameters
}
checkpoint = Checkpoint(state_path,
before_training=not fast_start,
every_n_batches=c['save_freq_batches'],
after_training=not fast_start,
**cp_args)
if c['checkpoint_every_n_batches']:
intermediate_cp = IntermediateCheckpoint(
state_path,
every_n_batches=c['checkpoint_every_n_batches'],
after_training=False,
**cp_args)
else:
load = (Load(main_loop_path, load_iteration_state=True,
load_log=True).set_conditions(
before_training=not new_training_job))
cp_args = {
'save_separately': ['iteration_state'],
'parameters': saved_parameters
}
checkpoint = Checkpoint(main_loop_path,
before_training=not fast_start,
every_n_batches=c['save_freq_batches'],
after_training=not fast_start,
**cp_args)
if c['checkpoint_every_n_batches']:
intermediate_cp = IntermediateCheckpoint(
main_loop_path,
every_n_batches=c['checkpoint_every_n_batches'],
after_training=False,
**cp_args)
checkpoint = checkpoint.add_condition(
['after_batch', 'after_epoch'],
OnLogRecord(track_the_best.notification_name), (best_tar_path, ))
extensions = [
load,
StartFuelServer(original_training_stream,
stream_path,
before_training=fuel_server),
Timing(every_n_batches=c['mon_freq_train'])
]
if retrieval:
extensions.append(
RetrievalPrintStats(retrieval=retrieval,
every_n_batches=c['mon_freq_train'],
before_training=not fast_start))
extensions.extend([
TrainingDataMonitoring(train_monitored_vars,
prefix="train",
every_n_batches=c['mon_freq_train']),
validation, track_the_best, checkpoint
])
if c['checkpoint_every_n_batches']:
extensions.append(intermediate_cp)
extensions.extend([
DumpTensorflowSummaries(save_path,
every_n_batches=c['mon_freq_train'],
after_training=True),
Printing(on_resumption=True, every_n_batches=c['mon_freq_train']),
FinishIfNoImprovementAfter(track_the_best.notification_name,
iterations=50 * c['mon_freq_valid'],
every_n_batches=c['mon_freq_valid']),
FinishAfter(after_n_batches=c['n_batches'])
])
logger.info("monitored variables during training:" + "\n" +
pprint.pformat(train_monitored_vars, width=120))
logger.info("monitored variables during valid:" + "\n" +
pprint.pformat(monitored_vars, width=120))
main_loop = MainLoop(algorithm,
training_stream,
model=Model(cost),
extensions=extensions)
main_loop.run()
if normalization == 'bn2':
for m,s,var in statistics_list:
parameters.extend([m,s])
algorithm = GradientDescent(
cost=cost, parameters=parameters, step_rule=Adam(0.01))
#update the M and S with batch statistics
alpha = 0.1
updates = []
if normalization == 'bn2':
for m,s,var in statistics_list:
updates.append((m, cast(alpha*m + (1-alpha)*var.mean(axis=0), floatX)))
updates.append((s, cast(alpha*s + (1-alpha)*var.std(axis=0) , floatX)))
algorithm.add_updates(updates)
# Since this line wont work with the extension to include parameters
# in the gradient descent. Here's an extension that will do the job.
from blocks.extensions import SimpleExtension
from theano import function
class UpdateExtraParams(SimpleExtension):
"""Adjusts shared variable parameter using some function.
"""
def __init__(self, input1, input2, updates, **kwargs):
kwargs.setdefault("after_batch", True)
super(UpdateExtraParams, self).__init__(**kwargs)
self.updates = updates
self.func = function([input1, input2],[],
updates = updates,
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(1.0), 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 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)
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)
def construct_main_loop(name, task_name, batch_size, max_epochs,
patience_epochs, learning_rate, hyperparameters,
**kwargs):
task = tasks.get_task(**hyperparameters)
hyperparameters["n_channels"] = task.n_channels
extensions = []
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
algorithm = GradientDescent(cost=outputs["train"]["cost"],
parameters=graphs["train"].parameters,
step_rule=CompositeRule([
StepClipping(1e1),
Adam(learning_rate=learning_rate),
StepClipping(1e2)
]),
on_unused_sources="warn")
algorithm.add_updates(updates["train"])
extensions.extend(
construct_monitors(algorithm=algorithm,
task=task,
model=model,
graphs=graphs,
outputs=outputs,
**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
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=5,
before_training=True,
use_cpickle=True),
ProgressBar(),
Timing(),
Printing(),
PrintingTo(name + "_log")
])
from blocks.main_loop import MainLoop
main_loop = MainLoop(data_stream=task.get_stream("train"),
algorithm=algorithm,
extensions=extensions,
model=model)
# note blocks will crash and burn because it cannot deal with an
# already-initialized Algorithm, so this should be enabled only for
# debugging
if False:
with open("graph", "w") as graphfile:
algorithm.initialize()
theano.printing.debugprint(algorithm._function, file=graphfile)
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
# param_nans = 0
# for i, p in enumerate(params):
# # cost += reg * abs(p).sum()
# cost += reg * (p ** 2).sum()
# param_nans += T.isnan(p).sum()
# name = params[i].name
# params[i] = params[i].mean()
# params[i].name = name + str(i)
cost.name = "final_cost"
# param_nans.name = 'params_nans'
## Training algorithm
# algorithm.initialize()
algorithm.add_updates(cg.updates)
extensions = [
FinishAfter(after_n_epochs=5000),
DataStreamMonitoring([cost, error_rate, mistake_rate],
data_stream=test,
updates=cg.updates,
prefix="test"),
TrainingDataMonitoring(
[
cost, # hidden_states, debug_val, param_nans,
aggregation.mean(algorithm.total_gradient_norm)
], #+ params,
prefix="train",
after_epoch=True),
Timing(),
def train(cli_params):
cli_params['save_dir'] = prepare_dir(cli_params['save_to'])
logfile = os.path.join(cli_params['save_dir'], 'log.txt')
# Log also DEBUG to a file
fh = logging.FileHandler(filename=logfile)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
logger.info('Logging into %s' % logfile)
p, loaded = load_and_log_params(cli_params)
in_dim, data, whiten, cnorm = setup_data(p, test_set=False)
if not loaded:
# Set the zero layer to match input dimensions
p.encoder_layers = (in_dim, ) + p.encoder_layers
ladder = setup_model(p)
# Training
all_params = ComputationGraph([ladder.costs.total]).parameters
logger.info('Found the following parameters: %s' % str(all_params))
# Fetch all batch normalization updates. They are in the clean path.
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
assert 'counter' in [u.name for u in bn_updates.keys()], \
'No batch norm params in graph - the graph has been cut?'
training_algorithm = GradientDescent(
cost=ladder.costs.total,
parameters=all_params,
step_rule=Adam(learning_rate=ladder.lr))
# In addition to actual training, also do BN variable approximations
training_algorithm.add_updates(bn_updates)
short_prints = {
"train": {
'T_C_class': ladder.costs.class_corr,
'T_C_de': ladder.costs.denois.values(),
},
"valid_approx":
OrderedDict([
('V_C_class', ladder.costs.class_clean),
('V_E', ladder.error.clean),
('V_C_de', ladder.costs.denois.values()),
]),
"valid_error":
OrderedDict([
('VE_C_class', ladder.costs.class_clean),
('VE_E', ladder.error.clean),
('VE_C_de', ladder.costs.denois.values()),
]),
}
main_loop = MainLoop(
training_algorithm,
# Datastream used for training
make_datastream(data.train,
data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
n_unlabeled=p.unlabeled_samples,
whiten=whiten,
cnorm=cnorm),
model=Model(ladder.costs.total),
extensions=[
#FinishAfter(after_n_epochs=p.num_epochs),
# This will estimate the validation error using
# running average estimates of the batch normalization
# parameters, mean and variance
#ApproxTestMonitoring(
# [ladder.costs.class_clean, ladder.error.clean]
# + ladder.costs.denois.values(),
# make_datastream(data.valid, data.valid_ind,
# p.valid_batch_size, whiten=whiten, cnorm=cnorm,
# scheme=ShuffledScheme),
# prefix="valid_approx"),
# This Monitor is slower, but more accurate since it will first
# estimate batch normalization parameters from training data and
# then do another pass to calculate the validation error.
FinalTestMonitoring(
[ladder.costs.class_clean, ladder.error.clean] +
ladder.costs.denois.values(),
make_datastream(data.train,
data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
whiten=whiten,
cnorm=cnorm,
scheme=ShuffledScheme),
make_datastream(data.valid,
data.valid_ind,
p.valid_batch_size,
n_labeled=len(data.valid_ind),
whiten=whiten,
cnorm=cnorm,
scheme=ShuffledScheme),
prefix="valid_error",
every_n_epochs=1),
TrainingDataMonitoring([
ladder.costs.total, ladder.costs.class_corr,
training_algorithm.total_gradient_norm
] + ladder.costs.denois.values(),
prefix="train",
after_epoch=True),
#SaveParams(None, all_params, p.save_dir, after_epoch=True),
#SaveExpParams(p, p.save_dir, before_training=True),
#SaveLog(p.save_dir, after_training=True),
#ShortPrinting(short_prints),
#ProgressBar(),
StopAfterNoImprovementValidation('valid_error_error_rate_clean',
10, all_params, p, p.save_dir),
#LRDecay(ladder.lr, p.num_epochs * p.lrate_decay, p.num_epochs,
# after_epoch=True),
])
main_loop.run()
# Get results
df = DataFrame.from_dict(main_loop.log, orient='index')
col = 'valid_error_error_rate_clean'
logger.info('%s %g' % (col, df[col].iloc[-1]))
if main_loop.log.status['epoch_interrupt_received']:
return None
return df
def run():
# Load Model
net_size = 256 #Hard-code instead of loading model (takes too long to set up network)
#net = vaegan.VAEGAN()
#network_saver = saver.NetworkSaver('vaegan/models/', net=net)
#network_saver.load()
# DATA
train_stream = get_stream(hdf5_file, 'train', batch_size) #TODO jonathan ?
test_stream = get_stream(hdf5_file, 'test', batch_size) #TODO jonathan ?
# MODEL
x = T.TensorType('floatX', [False] * 3)('features')
y = T.tensor3('targets', dtype='floatX')
train_flag = [theano.shared(0)]
x = x.swapaxes(0, 1)
y = y.swapaxes(0, 1)
# More Config
out_size = len(output_columns) - 1 # code_mode=RL-MDN
latent_size = net_size
in_size = latent_size + len(input_columns)
# NN fprop
y_hat, cost, cells = nn_fprop(x, y, in_size, out_size, hidden_size,
num_recurrent_layers, train_flag)
# COST
cg = ComputationGraph(cost)
extra_updates = []
# RMS Prop training optimizer
step_rules = [
RMSProp(learning_rate=learning_rate, decay_rate=decay_rate),
StepClipping(step_clipping)
]
parameters_to_update = cg.parameters
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=parameters_to_update,
step_rule=CompositeRule(step_rules))
algorithm.add_updates(
extra_updates) # TODO jonathan what is this, is this needed?
# Extensions
gradient_norm = aggregation.mean(algorithm.total_gradient_norm)
step_norm = aggregation.mean(algorithm.total_step_norm)
monitored_vars = [
cost, step_rules[0].learning_rate, gradient_norm, step_norm
]
test_monitor = DataStreamMonitoring(variables=[cost],
after_epoch=True,
before_first_epoch=True,
data_stream=test_stream,
prefix="test")
train_monitor = TrainingDataMonitoring(variables=monitored_vars,
after_epoch=True,
before_first_epoch=True,
prefix='train')
set_train_flag = SetTrainFlag(after_epoch=True,
before_epoch=True,
flag=train_flag)
# plot = Plot('Plotting example', channels=[['cost']], after_batch=True, open_browser=True)
extensions = [
set_train_flag,
test_monitor,
train_monitor,
Timing(),
Printing(after_epoch=True),
FinishAfter(after_n_epochs=nepochs),
saveload.Load(load_path),
saveload.Checkpoint(last_path, every_n_epochs=10000),
] + track_best('test_cost',
save_path) #+ track_best('train_cost', last_path)
if learning_rate_decay not in (0, 1):
extensions.append(
SharedVariableModifier(step_rules[0].learning_rate,
lambda n, lr: np.cast[theano.config.floatX]
(learning_rate_decay * lr),
after_epoch=False,
every_n_epochs=lr_decay_every_n_epochs,
after_batch=False))
print 'number of parameters in the model: ' + str(
T.sum([p.size for p in cg.parameters]).eval())
# Finally build the main loop and train the model
mainLoop = MainLoop(data_stream=train_stream,
algorithm=algorithm,
model=Model(cost),
extensions=extensions)
mainLoop.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(num_epochs=50, batch_normalized=True, alpha=0.1):
"""Run the example.
Parameters
----------
num_epochs : int, optional
Number of epochs for which to train.
batch_normalized : bool, optional
Batch-normalize the training graph. Defaults to `True`.
alpha : float, optional
Weight to apply to a new sample when calculating running
averages for population statistics (1 - alpha weight is
given to the existing average).
