def test_load_train_file():
"""
Loads a YAML file with and without environment variables.
"""
environ = {'PYLEARN2_DATA_PATH': '/just/a/test/path/'}
load_train_file(yaml_path + 'test_model.yaml')
load_train_file(yaml_path + 'test_model.yaml', environ=environ)
def test_load_train_file():
"""
Loads a YAML file with and without environment variables.
"""
environ = {
'PYLEARN2_DATA_PATH': '/just/a/test/path/'
}
load_train_file(yaml_path + 'test_model.yaml')
load_train_file(yaml_path + 'test_model.yaml', environ=environ)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--out",required=True,help="The name of the yaml file produced")
parser.add_argument("--template",required=True,help="YAML template")
parser.add_argument("--hparams",required=True,help="Hyper-parameters configuration")
parser.add_argument("--force",action='store_true',help="Force to overwrite the old yaml file produced")
parser.add_argument("--range",nargs=2,type=int,help="Subrange of files to execute")
options = parser.parse_args()
out = options.out
template = options.template
hparams = options.hparams
force = options.force
print options
# Generates a list of hyper-parameter names and a list of
# hyper-parameter values
hpnames, hpvalues = generate_params(hparamfile=hparams,
generate="uniform",
search_mode="fix-grid-search")
# Writes template with each hyper-parameter settings in
# succesive files and returns the name of the files
files = write_files(template="".join(open(template,"r")),hpnames=hpnames,
hpvalues=hpvalues,save_path=out,force=force)
if options.range:
if options.range[1]==-1:
options.range[1] = len(files)
assert options.range[0] < options.range[1]
iterator = xrange(*options.range)
else:
iterator = xrange(0,len(files))
print list(iterator)
print_error_message("errors\n",out,"w")
from pylearn2.utils import serial
for i in iterator:#xrange(0,len(files)):
f = files[i]
print f
try:
serial.load_train_file(f).main_loop()
except BaseException as e:
print traceback.format_exc()
print e
print_error_message("%s : %s\n" % (f,str(e)),out)
def test_mnist():
"""
Test the mnist.yaml file from the maxout
paper on random input
"""
skip_if_no_gpu()
train = load_train_file(
os.path.join(pylearn2.__path__[0], "scripts/papers/maxout/mnist.yaml"))
# Load fake MNIST data
init_value = control.load_data
control.load_data = [False]
train.dataset = MNIST(which_set='train',
one_hot=1,
axes=['c', 0, 1, 'b'],
start=0,
stop=100)
train.algorithm._set_monitoring_dataset(train.dataset)
control.load_data = init_value
# Train shortly and prevent saving
train.algorithm.termination_criterion = EpochCounter(max_epochs=1)
train.extensions.pop(0)
train.save_freq = 0
train.main_loop()
def train_again(yaml):
'''
Relaunch training of a model with conditions specified by the YAML
Looks for the model file defined by save path and replace the model
instanciated by the one that was trained before
-------------------------------------------------------------------
yaml : string, filename
YAML file defining the exp to be continued
'''
context = serial.load_train_file(yaml)
print "\tLoaded YAML"
# Load the trained model
model_file = context.save_path
if not os.path.isfile(model_file):
model_file = context.extensions[-1].save_path
with open(model_file, 'r') as m_f:
trained_model = pkl.load(m_f)
# Define the continuing one
new_model = push_monitor(trained_model,
'trained_model',
transfer_experience=True)
# Define it as the model to be trained
context.model = new_model
# Train again
context.main_loop()
def limited_epoch_train(file_path, max_epochs = 1):
try:
train = load_train_file(file_path)
train.algorithm.termination_criterion = EpochCounter(max_epochs = max_epochs)
train.main_loop()
except NoDataPathError:
raise SkipTest("PYLEARN2_DATA_PATH environment variable not defined")
def train_again(yaml):
'''
Relaunch training of a model with conditions specified by the YAML
Looks for the model file defined by save path and replace the model
instanciated by the one that was trained before
-------------------------------------------------------------------
yaml : string, filename
YAML file defining the exp to be continued
'''
context = serial.load_train_file(yaml)
print "\tLoaded YAML"
# Load the trained model
model_file = context.save_path
if not os.path.isfile(model_file):
model_file = context.extensions[-1].save_path
with open(model_file, 'r') as m_f:
trained_model = pkl.load(m_f)
# Define the continuing one
new_model = push_monitor(trained_model, 'trained_model',
transfer_experience=True)
# Define it as the model to be trained
context.model = new_model
# Train again
context.main_loop()
def main(argv):
try:
opts, args = getopt.getopt(argv, '')
student_yaml = args[0]
load_layer = int(args[1])
except getopt.GetoptError:
usage()
sys.exit(2)
# Load student
student = serial.load_train_file(student_yaml)
with open(student_yaml, "r") as sty:
student = yaml_parse.load(sty)
# Load hints
if student.algorithm.cost.hints is not None:
student_layers = list(zip(*student.algorithm.cost.hints)[0])
teacher_layers = list(zip(*student.algorithm.cost.hints)[1])
n_hints = len(student_layers)
else:
n_hints = 0
hint_path = student.save_path[0:-4] + "_hintlayer" + str(
load_layer) + ".pkl"
for ext in range(len(student.extensions)):
if isinstance(student.extensions[ext], MonitorBasedSaveBest):
hint_path = student.extensions[ext].save_path[
0:-9] + "_hintlayer" + str(load_layer) + "_best.pkl"
# Load pretrained student network
pretrained_model = serial.load(hint_path)
#fo = open(hint_path, 'r')
#pretrained_model = pkl.load(fo)
