def sanity_check(conf_file, result, **be_args):
experiment = deserialize(os.path.join(dir, conf_file))
backend = gen_backend(model=experiment.model, **be_args)
experiment.initialize(backend)
res = experiment.run()
print(float(res['test']['MisclassRate_TOP_1']))
assert float(res['test']['MisclassRate_TOP_1']) == result
def serialize_check(conf_file, result, tol, res_string, **be_args):
experiment = deserialize(conf_file)
backend = gen_backend(model=experiment.model, **be_args)
experiment.initialize(backend)
res = experiment.run()
print float(res[res_string]['MisclassPercentage_TOP_1']), result,
assert abs(
float(res[res_string]['MisclassPercentage_TOP_1']) - result) < tol
def speed_check(conf_file, num_epochs, **be_args):
experiment = deserialize(os.path.join(dir, conf_file))
experiment.model.num_epochs = num_epochs
backend = gen_backend(model=experiment.model, **be_args)
experiment.initialize(backend)
start = time.time()
experiment.run()
return (time.time() - start)
def speed_check(conf_file, num_epochs, **be_args):
experiment = deserialize(os.path.join(dir, conf_file))
experiment.model.num_epochs = num_epochs
backend = gen_backend(model=experiment.model, **be_args)
experiment.initialize(backend)
start = time.time()
experiment.run()
return (time.time() - start)
def model_compare(model1_file, model2_file, atol=0.0, rtol=0.0):
model1 = deserialize(model1_file)
model2 = deserialize(model2_file)
assert model1.keys().sort() == model2.keys().sort()
# remove the epochs from the dictionaries and compare them
assert model1.pop('epochs_complete') == model2.pop('epochs_complete')
# for MLP just layers should be left?
print('checking the 1 versus k step outputs...')
for ky in model1.keys():
print(ky)
assert TestSerialization.layer_compare(model1[ky],
model2[ky],
atol=atol,
rtol=rtol)
print('OK')
return True
def model_compare(model1_file, model2_file, atol=0.0, rtol=0.0):
model1 = deserialize(model1_file)
model2 = deserialize(model2_file)
assert model1.keys().sort() == model2.keys().sort()
# remove the epochs from the dictionaries and compare them
assert model1.pop('epochs_complete') == model2.pop('epochs_complete')
# for MLP just layers should be left?
print('checking the 1 versus k step outputs...')
for ky in model1.keys():
print(ky)
assert TestSerialization.layer_compare(model1[ky],
model2[ky],
atol=atol,
rtol=rtol)
print('OK')
return True
def load_file(self, filename, nclasses):
logger.info("loading: %s", filename)
dict = deserialize(filename)
full_image = np.float32(dict["data"])
full_image /= 255.0
labels = np.array(dict["labels"])
onehot = np.zeros((len(labels), nclasses), dtype="float32")
for col in range(nclasses):
onehot[:, col] = labels == col
return (full_image, onehot)
def load_file(self, filename, nclasses):
logger.info('loading: %s', filename)
dict = deserialize(filename)
full_image = np.float32(dict['data'])
full_image /= 255.
labels = np.array(dict['labels'])
onehot = np.zeros((len(labels), nclasses), dtype='float32')
for col in range(nclasses):
onehot[:, col] = (labels == col)
return (full_image, onehot)
def load_file(self, filename, nclasses):
logger.info('loading: %s', filename)
dict = deserialize(filename)
full_image = np.float32(dict['data'])
full_image /= 255.
labels = np.array(dict['labels'])
onehot = np.zeros((len(labels), nclasses), dtype='float32')
for col in range(nclasses):
onehot[:, col] = (labels == col)
return (full_image, onehot)
def run(self):
"""
Actually carry out each of the experiment steps.
"""
# load the dataset, save it to disk if specified
self.dataset.set_batch_size(self.model.batch_size)
self.dataset.backend = self.backend
self.dataset.load(backend=self.backend, experiment=self)
if hasattr(self.dataset,
'serialized_path') and (self.dataset.serialized_path
is not None):
logger.warning('Ability to serialize dataset has been deprecated.')