"""
if batch_normalized:
# Add an extra keyword argument that only BatchNormalizedMLP takes,
# in order to speed things up at the cost of a bit of extra memory.
mlp_class = BatchNormalizedMLP
extra_kwargs = {'conserve_memory': False}
else:
mlp_class = MLP
extra_kwargs = {}
mlp = mlp_class([Logistic(), Logistic(),
Logistic(), Softmax()], [2, 5, 5, 5, 3],
weights_init=IsotropicGaussian(0.2),
biases_init=Constant(0.),
**extra_kwargs)
mlp.initialize()
# Generate a dataset with 3 spiral arms, using 8000 examples for
# training and 2000 for testing.
dataset = Spiral(num_examples=10000,
classes=3,
sources=['features', 'label'],
noise=0.05)
train_stream = DataStream(dataset,
iteration_scheme=ShuffledScheme(examples=8000,
batch_size=20))
test_stream = DataStream(dataset,
iteration_scheme=SequentialScheme(
examples=list(range(8000, 10000)),
batch_size=2000))
# Build a cost graph; this contains BatchNormalization bricks that will
# by default run in inference mode.
features = tensor.matrix('features')
label = tensor.lvector('label')
prediction = mlp.apply(features)
cost = CategoricalCrossEntropy().apply(label, prediction)
misclass = MisclassificationRate().apply(label, prediction)
misclass.name = 'misclass' # The default name for this is annoyingly long
original_cg = ComputationGraph([cost, misclass])
if batch_normalized:
cg = apply_batch_normalization(original_cg)
# Add updates for population parameters
pop_updates = get_batch_normalization_updates(cg)
extra_updates = [(p, m * alpha + p * (1 - alpha))
for p, m in pop_updates]
else:
cg = original_cg
extra_updates = []
algorithm = GradientDescent(step_rule=Adam(0.001),
cost=cg.outputs[0],
parameters=cg.parameters)
algorithm.add_updates(extra_updates)
main_loop = MainLoop(
algorithm=algorithm,
data_stream=train_stream,
# Use the original cost and misclass variables so
# that we monitor the (original) inference-mode graph.
extensions=[
DataStreamMonitoring([cost, misclass],
train_stream,
prefix='train'),
DataStreamMonitoring([cost, misclass], test_stream, prefix='test'),
Printing(),
FinishAfter(after_n_epochs=num_epochs)
])
main_loop.run()
return main_loop
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 train_model(cost,
cross_entropy,
updates,
train_stream,
valid_stream,
args,
gate_values=None):
step_rule = learning_algorithm(args)
cg = ComputationGraph(cost)
# ADD REGULARIZATION
# WEIGHT NOISE
weight_noise = args.weight_noise
if weight_noise > 0:
weights = VariableFilter(roles=[WEIGHT])(cg.variables)
cg_train = apply_noise(cg, weights, weight_noise)
cost = cg_train.outputs[0]
cost.name = "cost_with_weight_noise"
cg = ComputationGraph(cost)
logger.info(cg.parameters)
algorithm = GradientDescent(cost=cost,
step_rule=step_rule,
params=cg.parameters)
algorithm.add_updates(updates)
# extensions to be added
extensions = []
if args.load_path is not None:
extensions.append(Load(args.load_path))
outputs = [
variable for variable in cg.variables if variable.name == "presoft"
]
if args.generate:
extensions.append(
TextGenerationExtension(
outputs=outputs,
generation_length=args.generated_text_lenght,
initial_text_length=args.initial_text_length,
every_n_batches=args.monitoring_freq,
ploting_path=os.path.join(args.save_path, 'prob_plot.png'),
softmax_sampling=args.softmax_sampling,
dataset=args.dataset,
updates=updates,
interactive_mode=args.interactive_mode))
extensions.extend([
TrainingDataMonitoring([cost],
prefix='train',
every_n_batches=args.monitoring_freq,
after_epoch=True),
DataStreamMonitoring([cost, cross_entropy],
valid_stream,
args.mini_batch_size_valid,
state_updates=updates,
prefix='valid',
before_first_epoch=not (args.visualize_gates),
every_n_batches=args.monitoring_freq),
ResetStates([v for v, _ in updates], every_n_batches=100),
ProgressBar()
])
# Creating directory for saving model.
if not args.interactive_mode:
if not os.path.exists(args.save_path):
os.makedirs(args.save_path)
else:
raise Exception('Directory already exists')
early_stopping = EarlyStopping('valid_cross_entropy',
args.patience,
args.save_path,
every_n_batches=args.monitoring_freq)
# Visualizing extensions
if args.interactive_mode:
extensions.append(InteractiveMode())
if args.visualize_gates and (gate_values is not None):
if args.rnn_type == "lstm":
extensions.append(
VisualizeGateLSTM(gate_values,
updates,
args.dataset,
ploting_path=None))
elif args.rnn_type == "soft":
extensions.append(
VisualizeGateSoft(gate_values,
updates,
args.dataset,
ploting_path=None))
else:
assert (False)
extensions.append(early_stopping)
extensions.append(Printing(every_n_batches=args.monitoring_freq))
main_loop = MainLoop(model=Model(cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=extensions)
main_loop.run()
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 main(num_epochs=50, batch_normalized=True, alpha=0.1):
"""Run the example.
Parameters
----------
num_epochs : int, optional
Number of epochs for which to train.
batch_normalized : bool, optional
Batch-normalize the training graph. Defaults to `True`.
alpha : float, optional
Weight to apply to a new sample when calculating running
averages for population statistics (1 - alpha weight is
given to the existing average).
"""
if batch_normalized:
# Add an extra keyword argument that only BatchNormalizedMLP takes,
# in order to speed things up at the cost of a bit of extra memory.
mlp_class = BatchNormalizedMLP
extra_kwargs = {'conserve_memory': False}
else:
mlp_class = MLP
extra_kwargs = {}
mlp = mlp_class([Logistic(), Logistic(), Logistic(), Softmax()],
[2, 5, 5, 5, 3],
weights_init=IsotropicGaussian(0.2),
biases_init=Constant(0.), **extra_kwargs)
mlp.initialize()
# Generate a dataset with 3 spiral arms, using 8000 examples for
# training and 2000 for testing.
dataset = Spiral(num_examples=10000, classes=3,
sources=['features', 'label'],
noise=0.05)
train_stream = DataStream(dataset,
iteration_scheme=ShuffledScheme(examples=8000,
batch_size=20))
test_stream = DataStream(dataset,
iteration_scheme=SequentialScheme(
examples=list(range(8000, 10000)),
batch_size=2000))
# Build a cost graph; this contains BatchNormalization bricks that will
# by default run in inference mode.
features = tensor.matrix('features')
label = tensor.lvector('label')
prediction = mlp.apply(features)
cost = CategoricalCrossEntropy().apply(label, prediction)
misclass = MisclassificationRate().apply(label, prediction)
misclass.name = 'misclass' # The default name for this is annoyingly long
original_cg = ComputationGraph([cost, misclass])
if batch_normalized:
cg = apply_batch_normalization(original_cg)
# Add updates for population parameters
pop_updates = get_batch_normalization_updates(cg)
extra_updates = [(p, m * alpha + p * (1 - alpha))
for p, m in pop_updates]
else:
cg = original_cg
extra_updates = []
algorithm = GradientDescent(step_rule=Adam(0.001),
cost=cg.outputs[0],
parameters=cg.parameters)
algorithm.add_updates(extra_updates)
main_loop = MainLoop(algorithm=algorithm,
data_stream=train_stream,
# Use the original cost and misclass variables so
# that we monitor the (original) inference-mode graph.
extensions=[DataStreamMonitoring([cost, misclass],
train_stream,
prefix='train'),
DataStreamMonitoring([cost, misclass],
test_stream,
prefix='test'),
Printing(),
FinishAfter(after_n_epochs=num_epochs)])
main_loop.run()
return main_loop
def main(save_to, num_epochs,
regularization=0.0001, subset=None, num_batches=None,
batch_size=None, histogram=None, resume=False):
output_size = 10
convnet = create_res_net()
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):
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]
# 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=4, which_sources=('features',))
test_batch_size = 500
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])
# from theano.compile.nanguardmode import NanGuardMode
# Train with simple SGD
algorithm = GradientDescent(
cost=train_cost, parameters=train_cg.parameters,
step_rule=momentum)
algorithm.add_updates(extra_updates)
#,
# theano_func_kwargs={
# 'mode': NanGuardMode(
# nan_is_error=True, inf_is_error=True, big_is_error=True)})
# `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
(1, 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)
]),
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=17),
# 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=17),
Plot('Test performance for ' + save_to,
channels=[[
'train_error_rate',
'test_error_rate',
]],
after_epoch=True),
Checkpoint(save_to, use_cpickle=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(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 train_ladder(cli_params, dataset=None, save_to='results/ova_all_full'):
cli_params['save_dir'] = prepare_dir(save_to)
logfile = os.path.join(cli_params['save_dir'], 'log.txt')
# Log also DEBUG to a file
fh = logging.FileHandler(filename=logfile)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
logger.info('Logging into %s' % logfile)
p, loaded = load_and_log_params(cli_params)
ladder = setup_model(p)
# Training
all_params = ComputationGraph([ladder.costs.total]).parameters
logger.info('Found the following parameters: %s' % str(all_params))
# Fetch all batch normalization updates. They are in the clean path.
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
assert 'counter' in [u.name for u in bn_updates.keys()], \
'No batch norm params in graph - the graph has been cut?'
training_algorithm = GradientDescent(
cost=ladder.costs.total,
params=all_params,
step_rule=Adam(learning_rate=ladder.lr))
# In addition to actual training, also do BN variable approximations
training_algorithm.add_updates(bn_updates)
short_prints = {
"train": {
'T_C_class': ladder.costs.class_corr,
'T_C_de': ladder.costs.denois.values(),
},
"valid_approx":
OrderedDict([
('V_C_class', ladder.costs.class_clean),
('V_E', ladder.error.clean),
('V_C_de', ladder.costs.denois.values()),
]),
"valid_final":
OrderedDict([
('VF_C_class', ladder.costs.class_clean),
('VF_E', ladder.error.clean),
('VF_C_de', ladder.costs.denois.values()),
]),
}
ovadataset = dataset['ovadataset']
train_indexes = dataset['train_indexes']
val_indexes = dataset['val_indexes']
main_loop = MainLoop(
training_algorithm,
# Datastream used for training
make_datastream(ovadataset,
train_indexes,
p.batch_size,
scheme=ShuffledScheme),
model=Model(ladder.costs.total),
extensions=[
FinishAfter(after_n_epochs=p.num_epochs),
# This will estimate the validation error using
# running average estimates of the batch normalization
# parameters, mean and variance
ApproxTestMonitoring(
[ladder.costs.class_clean, ladder.error.clean] +
ladder.costs.denois.values(),
make_datastream(ovadataset, val_indexes, p.batch_size),
prefix="valid_approx"),
# This Monitor is slower, but more accurate since it will first
# estimate batch normalization parameters from training data and
# then do another pass to calculate the validation error.
FinalTestMonitoring(
[ladder.costs.class_clean, ladder.error.clean_mc] +
ladder.costs.denois.values(),
make_datastream(ovadataset, train_indexes, p.batch_size),
make_datastream(ovadataset, val_indexes, p.batch_size),
prefix="valid_final",
after_n_epochs=p.num_epochs),
TrainingDataMonitoring([
ladder.costs.total, ladder.costs.class_corr,
training_algorithm.total_gradient_norm
] + ladder.costs.denois.values(),
prefix="train",
after_epoch=True),
ShortPrinting(short_prints),
LRDecay(ladder.lr,
p.num_epochs * p.lrate_decay,
p.num_epochs,
after_epoch=True),
])
main_loop.run()
# Get results
df = main_loop.log.to_dataframe()
col = 'valid_final_error_matrix_cost'
logger.info('%s %g' % (col, df[col].iloc[-1]))
ds = make_datastream(ovadataset, val_indexes, p.batch_size)
outputs = ladder.act.clean.labeled.h[len(ladder.layers) - 1]
outputreplacer = TestMonitoring()
_, _, outputs = outputreplacer._get_bn_params(outputs)
cg = ComputationGraph(outputs)
f = cg.get_theano_function()
it = ds.get_epoch_iterator(as_dict=True)
res = []
inputs = {
'features_labeled': [],
'targets_labeled': [],
'features_unlabeled': []
}
# Loop over one epoch
for d in it:
# Store all inputs
for k, v in d.iteritems():
inputs[k] += [v]
# Store outputs
res += [f(*[d[str(inp)] for inp in cg.inputs])]
# Concatenate all minibatches
res = [numpy.vstack(minibatches) for minibatches in zip(*res)]
inputs = {k: numpy.vstack(v) for k, v in inputs.iteritems()}
if main_loop.log.status['epoch_interrupt_received']:
return None
return res[0], inputs
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 run(batch_size, classifier, oldmodel, monitor_every, checkpoint_every, final_epoch,
dataset, color_convert, image_size, net_depth, allowed, stretch):
streams = create_custom_streams(filename=dataset,
training_batch_size=batch_size,
monitoring_batch_size=batch_size,
include_targets=True,
color_convert=color_convert,
allowed=allowed,
stretch=stretch)
main_loop_stream = streams[0]
train_monitor_stream = streams[1]
valid_monitor_stream = streams[2]
cg, bn_dropout_cg = create_training_computation_graphs(image_size, net_depth)
model = Model(bn_dropout_cg.outputs[0])
# 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=True, after_epoch=False,
every_n_epochs=monitor_every)
# Prepare checkpoint
checkpoint = Checkpoint(classifier, every_n_epochs=checkpoint_every,
use_cpickle=True)
extensions = [Timing(), FinishAfter(after_n_epochs=final_epoch), train_monitoring,
valid_monitoring, checkpoint, Printing(), ProgressBar()]
main_loop = MainLoop(model=model, data_stream=main_loop_stream,
algorithm=algorithm, extensions=extensions)
if oldmodel is not None:
print("Initializing parameters with old model {}".format(oldmodel))
with open(oldmodel, 'rb') as src:
saved_model = load(src)
main_loop.model.set_parameter_values(
saved_model.model.get_parameter_values())
del saved_model
main_loop.run()
def train(cli_params):
cli_params['save_dir'] = prepare_dir(cli_params['save_to'])
logfile = os.path.join(cli_params['save_dir'], 'log.txt')
# Log also DEBUG to a file
fh = logging.FileHandler(filename=logfile)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
logger.info('Logging into %s' % logfile)
p, loaded = load_and_log_params(cli_params)
in_dim, data, whiten, cnorm = setup_data(p, test_set=False)
if not loaded:
# Set the zero layer to match input dimensions
p.encoder_layers = (in_dim,) + p.encoder_layers
ladder = setup_model(p)
# Training
all_params = ComputationGraph([ladder.costs.total]).parameters
logger.info('Found the following parameters: %s' % str(all_params))
# Fetch all batch normalization updates. They are in the clean path.
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
assert 'counter' in [u.name for u in bn_updates.keys()], \
'No batch norm params in graph - the graph has been cut?'