#fo.close()
student.model.layers[0:load_layer +
1] = pretrained_model.layers[0:load_layer + 1]
pretrained_model = None
del pretrained_model
#student.algorithm.learning_rate.set_value(0.001)
#for i in range(0,load_layer+1):
# student.model.layers[i].W_lr_scale = 0.1
# student.model.layers[i].b_lr_scale = 0.1
student.save_path = student.save_path[0:-4] + "_hint" + str(
load_layer) + "_softmax.pkl"
for ext in range(len(student.extensions)):
if isinstance(student.extensions[ext], MonitorBasedSaveBest):
student.extensions[
ext].save_path = student.save_path[0:-4] + "_best.pkl"
student.main_loop()
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--out",
required=True,
help="The name of the yaml file produced")
parser.add_argument("--template", required=True, help="YAML template")
parser.add_argument("--hparams",
required=True,
help="Hyper-parameters configuration")
parser.add_argument("--force",
action='store_true',
help="Force to overwrite the old yaml file produced")
options = parser.parse_args()
out = options.out
template = options.template
hparams = options.hparams
force = options.force
print options
# Generates a list of hyper-parameter names and a list of
# hyper-parameter values
hpnames, hpvalues = generate_params(hparamfile=hparams,
generate="uniform",
search_mode="fix-grid-search")
# Writes template with each hyper-parameter settings in
# succesive files and returns the name of the files
files = write_files(template="".join(open(template, "r")),
hpnames=hpnames,
hpvalues=hpvalues,
save_path=out,
force=force)
print_error_message("errors\n", out, "w")
from pylearn2.utils import serial
for i in xrange(0, len(files)):
f = files[i]
try:
serial.load_train_file(f).main_loop()
except BaseException as e:
print traceback.format_exc()
print e
print_error_message("%s : %s\n" % (f, str(e)), out)
def main(argv):
try:
opts, args = getopt.getopt(argv, '')
student_yaml = args[0]
load_layer = int(args[1])
except getopt.GetoptError:
usage()
sys.exit(2)
# Load student
student = serial.load_train_file(student_yaml)
with open(student_yaml, "r") as sty:
student = yaml_parse.load(sty)
# Load hints
if student.algorithm.cost.hints is not None:
student_layers = list(zip(*student.algorithm.cost.hints)[0])
teacher_layers = list(zip(*student.algorithm.cost.hints)[1])
n_hints = len(student_layers)
else:
n_hints = 0
hint_path = student.save_path[0:-4] + "_hintlayer" + str(load_layer) + ".pkl"
for ext in range(len(student.extensions)):
if isinstance(student.extensions[ext],MonitorBasedSaveBest):
hint_path = student.extensions[ext].save_path[0:-9] + "_hintlayer" + str(load_layer) + "_best.pkl"
# Load pretrained student network
pretrained_model = serial.load(hint_path)
#fo = open(hint_path, 'r')
#pretrained_model = pkl.load(fo)
#fo.close()
student.model.layers[0:load_layer+1] = pretrained_model.layers[0:load_layer+1]
pretrained_model = None
del pretrained_model
#student.algorithm.learning_rate.set_value(0.001)
#for i in range(0,load_layer+1):
# student.model.layers[i].W_lr_scale = 0.1
# student.model.layers[i].b_lr_scale = 0.1
student.save_path = student.save_path[0:-4] + "_hint" + str(load_layer) + "_softmax.pkl"
for ext in range(len(student.extensions)):
if isinstance(student.extensions[ext],MonitorBasedSaveBest):
student.extensions[ext].save_path = student.save_path[0:-4] + "_best.pkl"
student.main_loop()
def instantiate_decoder_from_yaml(yaml_filepath):
decoder = serial.load_train_file(yaml_filepath)
assert isinstance(decoder, Pylearn2KaldiDecoderProvider)
assert isinstance(decoder.model, Model)
assert isinstance(decoder.preprocessor, OnlinePreprocessor) or decoder.preprocessor is None
return decoder
def main(dataset='data/pylearn2/mnist.pkl.gz', nn_config = "lib/pylearn2_log.yaml"):
print '... load and setup data'
load_data(dataset)
print '... building the model structure'
train_obj = serial.load_train_file(nn_config)
print '... training the model'
train_obj.main_loop()
def main(dataset='data/pylearn2/mnist.pkl.gz',
nn_config="lib/pylearn2_log.yaml"):
print '... load and setup data'
load_data(dataset)
print '... building the model structure'
train_obj = serial.load_train_file(nn_config)
print '... training the model'
train_obj.main_loop()
def main(argv, freeze):
try:
opts, args = getopt.getopt(argv, '')
yaml = args[0]
model = args[1]
except getopt.GetoptError:
usage()
sys.exit(2)
# Load yaml
with open(yaml, "r") as sty:
train = serial.load_train_file(yaml)
#train = yaml_parse.load(sty)
# Load pretrained model with bad sigmoid output
with open(model, 'r') as fo:
model = pkl.load(fo)
# Remove the last layer, puts a real sigmoid instead
if freeze:
for i in range(0, len(model.layers) - 2):
model.freeze(model.layers[i].get_params())
### Add last conv elemwise
layer = ConvElemwise(layer_name= 'out',
output_channels= 1,
kernel_shape=(1,1),
irange=0.05,
nonlinearity=IdentityConvNonlinearity(),
max_kernel_norm= 7.9,
tied_b=1)
layer.set_mlp(model)
layer.set_input_space(model.layers[-3].get_output_space())
model.layers[-2] = layer
### Add Sigmoid
layer = SigmoidExtended(layer_name='y', n_classes=1)
layer.set_mlp(model)
layer.set_input_space(model.layers[-2].get_output_space())
model.layers[-1] = layer
#print model.layers
#model.monitor = train.model.monitor
#train.model = model
train.model = push_monitor(model, "old")
print train.model
#train = Train(train.dataset, model, train.algorithm, train.save_path,
# train.save_freq, train.extensions, train.allow_overwrite)
train.main_loop()
def _load(self, fname):
"""
Internal load.
