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_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 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 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