training_algorithm = GradientDescent(
cost=ladder.costs.total, parameters=all_params,
step_rule=Adam(learning_rate=ladder.lr.get_value()))
# In addition to actual training, also do BN variable approximations
training_algorithm.add_updates(bn_updates)
short_prints = {
"train": {
'T_C_class': ladder.costs.class_corr,
'T_C_de': ladder.costs.denois.values(),
},
"valid_approx": OrderedDict([
('V_C_class', ladder.costs.class_clean),
('V_E', ladder.error.clean),
('V_C_de', ladder.costs.denois.values()),
]),
"valid_final": OrderedDict([
('VF_C_class', ladder.costs.class_clean),
('VF_E', ladder.error.clean),
('VF_C_de', ladder.costs.denois.values()),
]),
}
main_loop = MainLoop(
training_algorithm,
# Datastream used for training
make_datastream(data.train, data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
n_unlabeled=p.unlabeled_samples,
whiten=whiten,
cnorm=cnorm),
model=Model(ladder.costs.total),
extensions=[
FinishAfter(after_n_epochs=p.num_epochs),
# This will estimate the validation error using
# running average estimates of the batch normalization
# parameters, mean and variance
ApproxTestMonitoring(
[ladder.costs.class_clean, ladder.error.clean]
+ ladder.costs.denois.values(),
make_datastream(data.valid, data.valid_ind,
p.valid_batch_size, whiten=whiten, cnorm=cnorm,
scheme=ShuffledScheme),
prefix="valid_approx"),
# This Monitor is slower, but more accurate since it will first
# estimate batch normalization parameters from training data and
# then do another pass to calculate the validation error.
FinalTestMonitoring(
[ladder.costs.class_clean, ladder.error.clean]
+ ladder.costs.denois.values(),
make_datastream(data.train, data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
whiten=whiten, cnorm=cnorm,
scheme=ShuffledScheme),
make_datastream(data.valid, data.valid_ind,
p.valid_batch_size,
n_labeled=len(data.valid_ind),
whiten=whiten, cnorm=cnorm,
scheme=ShuffledScheme),
prefix="valid_final",
after_n_epochs=p.num_epochs),
TrainingDataMonitoring(
[ladder.costs.total, ladder.costs.class_corr,
training_algorithm.total_gradient_norm]
+ ladder.costs.denois.values(),
prefix="train", after_epoch=True),
SaveParams(None, all_params, p.save_dir, after_epoch=True),
SaveExpParams(p, p.save_dir, before_training=True),
SaveLog(p.save_dir, after_training=True),
ShortPrinting(short_prints),
LRDecay(ladder.lr, p.num_epochs * p.lrate_decay, p.num_epochs,
after_epoch=True),
])
main_loop.run()
# Get results
df = DataFrame.from_dict(main_loop.log, orient='index')
col = 'valid_final_error_rate_clean'
logger.info('%s %g' % (col, df[col].iloc[-1]))
if main_loop.log.status['epoch_interrupt_received']:
return None
return df
def main(mode, save_to, num_epochs, load_params=None,
feature_maps=None, mlp_hiddens=None,
conv_sizes=None, pool_sizes=None, stride=None, repeat_times=None,
batch_size=None, num_batches=None, algo=None,
test_set=None, valid_examples=None,
dropout=None, max_norm=None, weight_decay=None,
batch_norm=None):
if feature_maps is None:
feature_maps = [20, 50, 50]
if mlp_hiddens is None:
mlp_hiddens = [500]
if conv_sizes is None:
conv_sizes = [5, 5, 5]
if pool_sizes is None:
pool_sizes = [2, 2, 2]
if repeat_times is None:
repeat_times = [1, 1, 1]
if batch_size is None:
batch_size = 500
if valid_examples is None:
valid_examples = 2500
if stride is None:
stride = 1
if test_set is None:
test_set = 'test'
if algo is None:
algo = 'rmsprop'
if batch_norm is None:
batch_norm = False
image_size = (128, 128)
output_size = 2
if (len(feature_maps) != len(conv_sizes) or
len(feature_maps) != len(pool_sizes) or
len(feature_maps) != len(repeat_times)):
raise ValueError("OMG, inconsistent arguments")
# 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, 3, image_size,
stride=stride,
filter_sizes=zip(conv_sizes, conv_sizes),
feature_maps=feature_maps,
pooling_sizes=zip(pool_sizes, pool_sizes),
repeat_times=repeat_times,
top_mlp_activations=mlp_activations,
top_mlp_dims=mlp_hiddens + [output_size],
border_mode='full',
batch_norm=batch_norm,
weights_init=Glorot(),
biases_init=Constant(0))
# We push initialization config to set different initialization schemes
# for convolutional layers.
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')))
single_x = tensor.tensor3('image_features')
x = tensor.tensor4('image_features')
single_y = tensor.lvector('targets')
y = tensor.lmatrix('targets')
# Training
with batch_normalization(convnet):
probs = convnet.apply(x)
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])
extra_updates = []
if batch_norm: # batch norm:
logger.debug("Apply batch norm")
pop_updates = get_batch_normalization_updates(cg)
# p stands for population mean
# m stands for minibatch
alpha = 0.005
extra_updates = [(p, m * alpha + p * (1 - alpha))
for p, m in pop_updates]
population_statistics = [p for p, m in extra_updates]
if dropout:
relu_outputs = VariableFilter(bricks=[Rectifier], roles=[OUTPUT])(cg)
cg = apply_dropout(cg, relu_outputs, dropout)
cost, error_rate = cg.outputs
if weight_decay:
logger.debug("Apply weight decay {}".format(weight_decay))
cost += weight_decay * l2_norm(cg.parameters)
cost.name = 'cost'
# Validation
valid_probs = convnet.apply_5windows(single_x)
valid_cost = (CategoricalCrossEntropy().apply(single_y, valid_probs)
.copy(name='cost'))
valid_error_rate = (MisclassificationRate().apply(
single_y, valid_probs).copy(name='error_rate'))
model = Model([cost, error_rate])
if load_params:
logger.info("Loaded params from {}".format(load_params))
with open(load_params, 'r') as src:
model.set_parameter_values(load_parameters(src))
# Training stream with random cropping
train = DogsVsCats(("train",), subset=slice(None, 25000 - valid_examples, None))
train_str = DataStream(
train, iteration_scheme=ShuffledScheme(train.num_examples, batch_size))
train_str = add_transformers(train_str, random_crop=True)
# Validation stream without cropping
valid = DogsVsCats(("train",), subset=slice(25000 - valid_examples, None, None))
valid_str = DataStream(
valid, iteration_scheme=SequentialExampleScheme(valid.num_examples))
valid_str = add_transformers(valid_str)
if mode == 'train':
directory, _ = os.path.split(sys.argv[0])
env = dict(os.environ)
env['THEANO_FLAGS'] = 'floatX=float32'
port = numpy.random.randint(1025, 10000)
server = subprocess.Popen(
[directory + '/server.py',
str(25000 - valid_examples), str(batch_size), str(port)],
env=env, stderr=subprocess.STDOUT)
train_str = ServerDataStream(
('image_features', 'targets'), produces_examples=False,
port=port)
save_to_base, save_to_extension = os.path.splitext(save_to)
# Train with simple SGD
if algo == 'rmsprop':
step_rule = RMSProp(decay_rate=0.999, learning_rate=0.0003)
elif algo == 'adam':
step_rule = Adam()
else:
assert False
if max_norm:
conv_params = VariableFilter(bricks=[Convolutional], roles=[WEIGHT])(cg)
linear_params = VariableFilter(bricks=[Linear], roles=[WEIGHT])(cg)
step_rule = CompositeRule(
[step_rule,
Restrict(VariableClipping(max_norm, axis=0), linear_params),
Restrict(VariableClipping(max_norm, axis=(1, 2, 3)), conv_params)])
algorithm = GradientDescent(
cost=cost, parameters=model.parameters,
step_rule=step_rule)
algorithm.add_updates(extra_updates)
# `Timing` extension reports time for reading data, aggregating a batch
# and monitoring;
# `ProgressBar` displays a nice progress bar during training.
extensions = [Timing(every_n_batches=100),
FinishAfter(after_n_epochs=num_epochs,
after_n_batches=num_batches),
DataStreamMonitoring(
[valid_cost, valid_error_rate],
valid_str,
prefix="valid"),
TrainingDataMonitoring(
[cost, error_rate,
aggregation.mean(algorithm.total_gradient_norm)],
prefix="train",
after_epoch=True),
TrackTheBest("valid_error_rate"),
Checkpoint(save_to, save_separately=['log'],
parameters=cg.parameters +
(population_statistics if batch_norm else []),
before_training=True, after_epoch=True)
.add_condition(
['after_epoch'],
OnLogRecord("valid_error_rate_best_so_far"),
(save_to_base + '_best' + save_to_extension,)),
Printing(every_n_batches=100)]
model = Model(cost)
main_loop = MainLoop(
algorithm,
train_str,
model=model,
extensions=extensions)
try:
main_loop.run()
finally:
server.terminate()
elif mode == 'test':
classify = theano.function([single_x], valid_probs.argmax())
test = DogsVsCats((test_set,))
test_str = DataStream(
test, iteration_scheme=SequentialExampleScheme(test.num_examples))
test_str = add_transformers(test_str)
correct = 0
with open(save_to, 'w') as dst:
print("id", "label", sep=',', file=dst)
for index, example in enumerate(test_str.get_epoch_iterator()):
image = example[0]
prediction = classify(image)
print(index + 1, classify(image), sep=',', file=dst)
if len(example) > 1 and prediction == example[1]:
correct += 1
print(correct / float(test.num_examples))
else:
assert False
def train(cli_params):
fn = 'noname'
if 'save_to' in nodefaultargs or not cli_params.get('load_from'):
fn = cli_params['save_to']
cli_params['save_dir'] = prepare_dir(fn)
nodefaultargs.append('save_dir')
logfile = os.path.join(cli_params['save_dir'], 'log.txt')
# Log also DEBUG to a file
fh = logging.FileHandler(filename=logfile)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
logger.info('Logging into %s' % logfile)
p, loaded = load_and_log_params(cli_params)
in_dim, data, whiten, cnorm = setup_data(p, test_set=False)
if not loaded:
# Set the zero layer to match input dimensions
p.encoder_layers = (in_dim, ) + p.encoder_layers
ladder = setup_model(p)
# Training
all_params = ComputationGraph([ladder.costs.total]).parameters
logger.info('Found the following parameters: %s' % str(all_params))
# Fetch all batch normalization updates. They are in the clean path.
# you can turn off BN by setting is_normalizing = False in ladder.py
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
assert not bn_updates or 'counter' in [u.name for u in bn_updates.keys()], \
'No batch norm params in graph - the graph has been cut?'
training_algorithm = GradientDescent(
cost=ladder.costs.total,
parameters=all_params,
step_rule=Adam(learning_rate=ladder.lr))
# In addition to actual training, also do BN variable approximations
if bn_updates:
training_algorithm.add_updates(bn_updates)
short_prints = {
"train":
OrderedDict([
('T_E', ladder.error.clean),
('T_O', ladder.oos.clean),
('T_C_class', ladder.costs.class_corr),
('T_C_de', ladder.costs.denois.values()),
('T_T', ladder.costs.total),
]),
"valid_approx":
OrderedDict([
('V_C_class', ladder.costs.class_clean),
('V_E', ladder.error.clean),
('V_O', ladder.oos.clean),
('V_C_de', ladder.costs.denois.values()),
('V_T', ladder.costs.total),
]),
"valid_final":
OrderedDict([
('VF_C_class', ladder.costs.class_clean),
('VF_E', ladder.error.clean),
('VF_O', ladder.oos.clean),
('VF_C_de', ladder.costs.denois.values()),
('V_T', ladder.costs.total),
]),
}
if len(data.valid_ind):