"""
# load the train object
train_obj = serial.load_train_file(fname)
if not isinstance(train_obj, Train):
raise ValueError('Top level object must be a pylearn2.Train')
# prepare for training
pre_train(train_obj)
self.train_obj = train_obj
self.crt_num_steps = 0
def main(argv, freeze):
try:
opts, args = getopt.getopt(argv, '')
yaml = args[0]
model = args[1]
except getopt.GetoptError:
usage()
sys.exit(2)
# Load yaml
with open(yaml, "r") as sty:
train = serial.load_train_file(yaml)
#train = yaml_parse.load(sty)
# Load pretrained model with bad sigmoid output
with open(model, 'r') as fo:
model = pkl.load(fo)
# Remove the last layer, puts a real sigmoid instead
if freeze:
for i in range(0, len(model.layers) - 2):
model.freeze(model.layers[i].get_params())
### Add last conv elemwise
layer = ConvElemwise(layer_name='out',
output_channels=1,
kernel_shape=(1, 1),
irange=0.05,
nonlinearity=IdentityConvNonlinearity(),
max_kernel_norm=7.9,
tied_b=1)
layer.set_mlp(model)
layer.set_input_space(model.layers[-3].get_output_space())
model.layers[-2] = layer
### Add Sigmoid
layer = SigmoidExtended(layer_name='y', n_classes=1)
layer.set_mlp(model)
layer.set_input_space(model.layers[-2].get_output_space())
model.layers[-1] = layer
#print model.layers
#model.monitor = train.model.monitor
#train.model = model
train.model = push_monitor(model, "old")
print train.model
#train = Train(train.dataset, model, train.algorithm, train.save_path,
# train.save_freq, train.extensions, train.allow_overwrite)
train.main_loop()
def limited_epoch_train(file_path, max_epochs=1):
"""
This method trains a given YAML file for a single epoch
Parameters
----------
file_path : str
The path to the YAML file to be trained
max_epochs : int
The number of epochs to train this YAML file for.
Defaults to 1.
"""
train = load_train_file(file_path)
train.algorithm.termination_criterion = EpochCounter(max_epochs=max_epochs)
train.main_loop()
def main(argv):
try:
opts, args = getopt.getopt(argv, '')
model_yaml = args[0]
except getopt.GetoptError:
usage()
sys.exit(2)
# Load student
#with open(model_yaml, "r") as sty:
#model = yaml_parse.load(sty)
model = serial.load_train_file(model_yaml)
result = numberMult(model)
print 'Number of multiplications is %is' % (result)
def main(argv):
try:
opts, args = getopt.getopt(argv, '')
model_yaml = args[0]
except getopt.GetoptError:
usage()
sys.exit(2)
# Load student
#with open(model_yaml, "r") as sty:
#model = yaml_parse.load(sty)
model = serial.load_train_file(model_yaml)
result = numberParams(model)
print 'Number of parameters is %i' % (result)
def main(argv):
try:
opts, args = getopt.getopt(argv, '')
model_yaml = args[0]
except getopt.GetoptError:
usage()
sys.exit(2)
# Load student
model = serial.load_train_file(model_yaml)
#with open(model_yaml, "r") as sty:
# model = yaml_parse.load(sty)
result = numberMult(model)
print 'Number of multiplications is %is' % (result)
def test_train_example():
""" tests that the train example script runs correctly """
path = pylearn2.__path__[0]
train_example_path = path + '/scripts/train_example'
cwd = os.getcwd()
try:
os.chdir(train_example_path)
train_yaml_path = 'cifar_grbm_smd.yaml'
train_object = load_train_file(train_yaml_path)
#make the termination criterion really lax so the test won't run for long
train_object.algorithm.termination_criterion.prop_decrease = 0.5
train_object.algorithm.termination_criterion.N = 1
train_object.main_loop()
finally:
os.chdir(cwd)
def main(argv):
try:
opts, args = getopt.getopt(argv, '')
model_yaml = args[0]
except getopt.GetoptError:
usage()
sys.exit(2)
# Load student
model = serial.load_train_file(model_yaml)
#with open(model_yaml, "r") as sty:
# model = yaml_parse.load(sty)
result = numberParams(model)
print 'Number of parameters is %i' % (result)
def test_train_example():
""" tests that the grbm_smd example script runs correctly """
assert config.mode != "DEBUG_MODE"
path = pylearn2.__path__[0]
train_example_path = os.path.join(path, 'scripts', 'tutorials', 'grbm_smd')
cwd = os.getcwd()
try:
os.chdir(train_example_path)
train_yaml_path = os.path.join(train_example_path, 'cifar_grbm_smd.yaml')
train_object = load_train_file(train_yaml_path)
#make the termination criterion really lax so the test won't run for long
train_object.algorithm.termination_criterion.prop_decrease = 0.5
train_object.algorithm.termination_criterion.N = 1
train_object.main_loop()
finally:
os.chdir(cwd)
def train_again(yaml):
'''
Relaunch training of a model with conditions specified by the YAML
Looks for the model file defined by save path and replace the model
instanciated by the one that was trained before
-------------------------------------------------------------------
yaml : string, filename
YAML file defining the exp to be continued
'''
context = serial.load_train_file(yaml)
print "\tLoaded YAML"
# Load the trained model
model_file = context.save_path
for ext in range(len(context.extensions)):
if isinstance(context.extensions[ext], MonitorBasedSaveBest):
pos = ext
else:
raise AssertionError(
'No MonitorBasedSaveBest extension in the model!')
if not os.path.isfile(model_file):
model_file = context.extensions[pos].save_path
with open(model_file, 'r') as m_f:
trained_model = pkl.load(m_f)
# Define the continuing one
new_model = push_monitor(trained_model,
'trained_model',
transfer_experience=True)
# Define it as the model to be trained
context.model = new_model
context.save_path = context.extensions[pos].save_path[
0:-4] + "_continue.pkl"
context.extensions[pos].save_path = context.save_path[0:-4] + "_best.pkl"
# Train again
context.main_loop()
def test_train_example():
""" tests that the grbm_smd example script runs correctly """
assert config.mode != "DEBUG_MODE"
if 'TRAVIS' in os.environ and os.environ['TRAVIS'] == '1':
raise SkipTest()
path = pylearn2.__path__[0]
train_example_path = path + '/scripts/tutorials/grbm_smd'
cwd = os.getcwd()
try:
os.chdir(train_example_path)
train_yaml_path = 'cifar_grbm_smd.yaml'
train_object = load_train_file(train_yaml_path)
#make the termination criterion really lax so the test won't run for long
train_object.algorithm.termination_criterion.prop_decrease = 0.5
train_object.algorithm.termination_criterion.N = 1
train_object.main_loop()
finally:
os.chdir(cwd)
def test_train_example():
""" tests that the train example script runs correctly """
assert config.mode != "DEBUG_MODE"
if 'TRAVIS' in os.environ and os.environ['TRAVIS'] == '1':
raise SkipTest()
path = pylearn2.__path__[0]
train_example_path = path + '/scripts/train_example'
cwd = os.getcwd()
try:
os.chdir(train_example_path)
train_yaml_path = 'cifar_grbm_smd.yaml'
train_object = load_train_file(train_yaml_path)
#make the termination criterion really lax so the test won't run for long
train_object.algorithm.termination_criterion.prop_decrease = 0.5
train_object.algorithm.termination_criterion.N = 1
train_object.main_loop()
finally:
os.chdir(cwd)
def test_train_example():
""" tests that the grbm_smd example script runs correctly """
assert config.mode != "DEBUG_MODE"
path = pylearn2.__path__[0]
train_example_path = os.path.join(path, 'scripts', 'tutorials', 'grbm_smd')
cwd = os.getcwd()
try:
os.chdir(train_example_path)
train_yaml_path = os.path.join(train_example_path,
'cifar_grbm_smd.yaml')
train_object = load_train_file(train_yaml_path)
#make the termination criterion really lax so the test won't run for long
train_object.algorithm.termination_criterion.prop_decrease = 0.5
train_object.algorithm.termination_criterion.N = 1
train_object.main_loop()
finally:
os.chdir(cwd)
def train_again(yaml):
'''
Relaunch training of a model with conditions specified by the YAML
Looks for the model file defined by save path and replace the model
instanciated by the one that was trained before
-------------------------------------------------------------------
yaml : string, filename
YAML file defining the exp to be continued
'''
context = serial.load_train_file(yaml)
print "\tLoaded YAML"
# Load the trained model
model_file = context.save_path
for ext in range(len(context.extensions)):
if isinstance(context.extensions[ext],MonitorBasedSaveBest):
pos = ext
else:
raise AssertionError('No MonitorBasedSaveBest extension in the model!')
if not os.path.isfile(model_file):
model_file = context.extensions[pos].save_path
with open(model_file, 'r') as m_f:
trained_model = pkl.load(m_f)
# Define the continuing one
new_model = push_monitor(trained_model, 'trained_model',
transfer_experience=True)
# Define it as the model to be trained
context.model = new_model
context.save_path = context.extensions[pos].save_path[0:-4] + "_continue.pkl"
context.extensions[pos].save_path = context.save_path[0:-4] + "_best.pkl"
# Train again
context.main_loop()
def prepare_decoding_pipeline(options):
#prepare decoder from yaml or simply instantiate an empty instance
#with model and preprocessors = None
decoder = prepare_decoder(options.decoder_yaml)
if decoder.model is None:
if options.model_pkl is None:
raise Exception('Not specified how to build or load the model')
model = serial.load(options.model_pkl)
decoder.model = model
if decoder.preprocessor is None:
if options.prepr_yaml is None:
log.warning("Preprocessor is empty, feats will be fed into the model"
" as provided from the pipe or as read from archive.")
preprocessor = None
else:
preprocessor = serial.load_train_file(options.prepr_yaml)
decoder.preprocessor = preprocessor
if (options.model_pytables is not None) or \
(options.model_pytables_sd is not None):
params = load_model_params_from_pytables(options.model_pytables,
options.model_pytables_sd)
#TODO: add cross-check if parameters keys match 1:1 with the model
decoder.set_model_params(params)
log_priors = None
if options.priors_path is not None:
log_priors = load_kaldi_priors(options.priors_path)
decoder.build_fwdprop_function(log_priors)
assert isinstance(decoder, Pylearn2KaldiDecoderProvider)
assert isinstance(decoder.model, Model)
assert isinstance(decoder.preprocessor, OnlinePreprocessor) or decoder.preprocessor is None
assert decoder.fprop is not None
return decoder
def test_mnist_pi():
"""
Test the mnist_pi.yaml file from the maxout
paper on random input
"""
train = load_train_file(
os.path.join(pylearn2.__path__[0],
"scripts/papers/maxout/mnist_pi.yaml")
)
# Load fake MNIST data
init_value = control.load_data
control.load_data = [False]
train.dataset = MNIST(which_set='train', start=0, stop=100)
train.algorithm._set_monitoring_dataset(train.dataset)
control.load_data = init_value
# Train shortly and prevent saving
train.algorithm.termination_criterion = EpochCounter(max_epochs=1)
train.extensions.pop(0)
train.save_freq = 0
train.main_loop()
def train_vecs(self, source_vecs, target_vecs):
self.build_datasets(source_vecs, target_vecs)
from pylearn2.space import VectorSpace
from pylearn2.utils.serial import load_train_file
# TODO allow overrides via parameters
train = load_train_file(self.network_cfg)
# Change input layer size
model = train.model
if not isinstance(model, mlp.MLP):
raise RuntimeError("Provided network config does not use "
"MLP model -- not supported by this "
"translation model code")
input_size = self.source_vsm.layer1_size
model.set_input_space(VectorSpace(dim=input_size))
# Change output layer size
final_layer = model.layers[-1]
if not isinstance(final_layer, mlp.Linear):
raise RuntimeError("Provided network config does not have "
"a linear output layer -- not supported "
"by this translation model code")
output_size = self.target_vsm.layer1_size
# TODO is this sufficient for the linear layer? Might need to
# call some setter which updates internal state I don't
# understand
final_layer.dim = output_size
# Now begin training
train.main_loop()
self.network = train.model
self.make_network_fn()
def train(config, level_name=None, timestamp=None, time_budget=None,
verbose_logging=None, debug=None):
"""
Trains a given YAML file.
Parameters
----------
config : str
A YAML configuration file specifying the
training procedure.
level_name : bool, optional
Display the log level (e.g. DEBUG, INFO)
for each logged message.
timestamp : bool, optional
Display human-readable timestamps for
each logged message.
time_budget : int, optional
Time budget in seconds. Stop training at
the end of an epoch if more than this
number of seconds has elapsed.
verbose_logging : bool, optional
Display timestamp, log level and source
logger for every logged message
(implies timestamp and level_name are True).
debug : bool, optional
Display any DEBUG-level log messages,
False by default.