main_loop = MainLoop(
training_algorithm,
# Datastream used for training
make_datastream(data.train,
data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
n_unlabeled=p.unlabeled_samples,
whiten=whiten,
cnorm=cnorm,
balanced_classes=p.balanced_classes,
dseed=p.dseed),
model=Model(ladder.costs.total),
extensions=[
FinishAfter(after_n_epochs=p.num_epochs),
# This will estimate the validation error using
# running average estimates of the batch normalization
# parameters, mean and variance
ApproxTestMonitoring([
ladder.costs.class_clean, ladder.error.clean,
ladder.oos.clean, ladder.costs.total
] + ladder.costs.denois.values(),
make_datastream(
data.valid,
data.valid_ind,
p.valid_batch_size,
whiten=whiten,
cnorm=cnorm,
balanced_classes=p.balanced_classes,
scheme=ShuffledScheme),
prefix="valid_approx"),
# This Monitor is slower, but more accurate since it will first
# estimate batch normalization parameters from training data and
# then do another pass to calculate the validation error.
FinalTestMonitoring(
[
ladder.costs.class_clean, ladder.error.clean,
ladder.oos.clean, ladder.costs.total
] + ladder.costs.denois.values(),
make_datastream(data.train,
data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
whiten=whiten,
cnorm=cnorm,
balanced_classes=p.balanced_classes,
scheme=ShuffledScheme),
make_datastream(data.valid,
data.valid_ind,
p.valid_batch_size,
n_labeled=len(data.valid_ind),
whiten=whiten,
cnorm=cnorm,
balanced_classes=p.balanced_classes,
scheme=ShuffledScheme),
prefix="valid_final",
after_n_epochs=p.num_epochs,
after_training=True),
TrainingDataMonitoring([
ladder.error.clean, ladder.oos.clean, ladder.costs.total,
ladder.costs.class_corr,
training_algorithm.total_gradient_norm
] + ladder.costs.denois.values(),
prefix="train",
after_epoch=True),
# ladder.costs.class_clean - save model whenever we have best validation result another option `('train',ladder.costs.total)`
SaveParams(('valid_approx', ladder.error.clean),
all_params,
p.save_dir,
after_epoch=True),
SaveExpParams(p, p.save_dir, before_training=True),
SaveLog(p.save_dir, after_training=True),
ShortPrinting(short_prints),
LRDecay(ladder.lr,
p.num_epochs * p.lrate_decay,
p.num_epochs,
lrmin=p.lrmin,
after_epoch=True),
])
else:
main_loop = MainLoop(
training_algorithm,
# Datastream used for training
make_datastream(data.train,
data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
n_unlabeled=p.unlabeled_samples,
whiten=whiten,
cnorm=cnorm,
balanced_classes=p.balanced_classes,
dseed=p.dseed),
model=Model(ladder.costs.total),
extensions=[
FinishAfter(after_n_epochs=p.num_epochs),
TrainingDataMonitoring([
ladder.error.clean, ladder.oos.clean, ladder.costs.total,
ladder.costs.class_corr,
training_algorithm.total_gradient_norm
] + ladder.costs.denois.values(),
prefix="train",
after_epoch=True),
# ladder.costs.class_clean - save model whenever we have best validation result another option `('train',ladder.costs.total)`
SaveParams(('train', ladder.error.clean),
all_params,
p.save_dir,
after_epoch=True),
SaveExpParams(p, p.save_dir, before_training=True),
SaveLog(p.save_dir, after_training=True),
ShortPrinting(short_prints),
LRDecay(ladder.lr,
p.num_epochs * p.lrate_decay,
p.num_epochs,
lrmin=p.lrmin,
after_epoch=True),
])
main_loop.run()
# Get results
if len(data.valid_ind) == 0:
return None
df = DataFrame.from_dict(main_loop.log, orient='index')
col = 'valid_final_error_rate_clean'
logger.info('%s %g' % (col, df[col].iloc[-1]))
if main_loop.log.status['epoch_interrupt_received']:
return None
return df
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()
def main(name, epochs, batch_size, learning_rate, window_size, conv_sizes, num_filters, fc_dim, enc_dim, dec_dim, step, num_digits, num_classes,
oldmodel, live_plotting):
channels, img_height, img_width = 1, 100, 100
rnninits = {
'weights_init': Uniform(width=0.02),
'biases_init': Constant(0.),
}
inits = {
'weights_init': IsotropicGaussian(0.001),
'biases_init': Constant(0.),
}
rec_inits = {
'weights_init': IsotropicGaussian(0.001),
'biases_init': Constant(0.),
}
convinits = {
'weights_init': Uniform(width=.2),
'biases_init': Constant(0.),
}
n_iter = step * num_digits
filter_size1, filter_size2 = zip(conv_sizes, conv_sizes)[:]
w_height, w_width = window_size.split(',')
w_height = int(w_height)
w_width = int(w_width)
subdir = time.strftime("%Y-%m-%d") + "-" + name
if not os.path.exists(subdir):
os.makedirs(subdir)
lines = ["\n Running experiment",
" subdirectory: %s" % subdir,
" learning rate: %g" % learning_rate,
" attention size: %s" % window_size,
" n_iterations: %d" % n_iter,
" encoder dimension: %d" % enc_dim,
" decoder dimension: %d" % dec_dim,
" batch size: %d" % batch_size,
" epochs: %d" % epochs,
]
for line in lines:
print(line)
print()
rectifier = Rectifier()
conv1 = Convolutional(filter_size=filter_size2, num_filters=int(num_filters / 2), num_channels=channels, image_size=(w_height, w_width), border_mode='half',
name='conv1', **convinits)
conv1_bn = SpatialBatchNormalization(input_dim=(64, 26, 26), conserve_memory=False, n_iter=n_iter, name='conv1_bn')
conv2 = Convolutional(filter_size=filter_size2, num_channels=int(num_filters / 2), num_filters=int(num_filters / 2), image_size=(26, 26), name='conv2',
**convinits)
conv2_bn = SpatialBatchNormalization(input_dim=(64, 24, 24), conserve_memory=False, n_iter=n_iter, name='conv2_bn')
max_pooling = MaxPooling(pooling_size=(2, 2), step=(2, 2))
conv3 = Convolutional(filter_size=filter_size2, num_filters=num_filters, num_channels=int(num_filters / 2), image_size=(12, 12), border_mode='half',
name='conv3', **convinits)
conv3_bn = SpatialBatchNormalization(input_dim=(128, 12, 12), conserve_memory=False, n_iter=n_iter, name='conv3_bn')
conv4 = Convolutional(filter_size=filter_size2, num_filters=num_filters, num_channels=num_filters, image_size=(12, 12), border_mode='half', name='conv4',
**convinits)
conv4_bn = SpatialBatchNormalization(input_dim=(128, 12, 12), conserve_memory=False, n_iter=n_iter, name='conv4_bn')
# Max Pooling
conv5 = Convolutional(filter_size=filter_size2, num_filters=160, num_channels=num_filters, image_size=(6, 6), border_mode='half', name='conv5',
**convinits)
conv5_bn = SpatialBatchNormalization(input_dim=(160, 6, 6), conserve_memory=False, n_iter=n_iter, name='conv5_bn')
conv6 = Convolutional(filter_size=filter_size2, num_filters=192, num_channels=160, image_size=(6, 6), name='conv6', **convinits)
conv6_bn = SpatialBatchNormalization(input_dim=(192, 4, 4), conserve_memory=False, n_iter=n_iter, name='conv6_bn')
conv_mlp = MLP(activations=[Identity()], dims=[3072, fc_dim], name="MLP_conv", **inits)
conv_mlp_bn = BatchNormalization(input_dim=fc_dim, conserve_memory=False, n_iter=n_iter, name='conv_mlp_bn')
loc_mlp = MLP(activations=[Identity()], dims=[6, fc_dim], name="MLP_loc", **inits)
loc_mlp_bn = BatchNormalization(input_dim=fc_dim, conserve_memory=False, n_iter=n_iter, name='loc_mlp_bn')
encoder_mlp = MLP([Identity()], [fc_dim, 4 * enc_dim], name="MLP_enc", **rec_inits)
decoder_mlp = MLP([Identity()], [enc_dim, 4 * dec_dim], name="MLP_dec", **rec_inits)
encoder_rnn = LSTM(activation=Tanh(), dim=enc_dim, name="RNN_enc", **rnninits)
conv_init = ConvolutionalSequence(
[Convolutional(filter_size=filter_size1, num_filters=int(num_filters / 8), name='conv1_init'),
SpatialBatchNormalization(conserve_memory=False, name='conv1_bn_init'),
Convolutional(filter_size=filter_size2, num_filters=int(num_filters / 8), name='conv2_init'),
SpatialBatchNormalization(conserve_memory=False, name='conv2_bn_init'),
Convolutional(filter_size=filter_size2, num_filters=int(num_filters / 4), name='conv3_init'),
SpatialBatchNormalization(conserve_memory=False, name='conv3_bn_init'),
], image_size=(12, 12), num_channels=channels, name='conv_seq_init', **convinits)
decoder_rnn = LSTM(activation=Tanh(), dim=dec_dim, name="RNN_dec", **rnninits)
emit_mlp = MLP(activations=[Tanh()], dims=[dec_dim, 6], name='emit_mlp', weights_init=Constant(0.),
biases_init=Constant((1., 0., 0., 0., 1., 0.)))
classification_mlp1 = MLP(activations=[Identity()], dims=[enc_dim, fc_dim], name='MPL_class1', **inits)
classification_mlp1_bn = BatchNormalization(input_dim=fc_dim, conserve_memory=False, n_iter=n_iter, name='classification_mlp1_bn')
classification_mlp2 = MLP(activations=[Identity()], dims=[fc_dim, fc_dim], name='MPL_class2', **inits)
classification_mlp2_bn = BatchNormalization(input_dim=fc_dim, conserve_memory=False, n_iter=n_iter, name='classification_mlp2_bn')
classification_mlp3 = MLP(activations=[Softmax()], dims=[fc_dim, num_classes], name='MPL_class3', **inits)
edram = EDRAM(channels=channels, out_height=w_height, out_width=w_width, n_iter=n_iter, num_classes=num_classes, rectifier=rectifier, conv1=conv1,
conv1_bn=conv1_bn, conv2=conv2, conv2_bn=conv2_bn, max_pooling=max_pooling, conv3=conv3, conv3_bn=conv3_bn, conv4=conv4, conv4_bn=conv4_bn,
conv5=conv5, conv5_bn=conv5_bn, conv6=conv6, conv6_bn=conv6_bn, conv_mlp=conv_mlp, conv_mlp_bn=conv_mlp_bn,
loc_mlp=loc_mlp, loc_mlp_bn=loc_mlp_bn, conv_init=conv_init, encoder_mlp=encoder_mlp, encoder_rnn=encoder_rnn, decoder_mlp=decoder_mlp,
decoder_rnn=decoder_rnn, classification_mlp1=classification_mlp1, classification_mlp1_bn=classification_mlp1_bn,
classification_mlp2=classification_mlp2, classification_mlp2_bn=classification_mlp2_bn, classification_mlp3=classification_mlp3,
emit_mlp=emit_mlp)
edram.initialize()
# ------------------------------------------------------------------------
x = T.ftensor4('features')
x_coarse = T.ftensor4('features_coarse')
y = T.ivector('labels')
wr = T.fmatrix('locations')
with batch_normalization(edram):
bn_p, bn_l, m_c1_bn, s_c1_bn, m_c2_bn, s_c2_bn, m_c3_bn, s_c3_bn, m_c4_bn, s_c4_bn, m_c5_bn, s_c5_bn, m_c6_bn, s_c6_bn, \
m_c_bn, s_c_bn, m_l_bn, s_l_bn, m_cl1_bn, s_cl1_bn, m_cl2_bn, s_cl2_bn = edram.calculate_train(x, x_coarse)
def compute_cost(p, wr, y, l):
cost_where = T.dot(T.sqr(wr - l), [1, 0.5, 1, 0.5, 1, 1])
cost_y = T.stack([T.nnet.categorical_crossentropy(T.maximum(p[i, :], 1e-7), y) for i in range(0, n_iter)])
return cost_where, cost_y
cost_where, cost_y = compute_cost(bn_p, wr, y, bn_l)
bn_cost = cost_y + cost_where
bn_cost = bn_cost.sum(axis=0)
bn_cost = bn_cost.mean()
bn_cost.name = 'cost'
bn_error_rate = MisclassificationRate().apply(y, bn_p[-1])
bn_error_rate.name = 'error_rate'
# ------------------------------------------------------------
bn_cg = ComputationGraph([bn_cost, bn_error_rate])
# Prepare algorithm
algorithm = GradientDescent(
cost=bn_cg.outputs[0],
on_unused_sources='ignore',
parameters=bn_cg.parameters,
step_rule=CompositeRule([
RemoveNotFinite(),
StepClipping(10.),
Adam(learning_rate)
])
)
pop_updates = get_batch_normalization_updates(bn_cg)
update_params = [conv1_bn.population_mean, conv1_bn.population_stdev, conv2_bn.population_mean, conv2_bn.population_stdev, conv3_bn.population_mean,
conv3_bn.population_stdev, conv4_bn.population_mean, conv4_bn.population_stdev, conv5_bn.population_mean, conv5_bn.population_stdev,
conv6_bn.population_mean, conv6_bn.population_stdev, conv_mlp_bn.population_mean, conv_mlp_bn.population_stdev,
loc_mlp_bn.population_mean, loc_mlp_bn.population_stdev, classification_mlp1_bn.population_mean, classification_mlp1_bn.population_stdev,
classification_mlp2_bn.population_mean, classification_mlp2_bn.population_stdev]
update_values = [m_c1_bn, s_c1_bn, m_c2_bn, s_c2_bn, m_c3_bn, s_c3_bn, m_c4_bn, s_c4_bn, m_c5_bn, s_c5_bn, m_c6_bn, s_c6_bn, m_c_bn, s_c_bn, m_l_bn, s_l_bn,
m_cl1_bn, s_cl1_bn, m_cl2_bn, s_cl2_bn]
pop_updates.extend([(p, m) for p, m in zip(update_params, update_values)])
decay_rate = 0.05
extra_updates = [(p, m * decay_rate + p * (1 - decay_rate)) for p, m in pop_updates]
algorithm.add_updates(extra_updates)
# ------------------------------------------------------------------------
# Setup monitors
p, l = edram.calculate_test(x, x_coarse)
cost_where, cost_y = compute_cost(p, wr, y, l)
cost = cost_y + cost_where
cost = cost.sum(axis=0)
cost = cost.mean()
cost.name = 'cost'
error_rate = MisclassificationRate().apply(y, p[-1])
error_rate.name = 'error_rate'
monitors = [cost, error_rate]
plotting_extensions = []
# Live plotting...
if live_plotting:
plot_channels = [
['train_cost', 'test_cost'],
['train_error_rate', 'test_error_rate'],
]
plotting_extensions = [
Plot(subdir, channels=plot_channels, server_url='http://155.69.150.60:80/')
]
# ------------------------------------------------------------
mnist_cluttered_train = MNISTCluttered(which_sets=['train'], sources=('features', 'locations', 'labels'))
mnist_cluttered_test = MNISTCluttered(which_sets=['test'], sources=('features', 'locations', 'labels'))
main_loop = MainLoop(
model=Model([bn_cost]),
data_stream=DataStream.default_stream(mnist_cluttered_train, iteration_scheme=ShuffledScheme(mnist_cluttered_train.num_examples, batch_size)),
algorithm=algorithm,
extensions=[Timing(),
FinishAfter(after_n_epochs=epochs),
DataStreamMonitoring(monitors,
DataStream.default_stream(mnist_cluttered_train, iteration_scheme=SequentialScheme(mnist_cluttered_train.num_examples,
batch_size)), prefix='train'),
DataStreamMonitoring(monitors,
DataStream.default_stream(mnist_cluttered_test,
iteration_scheme=SequentialScheme(mnist_cluttered_test.num_examples, batch_size)),
prefix="test"),
PartsOnlyCheckpoint("{}/{}".format(subdir, name), before_training=False, after_epoch=True, save_separately=['log', ]),
TrackTheBest('test_error_rate', 'best_test_error_rate'),
BestCheckpount("{}/{}".format(subdir, name), 'best_test_error_rate', save_separately=['model', ]),
Printing(),
ProgressBar(),
PrintingTo("\n".join(lines), "{}/{}_log.txt".format(subdir, name)), ] + plotting_extensions)
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_parameter_values(oldmodel.get_parameter_values())
main_loop.model.get_top_bricks()[0].conv1_bn.population_mean.set_value(oldmodel.get_top_bricks()[0].conv1_bn.population_mean.get_value())
main_loop.model.get_top_bricks()[0].conv1_bn.population_stdev.set_value(oldmodel.get_top_bricks()[0].conv1_bn.population_stdev.get_value())
main_loop.model.get_top_bricks()[0].conv2_bn.population_mean.set_value(oldmodel.get_top_bricks()[0].conv2_bn.population_mean.get_value())
main_loop.model.get_top_bricks()[0].conv2_bn.population_stdev.set_value(oldmodel.get_top_bricks()[0].conv2_bn.population_stdev.get_value())
main_loop.model.get_top_bricks()[0].conv3_bn.population_mean.set_value(oldmodel.get_top_bricks()[0].conv3_bn.population_mean.get_value())
main_loop.model.get_top_bricks()[0].conv3_bn.population_stdev.set_value(oldmodel.get_top_bricks()[0].conv3_bn.population_stdev.get_value())
main_loop.model.get_top_bricks()[0].conv4_bn.population_mean.set_value(oldmodel.get_top_bricks()[0].conv4_bn.population_mean.get_value())
main_loop.model.get_top_bricks()[0].conv4_bn.population_stdev.set_value(oldmodel.get_top_bricks()[0].conv4_bn.population_stdev.get_value())
main_loop.model.get_top_bricks()[0].conv5_bn.population_mean.set_value(oldmodel.get_top_bricks()[0].conv5_bn.population_mean.get_value())
main_loop.model.get_top_bricks()[0].conv5_bn.population_stdev.set_value(oldmodel.get_top_bricks()[0].conv5_bn.population_stdev.get_value())
main_loop.model.get_top_bricks()[0].conv6_bn.population_mean.set_value(oldmodel.get_top_bricks()[0].conv6_bn.population_mean.get_value())
main_loop.model.get_top_bricks()[0].conv6_bn.population_stdev.set_value(oldmodel.get_top_bricks()[0].conv6_bn.population_stdev.get_value())
main_loop.model.get_top_bricks()[0].loc_mlp_bn.population_mean.set_value(oldmodel.get_top_bricks()[0].loc_mlp_bn.population_mean.get_value())
main_loop.model.get_top_bricks()[0].loc_mlp_bn.population_stdev.set_value(oldmodel.get_top_bricks()[0].loc_mlp_bn.population_stdev.get_value())
main_loop.model.get_top_bricks()[0].conv_mlp_bn.population_mean.set_value(oldmodel.get_top_bricks()[0].conv_mlp_bn.population_mean.get_value())
main_loop.model.get_top_bricks()[0].conv_mlp_bn.population_stdev.set_value(oldmodel.get_top_bricks()[0].conv_mlp_bn.population_stdev.get_value())
main_loop.model.get_top_bricks()[0].classification_mlp1_bn.population_mean.set_value(
oldmodel.get_top_bricks()[0].classification_mlp1_bn.population_mean.get_value())
main_loop.model.get_top_bricks()[0].classification_mlp1_bn.population_stdev.set_value(
oldmodel.get_top_bricks()[0].classification_mlp1_bn.population_stdev.get_value())
main_loop.model.get_top_bricks()[0].classification_mlp2_bn.population_mean.set_value(
oldmodel.get_top_bricks()[0].classification_mlp2_bn.population_mean.get_value())
main_loop.model.get_top_bricks()[0].classification_mlp2_bn.population_stdev.set_value(
oldmodel.get_top_bricks()[0].classification_mlp2_bn.population_stdev.get_value())
main_loop.model.get_top_bricks()[0].conv_init.layers[1].population_mean.set_value(
oldmodel.get_top_bricks()[0].conv_init.layers[1].population_mean.get_value())
main_loop.model.get_top_bricks()[0].conv_init.layers[1].population_stdev.set_value(
oldmodel.get_top_bricks()[0].conv_init.layers[1].population_stdev.get_value())
main_loop.model.get_top_bricks()[0].conv_init.layers[3].population_mean.set_value(
oldmodel.get_top_bricks()[0].conv_init.layers[3].population_mean.get_value())
main_loop.model.get_top_bricks()[0].conv_init.layers[3].population_stdev.set_value(
oldmodel.get_top_bricks()[0].conv_init.layers[3].population_stdev.get_value())
main_loop.model.get_top_bricks()[0].conv_init.layers[5].population_mean.set_value(
oldmodel.get_top_bricks()[0].conv_init.layers[5].population_mean.get_value())
main_loop.model.get_top_bricks()[0].conv_init.layers[5].population_stdev.set_value(
oldmodel.get_top_bricks()[0].conv_init.layers[5].population_stdev.get_value())
del oldmodel
main_loop.run()
def train(cli_params):
cli_params['save_dir'] = prepare_dir(cli_params['save_to'])
logfile = os.path.join(cli_params['save_dir'], 'log.txt')
# Log also DEBUG to a file
fh = logging.FileHandler(filename=logfile)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
logger.info('Logging into %s' % logfile)
p, loaded = load_and_log_params(cli_params)
in_dim, data, whiten, cnorm = setup_data(p, test_set=False)
if not loaded:
# Set the zero layer to match input dimensions
p.encoder_layers = (in_dim, ) + p.encoder_layers
ladder = setup_model(p)
# Training
all_params = ComputationGraph([ladder.costs.total]).parameters
logger.info('Found the following parameters: %s' % str(all_params))