"""
train_obj = serial.load_train_file(config)
try:
iter(train_obj)
iterable = True
except TypeError:
iterable = False
# Undo our custom logging setup.
restore_defaults()
# Set up the root logger with a custom handler that logs stdout for INFO
# and DEBUG and stderr for WARNING, ERROR, CRITICAL.
root_logger = logging.getLogger()
if verbose_logging:
formatter = logging.Formatter(fmt="%(asctime)s %(name)s %(levelname)s "
"%(message)s")
handler = CustomStreamHandler(formatter=formatter)
else:
if timestamp:
prefix = '%(asctime)s '
else:
prefix = ''
formatter = CustomFormatter(prefix=prefix, only_from='pylearn2')
handler = CustomStreamHandler(formatter=formatter)
root_logger.addHandler(handler)
# Set the root logger level.
if debug:
root_logger.setLevel(logging.DEBUG)
else:
root_logger.setLevel(logging.INFO)
if iterable:
for number, subobj in enumerate(iter(train_obj)):
# Publish a variable indicating the training phase.
phase_variable = 'PYLEARN2_TRAIN_PHASE'
phase_value = 'phase%d' % (number + 1)
os.environ[phase_variable] = phase_value
# Execute this training phase.
subobj.main_loop(time_budget=time_budget)
# Clean up, in case there's a lot of memory used that's
# necessary for the next phase.
del subobj
gc.collect()
else:
train_obj.main_loop(time_budget=time_budget)
def main():
parser = argparse.ArgumentParser(description='Pylearn2 lab.')
parser.add_argument('-s',
'--save',
action='store_true',
help='Save the resulting images')
parser.add_argument(
'-q',
'--quit',
action='store_true',
help='Quit after plotting instead of dropping into IPython')
parser.add_argument('directory',
type=str,
help='Which results directory to use')
args = parser.parse_args()
# OLD
#config_file_path = '/home/jason/s/deep_learning/pylearn/pred_net.yaml'
#train = yaml_parse.load_path(config_file_path)
#train = serial.load_train_file(config_file_path)
#result_prefix = '/home/jason/s/pylearn2/pylearn2/pred/results/'
result_prefix = '/u/yosinski/s/galatea/fish/results/'
result_dir = os.path.join(result_prefix, args.directory)
print 'loading train object...'
#train = serial.load_train_file(os.path.join(result_dir, 'pred_net.yaml'))
train = serial.load_train_file(os.path.join(result_dir, 'model.yaml'))
print 'loading saved model...'
#model = serial.load(os.path.join(result_dir, 'pred_net.pkl'))
model = serial.load(os.path.join(result_dir, 'model.pkl'))
print 'done.'
print 'model was trained on:'
print model.dataset_yaml_src
if train.algorithm.cost is not None:
data_specs = train.algorithm.cost.get_data_specs(model)
else:
data_specs = train.model.get_default_cost().get_data_specs(train.model)
mapping = DataSpecsMapping(data_specs)
space_tuple = mapping.flatten(data_specs[0], return_tuple=True)
source_tuple = mapping.flatten(data_specs[1], return_tuple=True)
flat_data_specs = (CompositeSpace(space_tuple), source_tuple)
num_frames = model.num_frames
num_batches = 100
batch_size = train.algorithm.batch_size if train.algorithm.batch_size else 20 * num_frames
train_dataset = train.dataset
valid_dataset = train.algorithm.monitoring_dataset['valid']
rng = train.algorithm.rng
if not is_stochastic(train.algorithm.train_iteration_mode):
rng = None
train_iterator = train_dataset.iterator(
mode=train.algorithm.train_iteration_mode,
batch_size=batch_size,
data_specs=flat_data_specs,
return_tuple=True,
rng=rng,
num_batches=num_batches * 10)
valid_iterator = valid_dataset.iterator(
mode=train.algorithm.train_iteration_mode,
batch_size=batch_size,
data_specs=flat_data_specs,
return_tuple=True, # No rng override
num_batches=num_batches * 10)
train_batches = [train_iterator.next() for ii in range(num_batches)]
valid_batches = [valid_iterator.next() for ii in range(num_batches)]
print 'got batches with shape:'
for dat in train_batches[0]:
print ' ', dat.shape
#########################
# Plot costs
#########################
# Plot costs over time
ch_train_objective = model.monitor.channels['train_objective']
ch_valid_objective = model.monitor.channels['valid_objective']
x_vals = ch_train_objective.epoch_record
x_label = 'epoch'
plot(x_vals, ch_train_objective.val_record, 'b-')
plot(x_vals, ch_valid_objective.val_record, 'r-')
legend(('train', 'valid'))
if args.save:
savefig(os.path.join(result_dir, 'costs_lin.png'))
savefig(os.path.join(result_dir, 'costs_lin.pdf'))
if args.save:
gca().set_yscale('log')
savefig(os.path.join(result_dir, 'costs_log.png'))
savefig(os.path.join(result_dir, 'costs_log.pdf'))
gca().set_yscale('linear')
#########################
# Compute some accuracies
#########################
try:
model.fns.feat_to_compout
except:
model.redo_theano()
all_acc_id = []
all_xy_errs = []
print 'Training set:'
print ' acc_id\tx_err\ty_err'
for bb, batch in enumerate(train_batches):
feat, ids, xy = batch
idsN_floatX = array(ids.argmax(1), dtype=theano.config.floatX)
acc_id = model.fns.wiskott_id_accuracy(feat, idsN_floatX)
all_acc_id.append(acc_id)
xy_errs = model.fns.wiskott_xy_errors(feat, xy[:, 0:2])
all_xy_errs.append(xy_errs)
# Old numpy way
#ids_hat,xy_hat = model.fns.feat_to_idxy(feat)
#idx_true = np.where( ids == 1 )[1]
#idx_hat = np.where(np.sign(ids_hat.T - ids_hat.max(1)).T + 1)[1]
#n_correct += (idx_true == idx_hat).sum()
#n_total += len(idx_true)
print '%2d:\t%g,\t%g,\t%g' % (bb, acc_id, xy_errs[0], xy_errs[1])
#########################
# Embed
#########################
if not args.quit:
# Start shell
ipshell()
print 'done.'
space.validate(data)
inputs, targets = data
outputs = model.logistic_regression(inputs)
loss = -(targets * T.log(outputs)).sum(axis=1)
return loss.mean()
class LogisticRegression(Model):
def __init__(self, nvis, nclasses):
super(LogisticRegression, self).__init__()
self.nvis = nvis
self.nclasses = nclasses
W_value = numpy.random.uniform(size=(self.nvis, self.nclasses))
self.W = sharedX(W_value, 'W')
b_value = numpy.zeros(self.nclasses)
self.b = sharedX(b_value, 'b')
self._params = [self.W, self.b]
self.input_space = VectorSpace(dim=self.nvis)
self.output_space = VectorSpace(dim=self.nclasses)
def logistic_regression(self, inputs):
return T.nnet.softmax(T.dot(inputs, self.W) + self.b)
train_obj = serial.load_train_file('../files/log_reg.yaml')
train_obj.main_loop()
import numpy as np
from pylearn2.utils import serial
from theano import function
CODE_DIR = '/home/nico/Code/kaggle/GenderWrite/'
DATA_DIR = '/home/nico/datasets/Kaggle/GenderWrite/'
# load configuration file and train model
train_obj = serial.load_train_file(CODE_DIR+'gwmaxout.yaml')
# layer-by-layer training
#layers = train_obj.model.layers[:]
## first remove all but first and last layers and train first layer
#for ii in range(len(layers)-2):
# train_obj.model.layers.pop(1)
train_obj.main_loop()
# now add layers and re-train
#for ii in range(len(layers)-2):
# train_obj.model.layers.insert(1+ii,layers[1+ii])
# train_obj.main_loop()
# generate model output
def get_output(model, tdata, layerindex=-1, batch_size=100):
Xb = model.get_input_space().make_batch_theano()
ymf = model.fprop(Xb, return_all=True)
data = tdata.get_topological_view()
propagate = function([Xb],ymf)
def main():
parser = argparse.ArgumentParser(description='Pylearn2 lab.')
parser.add_argument('-s', '--save', action='store_true', help = 'Save the resulting images')
parser.add_argument('-q', '--quit', action='store_true', help = 'Quit after plotting instead of dropping into IPython')
parser.add_argument('directory', type = str,
help = 'Which results directory to use')
args = parser.parse_args()
# OLD
#config_file_path = '/home/jason/s/deep_learning/pylearn/pred_net.yaml'
#train = yaml_parse.load_path(config_file_path)
#train = serial.load_train_file(config_file_path)
#result_prefix = '/home/jason/s/pylearn2/pylearn2/pred/results/'
result_prefix = '/u/yosinski/s/galatea/fish/results/'
result_dir = os.path.join(result_prefix, args.directory)
print 'loading train object...'
#train = serial.load_train_file(os.path.join(result_dir, 'pred_net.yaml'))
train = serial.load_train_file(os.path.join(result_dir, 'model.yaml'))
print 'loading saved model...'
#model = serial.load(os.path.join(result_dir, 'pred_net.pkl'))
model = serial.load(os.path.join(result_dir, 'model.pkl'))
print 'done.'
print 'model was trained on:'
print model.dataset_yaml_src
if train.algorithm.cost is not None:
data_specs = train.algorithm.cost.get_data_specs(model)
else:
data_specs = train.model.get_default_cost().get_data_specs(train.model)
mapping = DataSpecsMapping(data_specs)
space_tuple = mapping.flatten(data_specs[0], return_tuple=True)
source_tuple = mapping.flatten(data_specs[1], return_tuple=True)
flat_data_specs = (CompositeSpace(space_tuple), source_tuple)
num_frames = model.num_frames
num_batches = 100
batch_size = train.algorithm.batch_size if train.algorithm.batch_size else 20*num_frames
train_dataset = train.dataset
valid_dataset = train.algorithm.monitoring_dataset['valid']
rng = train.algorithm.rng
if not is_stochastic(train.algorithm.train_iteration_mode):
rng = None
train_iterator = train_dataset.iterator(mode = train.algorithm.train_iteration_mode,
batch_size = batch_size,
data_specs = flat_data_specs,
return_tuple = True, rng=rng,
num_batches = num_batches * 10)
valid_iterator = valid_dataset.iterator(mode = train.algorithm.train_iteration_mode,
batch_size = batch_size,
data_specs = flat_data_specs,
return_tuple = True, # No rng override
num_batches = num_batches * 10)
train_batches = [train_iterator.next() for ii in range(num_batches)]
valid_batches = [valid_iterator.next() for ii in range(num_batches)]
print 'got batches with shape:'
for dat in train_batches[0]:
print ' ', dat.shape
#########################
# Plot costs
#########################
# Plot costs over time
ch_train_objective = model.monitor.channels['train_objective']
ch_valid_objective = model.monitor.channels['valid_objective']
x_vals = ch_train_objective.epoch_record
x_label = 'epoch'
plot(x_vals, ch_train_objective.val_record, 'b-')
plot(x_vals, ch_valid_objective.val_record, 'r-')
legend(('train', 'valid'))
if args.save:
savefig(os.path.join(result_dir, 'costs_lin.png'))
savefig(os.path.join(result_dir, 'costs_lin.pdf'))
if args.save:
gca().set_yscale('log')
savefig(os.path.join(result_dir, 'costs_log.png'))
savefig(os.path.join(result_dir, 'costs_log.pdf'))
gca().set_yscale('linear')
#########################
# Compute some accuracies
#########################
try:
model.fns.feat_to_compout
except:
model.redo_theano()
all_acc_id = []
all_xy_errs = []
print 'Training set:'
print ' acc_id\tx_err\ty_err'
for bb,batch in enumerate(train_batches):
feat,ids,xy = batch
idsN_floatX = array(ids.argmax(1), dtype=theano.config.floatX)
acc_id = model.fns.wiskott_id_accuracy(feat, idsN_floatX)
all_acc_id.append(acc_id)
xy_errs = model.fns.wiskott_xy_errors(feat, xy[:,0:2])
all_xy_errs.append(xy_errs)
# Old numpy way
#ids_hat,xy_hat = model.fns.feat_to_idxy(feat)
#idx_true = np.where( ids == 1 )[1]
#idx_hat = np.where(np.sign(ids_hat.T - ids_hat.max(1)).T + 1)[1]
#n_correct += (idx_true == idx_hat).sum()
#n_total += len(idx_true)
print '%2d:\t%g,\t%g,\t%g' % (bb, acc_id, xy_errs[0], xy_errs[1])
#########################
# Embed
#########################
if not args.quit:
# Start shell
ipshell()
print 'done.'