# Fetch all batch normalization updates. They are in the clean path.
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
assert 'counter' in [u.name for u in bn_updates.keys()], \
'No batch norm params in graph - the graph has been cut?'
training_algorithm = GradientDescent(
cost=ladder.costs.total,
params=all_params,
step_rule=Adam(learning_rate=ladder.lr))
# In addition to actual training, also do BN variable approximations
training_algorithm.add_updates(bn_updates)
model = Model(ladder.costs.total)
monitored_variables = [
ladder.costs.class_corr,
ladder.costs.class_clean,
ladder.error,
# training_algorithm.total_gradient_norm,
ladder.costs.total] \
# + ladder.costs.denois.values()
# Make a global history recorder so that we can get summary at end of
# training when we write to Sentinel
# global_history records all relevant monitoring vars
# updated by SaveLog every time
global_history = {}
main_loop = MainLoop(
training_algorithm,
# Datastream used for training
make_datastream(data.train,
data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
n_unlabeled=p.unlabeled_samples,
whiten=whiten,
cnorm=cnorm),
model=model,
extensions=[
FinishAfter(after_n_epochs=p.num_epochs),
# write out to sentinel file for experiment automator to work
SentinelWhenFinish(save_dir=p.save_dir,
global_history=global_history),
# This will estimate the validation error using
# running average estimates of the batch normalization
# parameters, mean and variance
ApproxTestMonitoring(monitored_variables,
make_datastream(data.valid,
data.valid_ind,
p.valid_batch_size,
whiten=whiten,
cnorm=cnorm,
scheme=ShuffledScheme),
prefix="valid_approx"),
# This Monitor is slower, but more accurate since it will first
# estimate batch normalization parameters from training data and
# then do another pass to calculate the validation error.
FinalTestMonitoring(monitored_variables,
make_datastream(data.train,
data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
whiten=whiten,
cnorm=cnorm,
scheme=ShuffledScheme),
make_datastream(data.valid,
data.valid_ind,
p.valid_batch_size,
n_labeled=len(data.valid_ind),
whiten=whiten,
cnorm=cnorm,
scheme=ShuffledScheme),
prefix="valid_final",
after_n_epochs=p.num_epochs),
TrainingDataMonitoring(variables=monitored_variables,
prefix="train",
after_epoch=True),
SaveParams('valid_approx_cost_class_corr', model, p.save_dir),
# SaveParams(None, all_params, p.save_dir, after_epoch=True),
SaveExpParams(p, p.save_dir, before_training=True),
SaveLog(save_dir=p.save_dir,
after_epoch=True,
global_history=global_history),
Printing(),
# ShortPrinting(short_prints),
LRDecay(ladder.lr,
p.num_epochs * p.lrate_decay,
p.num_epochs,
after_epoch=True),
])
main_loop.run()
# Get results
df = main_loop.log.to_dataframe()
col = 'valid_final_error_rate'
logger.info('%s %g' % (col, df[col].iloc[-1]))
if main_loop.log.status['epoch_interrupt_received']:
return None
return df
def run(batch_size, save_path, z_dim, oldmodel, discriminative_regularization,
classifier, vintage, monitor_every, monitor_before, checkpoint_every, dataset, color_convert,
image_size, net_depth, subdir,
reconstruction_factor, kl_factor, discriminative_factor, disc_weights,
num_epochs):
if dataset:
streams = create_custom_streams(filename=dataset,
training_batch_size=batch_size,
monitoring_batch_size=batch_size,
include_targets=False,
color_convert=color_convert)
else:
streams = create_celeba_streams(training_batch_size=batch_size,
monitoring_batch_size=batch_size,
include_targets=False)
main_loop_stream, train_monitor_stream, valid_monitor_stream = streams[:3]
# Compute parameter updates for the batch normalization population
# statistics. They are updated following an exponential moving average.
rval = create_training_computation_graphs(
z_dim, image_size, net_depth, discriminative_regularization, classifier,
vintage, reconstruction_factor, kl_factor, discriminative_factor, disc_weights)
cg, bn_cg, variance_parameters = rval
pop_updates = list(
set(get_batch_normalization_updates(bn_cg, allow_duplicates=True)))
decay_rate = 0.05
extra_updates = [(p, m * decay_rate + p * (1 - decay_rate))
for p, m in pop_updates]
model = Model(bn_cg.outputs[0])
selector = Selector(
find_bricks(
model.top_bricks,
lambda brick: brick.name in ('encoder_convnet', 'encoder_mlp',
'decoder_convnet', 'decoder_mlp')))
parameters = list(selector.get_parameters().values()) + variance_parameters
# Prepare algorithm
step_rule = Adam()
algorithm = GradientDescent(cost=bn_cg.outputs[0],
parameters=parameters,
step_rule=step_rule)
algorithm.add_updates(extra_updates)
# Prepare monitoring
sys.setrecursionlimit(1000000)
monitored_quantities_list = []
for graph in [bn_cg, cg]:
# cost, kl_term, reconstruction_term, discriminative_term = graph.outputs
cost, kl_term, reconstruction_term, discriminative_term = graph.outputs[:4]
discriminative_layer_terms = graph.outputs[4:]
cost.name = 'nll_upper_bound'
avg_kl_term = kl_term.mean(axis=0)
avg_kl_term.name = 'avg_kl_term'
avg_reconstruction_term = -reconstruction_term.mean(axis=0)
avg_reconstruction_term.name = 'avg_reconstruction_term'
avg_discriminative_term = discriminative_term.mean(axis=0)
avg_discriminative_term.name = 'avg_discriminative_term'
num_layer_terms = len(discriminative_layer_terms)
avg_discriminative_layer_terms = [None] * num_layer_terms
for i, term in enumerate(discriminative_layer_terms):
avg_discriminative_layer_terms[i] = discriminative_layer_terms[i].mean(axis=0)
avg_discriminative_layer_terms[i].name = "avg_discriminative_term_layer_{:02d}".format(i)
monitored_quantities_list.append(
[cost, avg_kl_term, avg_reconstruction_term,
avg_discriminative_term] + avg_discriminative_layer_terms)
train_monitoring = DataStreamMonitoring(
monitored_quantities_list[0], train_monitor_stream, prefix="train",
updates=extra_updates, after_epoch=False, before_first_epoch=monitor_before,
every_n_epochs=monitor_every)
valid_monitoring = DataStreamMonitoring(
monitored_quantities_list[1], valid_monitor_stream, prefix="valid",
after_epoch=False, before_first_epoch=monitor_before,
every_n_epochs=monitor_every)
# Prepare checkpoint
checkpoint = Checkpoint(save_path, every_n_epochs=checkpoint_every,
before_training=True, use_cpickle=True)
sample_checkpoint = SampleCheckpoint(interface=DiscGenModel, z_dim=z_dim/2,
image_size=(image_size, image_size), channels=3,
dataset=dataset, split="valid", save_subdir=subdir,
before_training=True, after_epoch=True)
# TODO: why does z_dim=foo become foo/2?
extensions = [Timing(),
FinishAfter(after_n_epochs=num_epochs),
checkpoint,
sample_checkpoint,
train_monitoring, valid_monitoring,
Printing(),
ProgressBar()]
main_loop = MainLoop(model=model, data_stream=main_loop_stream,
algorithm=algorithm, extensions=extensions)
if oldmodel is not None:
print("Initializing parameters with old model {}".format(oldmodel))
try:
saved_model = load(oldmodel)
except AttributeError:
# newer version of blocks
with open(oldmodel, 'rb') as src:
saved_model = load(src)
main_loop.model.set_parameter_values(
saved_model.model.get_parameter_values())
del saved_model
main_loop.run()
def train(ladder,
batch_size=100,
num_train_examples=60000,
num_epochs=150,
lrate_decay=0.67):
# Setting Logger
timestr = time.strftime("%Y_%m_%d_at_%H_%M")
save_path = 'results/mnist_100_standard_' + 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
model = Model(ladder.costs.total)
all_params = model.parameters
print len(all_params)
print all_params
training_algorithm = GradientDescent(
cost=ladder.costs.total,
parameters=all_params,
step_rule=Adam(learning_rate=ladder.lr))
# Fetch all batch normalization updates. They are in the clean path.
# In addition to actual training, also do BN variable approximations
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
training_algorithm.add_updates(bn_updates)
monitored_variables = [
ladder.costs.class_corr, ladder.costs.class_clean, ladder.error,
training_algorithm.total_gradient_norm, ladder.costs.total
] + ladder.costs.denois.values()
train_data_stream, test_data_stream = get_mixed_streams(batch_size)
train_monitoring = TrainingDataMonitoring(variables=monitored_variables,
prefix="train",
after_epoch=True)
valid_monitoring = DataStreamMonitoring(variables=monitored_variables,
data_stream=test_data_stream,
prefix="test",
after_epoch=True)
main_loop = MainLoop(algorithm=training_algorithm,
data_stream=train_data_stream,
model=model,
extensions=[
train_monitoring, valid_monitoring,
FinishAfter(after_n_epochs=num_epochs),
SaveParams('test_CE_corr', model, save_path),
SaveLog(save_path, after_epoch=True),
LRDecay(lr=ladder.lr,
decay_first=num_epochs * lrate_decay,
decay_last=num_epochs,
after_epoch=True),
Printing()
])
main_loop.run()
def train_model(cost, cross_entropy, updates,
train_stream, valid_stream, args, gate_values=None):
step_rule = learning_algorithm(args)
cg = ComputationGraph(cost)
# ADD REGULARIZATION
# WEIGHT NOISE
weight_noise = args.weight_noise
if weight_noise > 0:
weights = VariableFilter(roles=[WEIGHT])(cg.variables)
cg_train = apply_noise(cg, weights, weight_noise)
cost = cg_train.outputs[0]
cost.name = "cost_with_weight_noise"
cg = ComputationGraph(cost)
logger.info(cg.parameters)
algorithm = GradientDescent(cost=cost, step_rule=step_rule,
params=cg.parameters)
algorithm.add_updates(updates)
# extensions to be added
extensions = []
if args.load_path is not None:
extensions.append(Load(args.load_path))
outputs = [
variable for variable in cg.variables if variable.name == "presoft"]
if args.generate:
extensions.append(TextGenerationExtension(
outputs=outputs,
generation_length=args.generated_text_lenght,
initial_text_length=args.initial_text_length,
every_n_batches=args.monitoring_freq,
ploting_path=os.path.join(args.save_path, 'prob_plot.png'),
softmax_sampling=args.softmax_sampling,
dataset=args.dataset,
updates=updates,
interactive_mode=args.interactive_mode))
extensions.extend([
TrainingDataMonitoring([cost], prefix='train',
every_n_batches=args.monitoring_freq,
after_epoch=True),
DataStreamMonitoring([cost, cross_entropy],
valid_stream, args.mini_batch_size_valid,
state_updates=updates,
prefix='valid',
before_first_epoch=not(args.visualize_gates),
every_n_batches=args.monitoring_freq),
ResetStates([v for v, _ in updates], every_n_batches=100),
ProgressBar()])
# Creating directory for saving model.
if not args.interactive_mode:
if not os.path.exists(args.save_path):
os.makedirs(args.save_path)
else:
raise Exception('Directory already exists')
early_stopping = EarlyStopping('valid_cross_entropy',
args.patience, args.save_path,
every_n_batches=args.monitoring_freq)
# Visualizing extensions
if args.interactive_mode:
extensions.append(InteractiveMode())
if args.visualize_gates and (gate_values is not None):
if args.rnn_type == "lstm":
extensions.append(VisualizeGateLSTM(gate_values, updates,
args.dataset,
ploting_path=None))
elif args.rnn_type == "soft":
extensions.append(VisualizeGateSoft(gate_values, updates,
args.dataset,
ploting_path=None))
else:
assert(False)
extensions.append(early_stopping)
extensions.append(Printing(every_n_batches=args.monitoring_freq))
main_loop = MainLoop(
model=Model(cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=extensions
)
main_loop.run()
def train(ladder, batch_size=100, labeled_samples=100,
unlabeled_samples=50000, valid_set_size=10000,
num_epochs=150, valid_batch_size=100, lrate_decay=0.67,
save_path='results/mnist_100_full0'):
# Setting Logger
log_path = os.path.join(save_path, 'log.txt')
fh = logging.FileHandler(filename=log_path)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
logger.info('Logging into %s' % log_path)
# Training
all_params = ComputationGraph([ladder.costs.total]).parameters
logger.info('Found the following parameters: %s' % str(all_params))
training_algorithm = GradientDescent(
cost=ladder.costs.total, params=all_params,
step_rule=Adam(learning_rate=ladder.lr))