def main():
config_yaml = open("toy.yaml").read()
trainer = serial.load_train_file(
config_yaml % dict(save_path='tmp.tmp'))
trainer.main_loop()
#!/usr/bin/env python import pylearn2.utils.serial as serial if __name__=='__main__': import sys assert(len(sys.argv) == 2) serial.load_train_file(sys.argv[1]).main_loop()
from pylearn2.utils import serial
train_obj = serial.load_train_file('small_dataset.yaml')
train_obj.main_loop()
parser = argparse.ArgumentParser(
description="Launch an experiment from a YAML configuration file.",
epilog='\n'.join(__doc__.strip().split('\n')[1:]).strip(),
formatter_class=argparse.RawTextHelpFormatter)
parser.add_argument('config',
action='store',
choices=None,
help='A YAML configuration file specifying the '
'training procedure')
return parser
if __name__ == "__main__":
parser = make_argument_parser()
args = parser.parse_args()
train_obj = serial.load_train_file(args.config)
try:
iter(train_obj)
iterable = True
except TypeError as e:
iterable = False
if iterable:
for number, subobj in enumerate(iter(train_obj)):
# Publish a variable indicating the training phase.
phase_variable = 'PYLEARN2_TRAIN_PHASE'
phase_value = 'phase%d' % (number + 1)
os.environ[phase_variable] = phase_value
os.putenv(phase_variable, phase_value)
# Execute this training phase.
subobj.main_loop()
def train(config, level_name=None, timestamp=None, time_budget=None,
verbose_logging=None, debug=None):
"""
Trains a given YAML file.
Parameters
----------
config : str
A YAML configuration file specifying the
training procedure.
level_name : bool, optional
Display the log level (e.g. DEBUG, INFO)
for each logged message.
timestamp : bool, optional
Display human-readable timestamps for
each logged message.
time_budget : int, optional
Time budget in seconds. Stop training at
the end of an epoch if more than this
number of seconds has elapsed.
verbose_logging : bool, optional
Display timestamp, log level and source
logger for every logged message
(implies timestamp and level_name are True).
debug : bool, optional
Display any DEBUG-level log messages,
False by default.
"""
train_obj = serial.load_train_file(config)
try:
iter(train_obj)
iterable = True
except TypeError:
iterable = False
# Undo our custom logging setup.
restore_defaults()
# Set up the root logger with a custom handler that logs stdout for INFO
# and DEBUG and stderr for WARNING, ERROR, CRITICAL.
root_logger = logging.getLogger()
if verbose_logging:
formatter = logging.Formatter(fmt="%(asctime)s %(name)s %(levelname)s "
"%(message)s")
handler = CustomStreamHandler(formatter=formatter)
else:
if timestamp:
prefix = '%(asctime)s '
else:
prefix = ''
formatter = CustomFormatter(prefix=prefix, only_from='pylearn2')
handler = CustomStreamHandler(formatter=formatter)
root_logger.addHandler(handler)
# Set the root logger level.
if debug:
root_logger.setLevel(logging.DEBUG)
else:
root_logger.setLevel(logging.INFO)
if iterable:
for number, subobj in enumerate(iter(train_obj)):
# Publish a variable indicating the training phase.
phase_variable = 'PYLEARN2_TRAIN_PHASE'
phase_value = 'phase%d' % (number + 1)
os.environ[phase_variable] = phase_value
# Execute this training phase.
subobj.main_loop(time_budget=time_budget)
# Clean up, in case there's a lot of memory used that's
# necessary for the next phase.
del subobj
gc.collect()
else:
train_obj.main_loop(time_budget=time_budget)
OUT = "/yaml/test.yaml"
TEMPLATE = DIR+"template.yaml"
HPARAMS = "hparams.conf"
if __name__ == "__main__":
# for transformation in ['translate','scale','rotate','flip','gaussian','sharpen','denoize','occlusion','halfface']:
# for transformation in ['scale','rotate','flip','gaussian','sharpen','denoize','occlusion','halfface']:
for transformation in ['denoize','sharpen']:
out = DIR+transformation+OUT
t_template = "".join(open(DIR+transformation+"/"+transformation+".yaml",'r'))
# Generates a list of hyper-parameter names and a list of
# hyper-parameter values
hpnames, hpvalues = generate_params(hparamfile=DIR+transformation+"/"+transformation+".conf",
generate="log-uniform",
search_mode="fix-grid-search")
template = "".join(open(TEMPLATE,'r')) % {'transformations': t_template,'save_path':'%(save_path)s'}
# Writes template with each hyper-parameter settings in
# succesive files and returns the name of the files
files = write_files(template=template,hpnames=hpnames,
hpvalues=hpvalues,save_path=out,force=True)
# files = write_files(template="".join(open(TEMPLATE),'r'),hpnames=hpnames,
# hpvalues=hpvalues,save_path=OUT)
for f in files:
serial.load_train_file(f).main_loop()
def main():
config_yaml = open("toy.yaml").read()
trainer = serial.load_train_file(config_yaml % dict(save_path='tmp.tmp'))
trainer.main_loop()
#!/usr/bin/env python
import pylearn2.utils.serial as serial
if __name__ == '__main__':
import sys
assert (len(sys.argv) == 2)
serial.load_train_file(sys.argv[1]).main_loop()
def create_model(yaml_path):
from pylearn2.models.model import Model
model = serial.load_train_file(yaml_path)
assert isinstance(model, Model)
return model
'(implies -T).')
parser.add_argument('--debug', '-D',
action='store_true',
help='Display any DEBUG-level log messages, '
'suppressed by default.')