# Fetch all batch normalization updates. They are in the clean path.
# In addition to actual training, also do BN variable approximations
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
training_algorithm.add_updates(bn_updates)
monitored_variables = [
ladder.costs.class_corr, ladder.costs.class_clean,
ladder.error, training_algorithm.total_gradient_norm,
ladder.costs.total] + ladder.costs.denois.values()
data = get_mnist_data_dict(unlabeled_samples=unlabeled_samples,
valid_set_size=valid_set_size)
train_data_stream = make_datastream(
data.train, data.train_ind, batch_size,
n_labeled=labeled_samples,
n_unlabeled=unlabeled_samples)
valid_data_stream = make_datastream(
data.valid, data.valid_ind, valid_batch_size,
n_labeled=len(data.valid_ind),
n_unlabeled=len(data.valid_ind))
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=Model(ladder.costs.total),
extensions=[
train_monitoring,
valid_monitoring,
FinishAfter(after_n_epochs=num_epochs),
SaveParams(None, all_params, save_path, after_epoch=True),
SaveLog(save_path, after_training=True),
LRDecay(lr=ladder.lr,
decay_first=num_epochs * lrate_decay,
decay_last=num_epochs,
after_epoch=True),
Printing()])
main_loop.run()
batch_size=m)))
# uwaga: dyskryminator bierze 2m probek, pierwsze m to generowane, kolejne m to z danych
# observables.append(cost_discriminator)
generator_cost = theano.shared(value=np.array(0., dtype=np.float32), name='g_cost')
discriminator_cost = theano.shared(value=np.array(0., dtype=np.float32), name='d_cost')
generator_step_norm = theano.shared(value=np.array(0., dtype=np.float32), name='g_step_norm')
generator_grad_norm = theano.shared(value=np.array(0., dtype=np.float32), name='g_grad_norm')
discriminator_step_norm = theano.shared(value=np.array(0., dtype=np.float32), name='d_step_norm')
discriminator_grad_norm = theano.shared(value=np.array(0., dtype=np.float32), name='d_grad_norm')
discriminator_descent.add_updates([
(discriminator_cost, ComputationGraph(cost_discriminator).outputs[0]),
(discriminator_step_norm, discriminator_descent.total_step_norm),
(discriminator_grad_norm, discriminator_descent.total_gradient_norm)])
generator_descent.add_updates([
(generator_cost, ComputationGraph(cost_generator).outputs[0]),
(generator_step_norm, generator_descent.total_step_norm),
(generator_grad_norm, generator_descent.total_gradient_norm)])
observables = []
observables.append(generator_cost)
observables.append(discriminator_cost)
observables.append(generator_step_norm)
observables.append(generator_grad_norm)
observables.append(discriminator_step_norm)
observables.append(discriminator_grad_norm)
# Build datastream
train_stream = datastream.setup_datastream(config.dataset,
config.num_seqs,
config.seq_len,
config.seq_div_size)
# Build model
m = config.Model(config)
# Train the model
cg = Model(m.sgd_cost)
algorithm = GradientDescent(cost=m.sgd_cost,
step_rule=config.step_rule,
parameters=cg.parameters)
algorithm.add_updates(m.states)
monitor_vars = list(set(v for p in m.monitor_vars for v in p))
extensions = [
ProgressBar(),
TrainingDataMonitoring(
monitor_vars,
prefix='train', every_n_batches=config.monitor_freq),
Printing(every_n_batches=config.monitor_freq, after_epoch=False),
ResetStates([v for v, _ in m.states], after_epoch=True)
]
if plot_avail:
plot_channels = [['train_' + v.name for v in p] for p in m.monitor_vars]
extensions.append(
Plot(document='text_'+model_name,
def train_model(cost,
unregularized_cost,
updates,
train_stream,
valid_stream,
args,
gate_values=None):
step_rule = learning_algorithm(args)
cg = ComputationGraph(cost)
# ADD REGULARIZATION
# WEIGHT NOISE
weight_noise = args.weight_noise
if weight_noise > 0:
weights = VariableFilter(roles=[WEIGHT])(cg.variables)
cg_train = apply_noise(cg, weights, weight_noise)
cost = cg_train.outputs[0]
cost.name = "cost_with_weight_noise"
cg = ComputationGraph(cost)
logger.info(cg.parameters)
# Define algorithm
algorithm = GradientDescent(cost=cost,
step_rule=step_rule,
parameters=cg.parameters)
# Add the updates to carry the hidden state
algorithm.add_updates(updates)
# Extensions to be added
extensions = []
# Load from a dumped model
if args.load_path is not None:
if args.fine_tuning:
cost = fine_tuning(cost, args)
else:
extensions.append(Load(args.load_path))
# Generation extension
if args.generate:
extensions.append(
TextGenerationExtension(
cost=cost,
generation_length=args.generated_text_lenght,
initial_text_length=args.initial_text_length,
every_n_batches=1,
ploting_path=os.path.join(args.save_path, 'prob_plot.png'),
softmax_sampling=args.softmax_sampling,
dataset=args.dataset,
updates=updates,
interactive_mode=args.interactive_mode))
# Training and Validation score monitoring
extensions.extend([
TrainingDataMonitoring([cost],
prefix='train',
every_n_batches=args.monitoring_freq),
DataStreamMonitoring([cost, unregularized_cost],
valid_stream,
args.mini_batch_size_valid,
args.dataset,
state_updates=updates,
prefix='valid',
before_first_epoch=(args.visualize is None),
every_n_batches=args.monitoring_freq)
])
# Creating directory for saving model.
if not args.interactive_mode:
if not os.path.exists(args.save_path):
os.makedirs(args.save_path)
elif 'test' in args.save_path:
print("Rewriting in " + args.save_path)
else:
raise Exception('Directory already exists')
# Early stopping
extensions.append(
EarlyStopping('valid_' + unregularized_cost.name,
args.patience,
args.save_path,
every_n_batches=args.monitoring_freq))
# Printing
extensions.append(ProgressBar())
extensions.append(Printing(every_n_batches=args.monitoring_freq))
# Reset the initial states
if args.dataset == "sine":
reset_frequency = 1
else:
reset_frequency = 100
extensions.append(
ResetStates([v for v, _ in updates], every_n_batches=reset_frequency))
# Visualizing extensions
if args.interactive_mode:
extensions.append(InteractiveMode())
main_loop = MainLoop(model=Model(cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=extensions)
# This is where the magic happens!
main_loop.run()
def train_rnnrbm(train,
rnnrbm,
epochs=1000,
test=None,
bokeh=True,
load_path=None):
cdk = theano.shared(10)
lr = theano.shared(float32(0.004))
cost, v_sample = rnnrbm.cost(examples=x, mask=x_mask, k=cdk)
error_rate = MismulitclassificationRate().apply(x, v_sample[-1], x_mask)
error_rate.name = "error on note as a whole"
mistake_rate = MismulitmistakeRate().apply(x, v_sample[-1], x_mask)
mistake_rate.name = "single error within note"
cost.name = 'rbm_cost'
model = Model(cost)
cg = ComputationGraph([cost])
step_rule = CompositeRule([
RemoveNotFinite(),
StepClipping(30.0),
Adam(learning_rate=lr),
StepClipping(6.0),
RemoveNotFinite()
]) # Scale(0.01)
gradients = dict(
equizip(cg.parameters,
T.grad(cost, cg.parameters, consider_constant=[v_sample])))
algorithm = GradientDescent(step_rule=step_rule,
gradients=gradients,
cost=cost,
params=cg.parameters)
algorithm.add_updates(cg.updates)
extensions = [
SharedVariableModifier(parameter=cdk,
function=lambda n, v: rnnrbm_cdk[n]
if rnnrbm_cdk.get(n) else v),
SharedVariableModifier(parameter=lr,
function=lambda n, v: float32(0.78 * v)
if n % (200 * 5) == 0 else v),
FinishAfter(after_n_epochs=epochs),
TrainingDataMonitoring(
[
cost,
error_rate,
mistake_rate,
], # hidden_states, debug_val, param_nans,
# aggregation.mean(algorithm.total_gradient_norm)], #+ params,
prefix="train",
after_epoch=False,
every_n_batches=40),
Timing(),
Printing(),
ProgressBar()
]
if test is not None:
extensions.append(
DataStreamMonitoring([cost, error_rate, mistake_rate],
data_stream=test,
updates=cg.updates,
prefix="test",
after_epoch=False,
every_n_batches=40))
if bokeh:
extensions.append(
Plot(
'Training RNN-RBM',
channels=[
[
'train_error on note as a whole',
'train_single error within note',
'test_error on note as a whole',
'test_single error within note'
],
['train_final_cost'],
# ['train_total_gradient_norm'],
]))
main_loop = MainLoop(algorithm=algorithm,
data_stream=train,
model=model,
extensions=extensions)
return main_loop
def train(cli_params):
cli_params['save_dir'] = prepare_dir(cli_params['save_to'])
logfile = os.path.join(cli_params['save_dir'], 'log.txt')
# Log also DEBUG to a file
fh = logging.FileHandler(filename=logfile)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
logger.info('Logging into %s' % logfile)
p, loaded = load_and_log_params(cli_params)
in_dim, data = setup_data(p, test_set=False)
if not loaded:
# Set the zero layer to match input dimensions
p.encoder_layers = (in_dim, ) + p.encoder_layers
ladder = setup_model(p)
# Training
all_params = ComputationGraph([ladder.costs.total]).parameters
logger.info('Found the following parameters: %s' % str(all_params))
# Fetch all batch normalization updates. They are in the clean path.
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
assert 'counter' in [u.name for u in list(bn_updates.keys())], \
'No batch norm params in graph - the graph has been cut?'
training_algorithm = GradientDescent(
cost=ladder.costs.total,
params=all_params,
step_rule=Adam(learning_rate=ladder.lr))
# In addition to actual training, also do BN variable approximations
training_algorithm.add_updates(bn_updates)
short_prints = {
"train": {
'T_C_class': ladder.costs.class_corr,
'T_C_de': list(ladder.costs.denois.values()),
},
"valid_approx":
OrderedDict([
('V_C_class', ladder.costs.class_clean),
('V_E', ladder.error.clean),
('V_C_de', list(ladder.costs.denois.values())),
]),
"valid_final":
OrderedDict([
('VF_C_class', ladder.costs.class_clean),
('VF_E', ladder.error.clean),
('VF_C_de', list(ladder.costs.denois.values())),
]),
}
main_loop = MainLoop(
training_algorithm,
# Datastream used for training
make_datastream(data.train,
data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
n_unlabeled=p.unlabeled_samples),
model=Model(theano.tensor.cast(ladder.costs.total, "float32")),
extensions=[
FinishAfter(after_n_epochs=p.num_epochs),
# This will estimate the validation error using
# running average estimates of the batch normalization
# parameters, mean and variance
ApproxTestMonitoring(
[ladder.costs.class_clean, ladder.error.clean] +
list(ladder.costs.denois.values()),
make_datastream(data.valid,
data.valid_ind,
p.valid_batch_size,
scheme=ShuffledScheme),
prefix="valid_approx"),
TrainingDataMonitoring([
ladder.costs.total, ladder.costs.class_corr,
training_algorithm.total_gradient_norm
] + list(ladder.costs.denois.values()),
prefix="train",
after_epoch=True),
SaveParams(None, all_params, p.save_dir, after_epoch=True),
SaveExpParams(p, p.save_dir, before_training=True),
ShortPrinting(short_prints),
LRDecay(ladder.lr,
p.num_epochs * p.lrate_decay,
p.num_epochs,
after_epoch=True),
])
main_loop.run()
# Get results
df = main_loop.log.to_dataframe()
col = 'valid_final_error_rate_clean'
logger.info('%s %g' % (col, df[col].iloc[-1]))
if main_loop.log.status['epoch_interrupt_received']:
return None
return df
def train_model(cost, unregularized_cost, updates,
train_stream, valid_stream, args, gate_values=None):
step_rule = learning_algorithm(args)
cg = ComputationGraph(cost)
# ADD REGULARIZATION
# WEIGHT NOISE
weight_noise = args.weight_noise
if weight_noise > 0:
weights = VariableFilter(roles=[WEIGHT])(cg.variables)
cg_train = apply_noise(cg, weights, weight_noise)
cost = cg_train.outputs[0]
cost.name = "cost_with_weight_noise"
cg = ComputationGraph(cost)
logger.info(cg.parameters)
# Define algorithm
algorithm = GradientDescent(cost=cost, step_rule=step_rule,
parameters=cg.parameters)
# Add the updates to carry the hidden state
algorithm.add_updates(updates)
# Extensions to be added
extensions = []
# Load from a dumped model
if args.load_path is not None:
extensions.append(Load(args.load_path))
# Generation extension
if args.generate:
extensions.append(TextGenerationExtension(
cost=cost,
generation_length=args.generated_text_lenght,
initial_text_length=args.initial_text_length,
every_n_batches=1,
ploting_path=os.path.join(args.save_path, 'prob_plot.png'),
softmax_sampling=args.softmax_sampling,
dataset=args.dataset,
updates=updates,
interactive_mode=args.interactive_mode))
# Training and Validation score monitoring
extensions.extend([
TrainingDataMonitoring([cost], prefix='train',
every_n_batches=args.monitoring_freq),
DataStreamMonitoring([cost, unregularized_cost],
valid_stream, args.mini_batch_size_valid,
args.dataset,
state_updates=updates,
prefix='valid',
before_first_epoch=(args.visualize == "nothing"),
every_n_batches=args.monitoring_freq)])
# Creating directory for saving model.
if not args.interactive_mode:
if not os.path.exists(args.save_path):
os.makedirs(args.save_path)
elif 'test' in args.save_path:
print "Rewriting in " + args.save_path
else:
raise Exception('Directory already exists')
# Early stopping
extensions.append(EarlyStopping('valid_' + unregularized_cost.name,
args.patience, args.save_path,
every_n_batches=args.monitoring_freq))
# Printing
extensions.append(ProgressBar())
extensions.append(Printing(every_n_batches=args.monitoring_freq))
# Reset the initial states
if args.dataset == "sine":
reset_frequency = 1
else:
reset_frequency = 100
extensions.append(ResetStates([v for v, _ in updates],
every_n_batches=reset_frequency))
# Visualizing extensions
if args.interactive_mode:
extensions.append(InteractiveMode())
main_loop = MainLoop(
model=Model(cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=extensions
)
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 train(ladder, batch_size=100, num_train_examples=60000,
num_epochs=150, lrate_decay=0.67):
# Setting Logger
timestr = time.strftime("%Y_%m_%d_at_%H_%M")
save_path = 'results/mnist_100_standard_' + 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
model = Model(ladder.costs.total)
all_params = model.parameters
print len(all_params)
print all_params
training_algorithm = GradientDescent(
cost=ladder.costs.total, parameters=all_params,
step_rule=Adam(learning_rate=ladder.lr))