parser.add_argument('config', action='store',
choices=None,
help='A YAML configuration file specifying the '
'training procedure')
return parser
if __name__ == "__main__":
parser = make_argument_parser()
args = parser.parse_args()
train_obj = serial.load_train_file(args.config)
try:
iter(train_obj)
iterable = True
except TypeError as e:
iterable = False
# Undo our custom logging setup.
restore_defaults()
# Set up the root logger with a custom handler that logs stdout for INFO
# and DEBUG and stderr for WARNING, ERROR, CRITICAL.
root_logger = logging.getLogger()
if args.verbose_logging:
formatter = logging.Formatter(fmt="%(asctime)s %(name)s %(levelname)s "
"%(message)s")
handler = CustomStreamHandler(formatter=formatter)
#! /usr/bin/env python
from pylearn2.utils import serial
import sys
if __name__ == "__main__":
train_obj = serial.load_train_file(sys.argv[1])
train_obj.main_loop()
import numpy as np
from pylearn2.utils import serial
from theano import function
CODE_DIR = '/home/nico/Code/kaggle/GenderWrite/'
DATA_DIR = '/home/nico/datasets/Kaggle/GenderWrite/'
# load configuration file and train model
train_obj = serial.load_train_file(CODE_DIR + 'gwmaxout.yaml')
# layer-by-layer training
#layers = train_obj.model.layers[:]
## first remove all but first and last layers and train first layer
#for ii in range(len(layers)-2):
# train_obj.model.layers.pop(1)
train_obj.main_loop()
# now add layers and re-train
#for ii in range(len(layers)-2):
# train_obj.model.layers.insert(1+ii,layers[1+ii])
# train_obj.main_loop()
# generate model output
def get_output(model, tdata, layerindex=-1, batch_size=100):
Xb = model.get_input_space().make_batch_theano()
ymf = model.fprop(Xb, return_all=True)
data = tdata.get_topological_view()
from __future__ import print_function
from pylearn2.utils import serial
import matplotlib.pyplot as plt
import numpy as np
train_obj = serial.load_train_file('pylearn_svm.yaml')
train_obj.main_loop()
W, b = train_obj.model.get_param_values()
print(W)
print(b)
X = train_obj.dataset.X
x = np.linspace(-10, 10, 200)
f_x = (-1.0 * W[0][0] / W[1][0]) * x - (b[0][0] / W[1][0])
plt.plot(x, f_x)
plt.scatter(X[:, 0], X[:, 1])
plt.show()
def main(args=None):
usage = "pool_adaptation.py [options] <si-model-dir> <sa-model-dir> <feats-scp> <targets-pdf> "
parser = OptionParser()
parser.add_option("--adapt-yaml", dest="adapt_yaml", default="",
help="Provide the adaptation yaml template to start with")
parser.add_option("--freeze-means", dest="freeze_means", default=False,
help="Skip means while updating pools")
parser.add_option("--freeze-betas", dest="freeze_betas", default=False,
help="Skip precisions while updating pools")
parser.add_option("--freeze-amp", dest="freeze_amp", default="true",
help="Skip activation function amplitudes while updating pools")
parser.add_option("--freeze-slopes", dest="freeze_slopes", default="true",
help="Skip activation function slopes while updating pools")
parser.add_option("--freeze-layer-ids", dest="freeze_layer_ids", default="",
help="update params only in this layers, i.e. --layer-ids 012")
parser.add_option("--job", dest="JOB", default=0,
help="JOB ID used to store model in")
parser.add_option("--debug", dest="debug", default=False,
help="Prints activations and shapes in text format rather than binary Kaldi archives")
(options,args) = parser.parse_args(args=args)
print options.adapt_yaml
print options.freeze_means
print options.freeze_betas
print options.freeze_amp
print options.freeze_slopes
print options.freeze_layer_ids
print options.JOB
print 'ARGS: ', args
#if options.adapt_yaml!='':
# NotImplementedError('Lodaing from pkl not yet supported due to GPU/CPU pickle issues.')
if len(args) != 5:
print usage
exit(1)
si_model_dir = args[1]
sa_model_dir = args[2]
feats_scp = args[3]
targets_pdf = args[4]
#print "si model dir is %s"%si_model_dir
model_yaml = "%s/adapt_final%s.yaml"%(si_model_dir, options.JOB)
model_params = "%s/cnn_best.h5"%si_model_dir
#print 'Yaml path', model_yaml
#print 'Params path', model_params
if not os.path.isfile(options.adapt_yaml):
raise Exception('File %s not found'%options.adapt_yaml)
if not os.path.isfile(model_params):
raise Exception('File %s not found'%model_params)
vars={}
vars['adapt_flist']=feats_scp
vars['adapt_pdfs']=targets_pdf
vars['adapt_lr']=0.05
vars['adapt_momentum']=0.5
vars['sa_dir'] = sa_model_dir
vars['JOB'] = options.JOB
#print vars
#print 'Locals: ',locals()
adapt_template = open(options.adapt_yaml, 'r').read()
adapt_template_str = adapt_template % vars
f = open(model_yaml, 'w')
f.write(adapt_template_str)
f.close()
print 'Building model %s'%model_yaml
train_obj = serial.load_train_file(model_yaml)
print 'Loading params from %s'%model_params
params = ModelPyTables.pytables_to_params(model_params, name='Model')
train_obj.model.set_params(params)
freeze_regex='softmax_[Wb]|h[0-9]_[Wb]|nlrf_[Wb]'
if options.freeze_layer_ids!='':
layers = options.freeze_layer_ids
freeze_regex = "%s|g[%s]p_u|g[%s]p_beta"%(freeze_regex, layers, layers)
if options.freeze_means == 'true':
freeze_regex = "%s|g[0-9]p_u"%(freeze_regex)
if options.freeze_betas == 'true':
freeze_regex = "%s|g[0-9]p_beta"%(freeze_regex)
if options.freeze_amp == 'true':
freeze_regex = "%s|g[0-9]p_amp"%(freeze_regex)
if options.freeze_slopes == 'true':
freeze_regex = "%s|g[0-9]p_arg"%(freeze_regex)
#print "Freeze regex is", freeze_regex
model_params = train_obj.model.get_params()
params_to_freeze = {}
for param in model_params:
if re.match(freeze_regex, str(param)) is not None:
if param not in params_to_freeze:
params_to_freeze[param] = param
#print params_to_freeze
if len(params_to_freeze)==len(model_params):
print 'None of the parameters were set to be updated. Freeze list is', params_to_freeze
exit(0)
train_obj.model.freeze(params_to_freeze.values())
print 'Will update those params only: ', train_obj.model.get_params()
train_obj.main_loop()
train_obj.model.freeze_set = set([]) #unfreeze so get_params will return all model params
print 'Unfreezed params are ', train_obj.model.get_params()