# Fetch all batch normalization updates. They are in the clean path.
# In addition to actual training, also do BN variable approximations
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
training_algorithm.add_updates(bn_updates)
monitored_variables = [
ladder.costs.class_corr, ladder.costs.class_clean,
ladder.error, training_algorithm.total_gradient_norm,
ladder.costs.total] + ladder.costs.denois.values()
train_data_stream, test_data_stream = get_mixed_streams(
batch_size)
train_monitoring = TrainingDataMonitoring(
variables=monitored_variables,
prefix="train",
after_epoch=True)
valid_monitoring = DataStreamMonitoring(
variables=monitored_variables,
data_stream=test_data_stream,
prefix="test",
after_epoch=True)
main_loop = MainLoop(
algorithm=training_algorithm,
data_stream=train_data_stream,
model=model,
extensions=[
train_monitoring,
valid_monitoring,
FinishAfter(after_n_epochs=num_epochs),
SaveParams('test_CE_corr', model, save_path),
SaveLog(save_path, after_epoch=True),
LRDecay(lr=ladder.lr,
decay_first=num_epochs * lrate_decay,
decay_last=num_epochs,
after_epoch=True),
Printing()])
main_loop.run()
graphs, extensions, updates = construct_graphs(args, nclasses,
sequence_length)
#graph, extension, update = construct_graphs(args, nclasses, sequence_length)
### optimization algorithm definition
step_rule = CompositeRule([
StepClipping(1.),
#Momentum(learning_rate=args.learning_rate, momentum=0.9),
RMSProp(learning_rate=args.learning_rate, decay_rate=0.5),
])
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 (after epoch to limit the log size)
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")
def train(cli_params):
cli_params['save_dir'] = prepare_dir(cli_params['save_to'])
logfile = os.path.join(cli_params['save_dir'], 'log.txt')
# Log also DEBUG to a file
fh = logging.FileHandler(filename=logfile)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
logger.info('Logging into %s' % logfile)
p, loaded = load_and_log_params(cli_params)
in_dim, data, whiten, cnorm = setup_data(p, test_set=True)
if not loaded:
# Set the zero layer to match input dimensions
p.encoder_layers = (in_dim,) + p.encoder_layers
ladder = setup_model(p)
# Training
all_params = ComputationGraph([ladder.costs.total]).parameters
logger.info('Found the following parameters: %s' % str(all_params))
# Fetch all batch normalization updates. They are in the clean path.
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
assert 'counter' in [u.name for u in bn_updates.keys()], \
'No batch norm params in graph - the graph has been cut?'
training_algorithm = GradientDescent(
cost=ladder.costs.total, params=all_params,
step_rule=Adam(learning_rate=ladder.lr))
# In addition to actual training, also do BN variable approximations
training_algorithm.add_updates(bn_updates)
model=Model(ladder.costs.total)
monitored_variables = [
ladder.costs.class_corr,
ladder.costs.class_clean,
ladder.error,
# training_algorithm.total_gradient_norm,
ladder.costs.total] \
# + ladder.costs.denois.values()
# Make a global history recorder so that we can get summary at end of
# training when we write to Sentinel
# global_history records all relevant monitoring vars
# updated by SaveLog every time
global_history = {}
main_loop = MainLoop(
training_algorithm,
# Datastream used for training
make_datastream(data.train, data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
n_unlabeled=p.unlabeled_samples,
whiten=whiten,
cnorm=cnorm),
model=model,
extensions=[
FinishAfter(after_n_epochs=p.num_epochs),
# This will estimate the validation error using
# running average estimates of the batch normalization
# parameters, mean and variance
ApproxTestMonitoring(
monitored_variables,
make_datastream(data.valid, data.valid_ind,
p.valid_batch_size, whiten=whiten, cnorm=cnorm,
scheme=ShuffledScheme),
prefix="valid_approx"),
# This Monitor is slower, but more accurate since it will first
# estimate batch normalization parameters from training data and
# then do another pass to calculate the validation error.
FinalTestMonitoring(
monitored_variables,
make_datastream(data.train, data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
whiten=whiten, cnorm=cnorm,
scheme=ShuffledScheme),
# DEPREC: we directly test now
# make_datastream(data.valid, data.valid_ind,
# p.valid_batch_size,
# n_labeled=len(data.valid_ind),
# whiten=whiten, cnorm=cnorm,
# scheme=ShuffledScheme),
# prefix="valid_final",
make_datastream(data.test, data.test_ind,
p.batch_size,
n_labeled=len(data.test_ind),
whiten=whiten, cnorm=cnorm,
scheme=ShuffledScheme),
prefix="final_test",
after_n_epochs=p.num_epochs),
TrainingDataMonitoring(
variables=monitored_variables,
prefix="train", after_epoch=True),
# write out to sentinel file for experiment automator to work
# REMOVE THIS if you're running test mode with early stopping immediately after
SentinelWhenFinish(save_dir=p.save_dir,
global_history=global_history),
# originally use 'valid_approx_cost_class_clean'
# turns out should use ER as early stopping
# use CE as a fallback (secondary early stopvar) if ER is the same
# SaveParams(('valid_approx_error_rate', 'valid_approx_cost_class_clean'),
# model, p.save_dir),
# doesn't do early stopping now
SaveParams(None, model, p.save_dir, after_epoch=True),
SaveExpParams(p, p.save_dir, before_training=True),
SaveLog(save_dir=p.save_dir,
after_epoch=True,
global_history=global_history),
Printing(),
# ShortPrinting(short_prints),
LRDecay(ladder.lr, p.num_epochs * p.lrate_decay, p.num_epochs,
after_epoch=True),
])
main_loop.run()
# ================= Add testing at end of training =================
# DEPREC don't do early stopping anymore
if False:
p.load_from = p.save_dir
ladder = setup_model(p)
logger.info('Start testing on trained_params_best')
main_loop = DummyLoop(
extensions=[
# write to global history
SaveLog(save_dir=p.save_dir,
after_training=True,
global_history=global_history),
# write out to sentinel file for experiment automator to work
SentinelWhenFinish(save_dir=p.save_dir,
global_history=global_history),
FinalTestMonitoring(
[ladder.costs.class_clean, ladder.error]
+ ladder.costs.denois.values(),
make_datastream(data.train, data.train_ind,
# These need to match with the training
p.batch_size,
n_labeled=p.labeled_samples,
n_unlabeled=len(data.train_ind),
cnorm=cnorm,
whiten=whiten, scheme=ShuffledScheme),
make_datastream(data.test, data.test_ind,
p.batch_size,
n_labeled=len(data.test_ind),
n_unlabeled=len(data.test_ind),
cnorm=cnorm,
whiten=whiten, scheme=ShuffledScheme),
prefix="test",
before_training=True)
])
main_loop.run()
# Get results
df = main_loop.log.to_dataframe()
# col = 'valid_final_error_rate'
# logger.info('%s %g' % (col, df[col].iloc[-1]))
if main_loop.log.status['epoch_interrupt_received']:
return None
return df
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()
def train_snli_model(new_training_job,
config,
save_path,
params,
fast_start,
fuel_server,
seed,
model='simple'):
if config['exclude_top_k'] > config['num_input_words'] and config[
'num_input_words'] > 0:
raise Exception("Some words have neither word nor def embedding")
c = config
logger = configure_logger(name="snli_baseline_training",
log_file=os.path.join(save_path, "log.txt"))
if not os.path.exists(save_path):
logger.info("Start a new job")
os.mkdir(save_path)
else:
logger.info("Continue an existing job")
with open(os.path.join(save_path, "cmd.txt"), "w") as f:
f.write(" ".join(sys.argv))
# Make data paths nice
for path in [
'dict_path', 'embedding_def_path', 'embedding_path', 'vocab',
'vocab_def', 'vocab_text'
]:
if c.get(path, ''):
if not os.path.isabs(c[path]):
c[path] = os.path.join(fuel.config.data_path[0], c[path])
main_loop_path = os.path.join(save_path, 'main_loop.tar')
main_loop_best_val_path = os.path.join(save_path, 'main_loop_best_val.tar')
stream_path = os.path.join(save_path, 'stream.pkl')
# Save config to save_path
json.dump(config, open(os.path.join(save_path, "config.json"), "w"))
if model == 'simple':
nli_model, data, used_dict, used_retrieval, _ = _initialize_simple_model_and_data(
c)
elif model == 'esim':
nli_model, data, used_dict, used_retrieval, _ = _initialize_esim_model_and_data(
c)
else:
raise NotImplementedError()
# Compute cost
s1, s2 = T.lmatrix('sentence1'), T.lmatrix('sentence2')
if c['dict_path']:
assert os.path.exists(c['dict_path'])
s1_def_map, s2_def_map = T.lmatrix('sentence1_def_map'), T.lmatrix(
'sentence2_def_map')
def_mask = T.fmatrix("def_mask")
defs = T.lmatrix("defs")
else:
s1_def_map, s2_def_map = None, None
def_mask = None
defs = None
s1_mask, s2_mask = T.fmatrix('sentence1_mask'), T.fmatrix('sentence2_mask')
y = T.ivector('label')
cg = {}
for train_phase in [True, False]:
# NOTE: Please don't change outputs of cg
if train_phase:
with batch_normalization(nli_model):
pred = nli_model.apply(s1,
s1_mask,
s2,
s2_mask,
def_mask=def_mask,
defs=defs,
s1_def_map=s1_def_map,
s2_def_map=s2_def_map,
train_phase=train_phase)
else:
pred = nli_model.apply(s1,
s1_mask,
s2,
s2_mask,
def_mask=def_mask,
defs=defs,
s1_def_map=s1_def_map,
s2_def_map=s2_def_map,
train_phase=train_phase)
cost = CategoricalCrossEntropy().apply(y.flatten(), pred)
error_rate = MisclassificationRate().apply(y.flatten(), pred)
cg[train_phase] = ComputationGraph([cost, error_rate])
# Weight decay (TODO: Make it less bug prone)
if model == 'simple':
weights_to_decay = VariableFilter(
bricks=[dense for dense, relu, bn in nli_model._mlp],
roles=[WEIGHT])(cg[True].variables)
weight_decay = np.float32(c['l2']) * sum(
(w**2).sum() for w in weights_to_decay)
elif model == 'esim':
weight_decay = 0.0
else:
raise NotImplementedError()
final_cost = cg[True].outputs[0] + weight_decay
final_cost.name = 'final_cost'
# Add updates for population parameters
if c.get("bn", True):
pop_updates = get_batch_normalization_updates(cg[True])
extra_updates = [(p, m * 0.1 + p * (1 - 0.1)) for p, m in pop_updates]
else:
pop_updates = []
extra_updates = []
if params:
logger.debug("Load parameters from {}".format(params))
with open(params) as src:
loaded_params = load_parameters(src)
cg[True].set_parameter_values(loaded_params)
for param, m in pop_updates:
param.set_value(loaded_params[get_brick(
param).get_hierarchical_name(param)])
if os.path.exists(os.path.join(save_path, "main_loop.tar")):
logger.warning("Manually loading BN stats :(")
with open(os.path.join(save_path, "main_loop.tar")) as src:
loaded_params = load_parameters(src)
for param, m in pop_updates:
param.set_value(
loaded_params[get_brick(param).get_hierarchical_name(param)])
if theano.config.compute_test_value != 'off':
test_value_data = next(
data.get_stream('train', batch_size=4).get_epoch_iterator())
s1.tag.test_value = test_value_data[0]
s1_mask.tag.test_value = test_value_data[1]
s2.tag.test_value = test_value_data[2]
s2_mask.tag.test_value = test_value_data[3]
y.tag.test_value = test_value_data[4]
# Freeze embeddings
if not c['train_emb']:
frozen_params = [
p for E in nli_model.get_embeddings_lookups() for p in E.parameters
]
train_params = [p for p in cg[True].parameters]
assert len(set(frozen_params) & set(train_params)) > 0
else:
frozen_params = []
if not c.get('train_def_emb', 1):
frozen_params_def = [
p for E in nli_model.get_def_embeddings_lookups()
for p in E.parameters
]
train_params = [p for p in cg[True].parameters]
assert len(set(frozen_params_def) & set(train_params)) > 0
frozen_params += frozen_params_def
train_params = [p for p in cg[True].parameters if p not in frozen_params]
train_params_keys = [
get_brick(p).get_hierarchical_name(p) for p in train_params
]
# Optimizer
algorithm = GradientDescent(cost=final_cost,
on_unused_sources='ignore',
parameters=train_params,
step_rule=Adam(learning_rate=c['lr']))
algorithm.add_updates(extra_updates)
m = Model(final_cost)
parameters = m.get_parameter_dict() # Blocks version mismatch
logger.info("Trainable parameters" + "\n" +
pprint.pformat([(key, parameters[key].get_value().shape)
for key in sorted(train_params_keys)],
width=120))
logger.info("# of parameters {}".format(
sum([
np.prod(parameters[key].get_value().shape)
for key in sorted(train_params_keys)
])))
### Monitored args ###
train_monitored_vars = [final_cost] + cg[True].outputs
monitored_vars = cg[False].outputs
val_acc = monitored_vars[1]
to_monitor_names = [
'def_unk_ratio', 's1_merged_input_rootmean2', 's1_def_mean_rootmean2',
's1_gate_rootmean2', 's1_compose_gate_rootmean2'
]
for k in to_monitor_names:
train_v, valid_v = VariableFilter(name=k)(
cg[True]), VariableFilter(name=k)(cg[False])
if len(train_v):
logger.info("Adding {} tracking".format(k))
train_monitored_vars.append(train_v[0])
monitored_vars.append(valid_v[0])
else:
logger.warning("Didnt find {} in cg".format(k))
if c['monitor_parameters']:
for name in train_params_keys:
param = parameters[name]
num_elements = numpy.product(param.get_value().shape)
norm = param.norm(2) / num_elements
grad_norm = algorithm.gradients[param].norm(2) / num_elements
step_norm = algorithm.steps[param].norm(2) / num_elements
stats = tensor.stack(norm, grad_norm, step_norm,
step_norm / grad_norm)
stats.name = name + '_stats'
train_monitored_vars.append(stats)
regular_training_stream = data.get_stream('train',
batch_size=c['batch_size'],
seed=seed)
if fuel_server:
# the port will be configured by the StartFuelServer extension
training_stream = ServerDataStream(
sources=regular_training_stream.sources,
hwm=100,
produces_examples=regular_training_stream.produces_examples)
else:
training_stream = regular_training_stream
### Build extensions ###
extensions = [
# Load(main_loop_path, load_iteration_state=True, load_log=True)
# .set_conditions(before_training=not new_training_job),
StartFuelServer(regular_training_stream,
stream_path,
hwm=100,
script_path=os.path.join(
os.path.dirname(__file__),
"../bin/start_fuel_server.py"),
before_training=fuel_server),
Timing(every_n_batches=c['mon_freq']),
ProgressBar(),
RetrievalPrintStats(retrieval=used_retrieval,
every_n_batches=c['mon_freq_valid'],
before_training=not fast_start),
Timestamp(),
TrainingDataMonitoring(train_monitored_vars,
prefix="train",
every_n_batches=c['mon_freq']),
]
if c['layout'] == 'snli':
validation = DataStreamMonitoring(monitored_vars,
data.get_stream('valid',
batch_size=14,
seed=seed),
before_training=not fast_start,
on_resumption=True,
after_training=True,
every_n_batches=c['mon_freq_valid'],
prefix='valid')
extensions.append(validation)
elif c['layout'] == 'mnli':
validation = DataStreamMonitoring(monitored_vars,
data.get_stream('valid_matched',
batch_size=14,
seed=seed),
every_n_batches=c['mon_freq_valid'],
on_resumption=True,
after_training=True,
prefix='valid_matched')
validation_mismatched = DataStreamMonitoring(
monitored_vars,
data.get_stream('valid_mismatched', batch_size=14, seed=seed),
every_n_batches=c['mon_freq_valid'],
before_training=not fast_start,
on_resumption=True,
after_training=True,
prefix='valid_mismatched')
extensions.extend([validation, validation_mismatched])
else:
raise NotImplementedError()
# Similarity trackers for embeddings
if len(c.get('vocab_def', '')):
retrieval_vocab = Vocabulary(c['vocab_def'])
else:
retrieval_vocab = data.vocab
retrieval_all = Retrieval(vocab_text=retrieval_vocab,
dictionary=used_dict,
max_def_length=c['max_def_length'],
exclude_top_k=0,
max_def_per_word=c['max_def_per_word'])
for name in [
's1_word_embeddings', 's1_dict_word_embeddings',
's1_translated_word_embeddings'
]:
variables = VariableFilter(name=name)(cg[False])
if len(variables):
s1_emb = variables[0]
logger.info("Adding similarity tracking for " + name)
# A bit sloppy about downcast
if "dict" in name:
embedder = construct_dict_embedder(theano.function(
[s1, defs, def_mask, s1_def_map],
s1_emb,
allow_input_downcast=True),
vocab=data.vocab,
retrieval=retrieval_all)
extensions.append(
SimilarityWordEmbeddingEval(
embedder=embedder,
prefix=name,
every_n_batches=c['mon_freq_valid'],
before_training=not fast_start))
else:
embedder = construct_embedder(theano.function(
[s1], s1_emb, allow_input_downcast=True),
vocab=data.vocab)
extensions.append(
SimilarityWordEmbeddingEval(
embedder=embedder,
prefix=name,
every_n_batches=c['mon_freq_valid'],
before_training=not fast_start))
track_the_best = TrackTheBest(validation.record_name(val_acc),
before_training=not fast_start,
every_n_epochs=c['save_freq_epochs'],
after_training=not fast_start,
every_n_batches=c['mon_freq_valid'],
choose_best=min)
extensions.append(track_the_best)
# Special care for serializing embeddings
if len(c.get('embedding_path', '')) or len(c.get('embedding_def_path',
'')):
extensions.insert(
0,
LoadNoUnpickling(main_loop_path,
load_iteration_state=True,
load_log=True).set_conditions(
before_training=not new_training_job))
extensions.append(
Checkpoint(main_loop_path,
parameters=train_params + [p for p, m in pop_updates],
save_main_loop=False,
save_separately=['log', 'iteration_state'],
before_training=not fast_start,
every_n_epochs=c['save_freq_epochs'],
after_training=not fast_start).add_condition(
['after_batch', 'after_epoch'],
OnLogRecord(track_the_best.notification_name),
(main_loop_best_val_path, )))
else:
extensions.insert(
0,
Load(main_loop_path, load_iteration_state=True,
load_log=True).set_conditions(
before_training=not new_training_job))
extensions.append(
Checkpoint(main_loop_path,
parameters=cg[True].parameters +
[p for p, m in pop_updates],
before_training=not fast_start,
every_n_epochs=c['save_freq_epochs'],
after_training=not fast_start).add_condition(
['after_batch', 'after_epoch'],
OnLogRecord(track_the_best.notification_name),
(main_loop_best_val_path, )))
extensions.extend([
DumpCSVSummaries(save_path,
every_n_batches=c['mon_freq_valid'],
after_training=True),
DumpTensorflowSummaries(save_path,
after_epoch=True,
every_n_batches=c['mon_freq_valid'],
after_training=True),
Printing(every_n_batches=c['mon_freq_valid']),
PrintMessage(msg="save_path={}".format(save_path),
every_n_batches=c['mon_freq']),
FinishAfter(after_n_batches=c['n_batches']).add_condition(
['after_batch'],
OnLogStatusExceed('iterations_done', c['n_batches']))
])
logger.info(extensions)
### Run training ###
if "VISDOM_SERVER" in os.environ:
print("Running visdom server")
ret = subprocess.Popen([
os.path.join(os.path.dirname(__file__), "../visdom_plotter.py"),
"--visdom-server={}".format(os.environ['VISDOM_SERVER']),
"--folder={}".format(save_path)
])
time.sleep(0.1)
if ret.returncode is not None:
raise Exception()
atexit.register(lambda: os.kill(ret.pid, signal.SIGINT))
model = Model(cost)
for p, m in pop_updates:
model._parameter_dict[get_brick(p).get_hierarchical_name(p)] = p
main_loop = MainLoop(algorithm,
training_stream,
model=model,
extensions=extensions)
assert os.path.exists(save_path)
main_loop.run()
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()
mean_sigma, min_sigma, min_mu, max_mu, mean_mu, max_sigma, mean_coeff,
min_coeff, max_coeff
]
#################
# Algorithm
#################
n_batches = 200
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule(
[StepClipping(10.0),
Adam(lr)]))
algorithm.add_updates(extra_updates)
train_monitor = TrainingDataMonitoring(variables=monitoring_variables +
data_monitoring,
every_n_batches=n_batches,
prefix="train")
valid_monitor = DataStreamMonitoring(monitoring_variables,
valid_stream,
every_n_batches=n_batches,
prefix="valid")
extensions = [
ProgressBar(),
Timing(every_n_batches=n_batches), train_monitor, valid_monitor,
TrackTheBest('valid_nll', every_n_batches=n_batches),
def run(discriminative_regularization=True):
streams = create_celeba_streams(training_batch_size=100,
monitoring_batch_size=500,
include_targets=False)
main_loop_stream, train_monitor_stream, valid_monitor_stream = streams[:3]
# Compute parameter updates for the batch normalization population
# statistics. They are updated following an exponential moving average.
rval = create_training_computation_graphs(discriminative_regularization)
cg, bn_cg, variance_parameters = rval
pop_updates = list(
set(get_batch_normalization_updates(bn_cg, allow_duplicates=True)))
decay_rate = 0.05
extra_updates = [(p, m * decay_rate + p * (1 - decay_rate))
for p, m in pop_updates]
model = Model(bn_cg.outputs[0])
selector = Selector(
find_bricks(
model.top_bricks, lambda brick: brick.name in
('encoder_convnet', 'encoder_mlp', 'decoder_convnet', 'decoder_mlp'
)))
parameters = list(selector.get_parameters().values()) + variance_parameters
# Prepare algorithm
step_rule = Adam()
algorithm = GradientDescent(cost=bn_cg.outputs[0],
parameters=parameters,
step_rule=step_rule)
algorithm.add_updates(extra_updates)
# Prepare monitoring
monitored_quantities_list = []
for graph in [bn_cg, cg]:
cost, kl_term, reconstruction_term = graph.outputs
cost.name = 'nll_upper_bound'
avg_kl_term = kl_term.mean(axis=0)
avg_kl_term.name = 'avg_kl_term'
avg_reconstruction_term = -reconstruction_term.mean(axis=0)
avg_reconstruction_term.name = 'avg_reconstruction_term'
monitored_quantities_list.append(
[cost, avg_kl_term, avg_reconstruction_term])
train_monitoring = DataStreamMonitoring(monitored_quantities_list[0],
train_monitor_stream,
prefix="train",
updates=extra_updates,
after_epoch=False,
before_first_epoch=False,
every_n_epochs=5)
valid_monitoring = DataStreamMonitoring(monitored_quantities_list[1],
valid_monitor_stream,
prefix="valid",
after_epoch=False,
before_first_epoch=False,
every_n_epochs=5)
# Prepare checkpoint
save_path = 'celeba_vae_{}regularization.zip'.format(
'' if discriminative_regularization else 'no_')
checkpoint = Checkpoint(save_path, every_n_epochs=5, use_cpickle=True)
extensions = [
Timing(),
FinishAfter(after_n_epochs=75), train_monitoring, valid_monitoring,
checkpoint,
Printing(),
ProgressBar()
]
main_loop = MainLoop(data_stream=main_loop_stream,
algorithm=algorithm,
extensions=extensions)
main_loop.run()
min_binary = binary.min().copy(name="binary_min")
mean_binary = binary.mean().copy(name="binary_mean")
max_binary = binary.max().copy(name="binary_max")
data_monitoring += [mean_sigma, min_sigma, min_mu, max_mu, mean_mu, max_sigma, mean_binary, min_binary, max_binary]
#################
# Algorithm
#################
n_batches = 200
algorithm = GradientDescent(
cost=cost, parameters=cg.parameters, step_rule=CompositeRule([StepClipping(10.0), Adam(lr)])
)
algorithm.add_updates(extra_updates)
train_monitor = TrainingDataMonitoring(
variables=monitoring_variables + data_monitoring, every_n_batches=n_batches, prefix="train"
)
valid_monitor = DataStreamMonitoring(
monitoring_variables + data_monitoring, valid_stream, every_n_batches=n_batches, prefix="valid"
)
extensions = [
ProgressBar(),
Timing(every_n_batches=n_batches),
train_monitor,
valid_monitor,
TrackTheBest("valid_nll", every_n_batches=n_batches),
# # cost += reg * abs(p).sum()
# cost += reg * (p ** 2).sum()
# param_nans += T.isnan(p).sum()
# name = params[i].name
# params[i] = params[i].mean()
# params[i].name = name + str(i)
cost.name = "final_cost"
# param_nans.name = 'params_nans'
## Training algorithm
# algorithm.initialize()
algorithm.add_updates(cg.updates)
extensions = [FinishAfter(after_n_epochs=5000),
DataStreamMonitoring(
[cost, error_rate, mistake_rate],
data_stream=test,
updates=cg.updates,
prefix="test"),
TrainingDataMonitoring(
[cost, # hidden_states, debug_val, param_nans,
aggregation.mean(algorithm.total_gradient_norm)], #+ params,
prefix="train",
after_epoch=True),
Timing(),
Printing(),
ProgressBar()]
main_loop.run()
sys.exit(0)
graphs, extensions, updates = construct_graphs(args, nclasses, sequence_length)
### optimization algorithm definition
step_rule = CompositeRule([
StepClipping(1.),
#Momentum(learning_rate=args.learning_rate, momentum=0.9),
RMSProp(learning_rate=args.learning_rate, decay_rate=0.5),
])
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 (after epoch to limit the log size)
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=False))
def train(cli_params):
cli_params["save_dir"] = prepare_dir(cli_params["save_to"])
logfile = os.path.join(cli_params["save_dir"], "log.txt")
# Log also DEBUG to a file
fh = logging.FileHandler(filename=logfile)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
logger.info("Logging into %s" % logfile)
p, loaded = load_and_log_params(cli_params)
in_dim, data, whiten, cnorm = setup_data(p, test_set=False)
if not loaded:
# Set the zero layer to match input dimensions
p.encoder_layers = (in_dim,) + p.encoder_layers
ladder = setup_model(p)
# Training
all_params = ComputationGraph([ladder.costs.total]).parameters
logger.info("Found the following parameters: %s" % str(all_params))
# Fetch all batch normalization updates. They are in the clean path.
bn_updates = ComputationGraph([ladder.costs.class_clean]).updates
assert "counter" in [u.name for u in bn_updates.keys()], "No batch norm params in graph - the graph has been cut?"
training_algorithm = GradientDescent(
cost=ladder.costs.total, params=all_params, step_rule=Adam(learning_rate=ladder.lr)
)
# In addition to actual training, also do BN variable approximations
training_algorithm.add_updates(bn_updates)
model = Model(ladder.costs.total)
monitored_variables = [
ladder.costs.class_corr,
ladder.costs.class_clean,
ladder.error,
# training_algorithm.total_gradient_norm,
ladder.costs.total,
]
# + ladder.costs.denois.values()
# Make a global history recorder so that we can get summary at end of
# training when we write to Sentinel
# global_history records all relevant monitoring vars
# updated by SaveLog every time
global_history = {}
main_loop = MainLoop(
training_algorithm,
# Datastream used for training
make_datastream(
data.train,
data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
n_unlabeled=p.unlabeled_samples,
whiten=whiten,
cnorm=cnorm,
),
model=model,
extensions=[
FinishAfter(after_n_epochs=p.num_epochs),
# write out to sentinel file for experiment automator to work
SentinelWhenFinish(save_dir=p.save_dir, global_history=global_history),
# This will estimate the validation error using
# running average estimates of the batch normalization
# parameters, mean and variance
ApproxTestMonitoring(
monitored_variables,
make_datastream(
data.valid, data.valid_ind, p.valid_batch_size, whiten=whiten, cnorm=cnorm, scheme=ShuffledScheme
),
prefix="valid_approx",
),
# This Monitor is slower, but more accurate since it will first
# estimate batch normalization parameters from training data and
# then do another pass to calculate the validation error.
FinalTestMonitoring(
monitored_variables,
make_datastream(
data.train,
data.train_ind,
p.batch_size,
n_labeled=p.labeled_samples,
whiten=whiten,
cnorm=cnorm,
scheme=ShuffledScheme,
),
make_datastream(
data.valid,
data.valid_ind,
p.valid_batch_size,
n_labeled=len(data.valid_ind),
whiten=whiten,
cnorm=cnorm,
scheme=ShuffledScheme,
),
prefix="valid_final",
after_n_epochs=p.num_epochs,
),
TrainingDataMonitoring(variables=monitored_variables, prefix="train", after_epoch=True),
SaveParams("valid_approx_cost_class_corr", model, p.save_dir),
# SaveParams(None, all_params, p.save_dir, after_epoch=True),
SaveExpParams(p, p.save_dir, before_training=True),
SaveLog(save_dir=p.save_dir, after_epoch=True, global_history=global_history),
Printing(),
# ShortPrinting(short_prints),
LRDecay(ladder.lr, p.num_epochs * p.lrate_decay, p.num_epochs, after_epoch=True),
],
)
main_loop.run()
# Get results
df = main_loop.log.to_dataframe()
col = "valid_final_error_rate"
logger.info("%s %g" % (col, df[col].iloc[-1]))
if main_loop.log.status["epoch_interrupt_received"]:
return None
return df