# fit the model to the data, save it if specified
if not hasattr(self.model, 'backend'):
self.model.backend = self.backend
if not hasattr(self.model, 'epochs_complete'):
self.model.epochs_complete = 0
mfile = ''
if hasattr(self.model, 'deserialized_path'):
mfile = os.path.expandvars(
os.path.expanduser(self.model.deserialized_path))
elif hasattr(self.model, 'serialized_path'):
mfile = os.path.expandvars(
os.path.expanduser(self.model.serialized_path))
elif self.live:
raise RuntimeError('Live inference requires a saved model')
if os.access(mfile, os.R_OK):
if self.backend.is_distributed():
raise NotImplementedError('Deserializing models not supported '
'in distributed mode')
self.model.set_params(deserialize(mfile))
elif mfile != '':
logger.info('Unable to find saved model %s, starting over', mfile)
if self.live:
raise RuntimeError('Live inference requires a saved model')
if self.model.epochs_complete >= self.model.num_epochs:
return
if self.live:
return
self.model.fit(self.dataset)
if hasattr(self.model, 'serialized_path'):
if self.backend.rank() == 0:
serialize(self.model.get_params(), self.model.serialized_path)
def call_neon(params):
"""
runs the system call to neon and reads the result to give back to sm
"""
timestring = str(int(time.time()))
experiment_dir = os.path.realpath(os.environ['HYPEROPT_PATH'])
# Generate the yaml file
hyper_file = os.path.join(experiment_dir, 'hyperyaml.yaml')
yaml_file = os.path.join(experiment_dir, 'yamels',
'temp' + timestring + '.yaml')
try:
os.mkdir(os.path.join(experiment_dir, 'yamels'))
except OSError:
"Directory exists"
write_params(hyper_file, yaml_file, params)
# Initialize the neon experiment
logging.basicConfig(level=20)
experiment = deserialize(yaml_file)
backend = gen_backend(model=experiment.model) # , gpu='nervanagpu'
experiment.initialize(backend)
# ensure TOP1 error is calculated
if not hasattr(experiment, 'metrics'):
experiment.metrics = {
'validation': [MisclassPercentage(error_rank=1)],
'test': [MisclassPercentage(error_rank=1)]
}
for item in ['validation', 'test']:
if item not in experiment.metrics:
experiment.metrics[item] = [MisclassPercentage(error_rank=1)]
metriclist = [str(x) for x in experiment.metrics[item]]
if 'MisclassPercentage_TOP_1' not in metriclist:
experiment.metrics[item].append(MisclassPercentage(error_rank=1))
result = experiment.run()
# check if validation set is available
if experiment.dataset.has_set('validation'):
hyperopt_set = 'validation'
elif experiment.dataset.has_set('test'):
hyperopt_set = 'test'
print("Warning: No validation set found, performing hyperparameter "
"optimization on test set.")
else:
raise AttributeError("No error found.")
return result[hyperopt_set]['MisclassPercentage_TOP_1']
def run(self):
"""
Actually carry out each of the experiment steps.
"""
# load the dataset, save it to disk if specified
self.dataset.set_batch_size(self.model.batch_size)
self.dataset.backend = self.backend
self.dataset.load(backend=self.backend, experiment=self)
if hasattr(self.dataset, 'serialized_path') and (
self.dataset.serialized_path is not None):
logger.warning('Ability to serialize dataset has been deprecated.')
# fit the model to the data, save it if specified
if not hasattr(self.model, 'backend'):
self.model.backend = self.backend
if not hasattr(self.model, 'epochs_complete'):
self.model.epochs_complete = 0
mfile = ''
if hasattr(self.model, 'deserialized_path'):
mfile = os.path.expandvars(os.path.expanduser(
self.model.deserialized_path))
elif hasattr(self.model, 'serialized_path'):
mfile = os.path.expandvars(os.path.expanduser(
self.model.serialized_path))
elif self.live:
raise RuntimeError('Live inference requires a saved model')
if os.access(mfile, os.R_OK):
if self.backend.is_distributed():
raise NotImplementedError('Deserializing models not supported '
'in distributed mode')
self.model.set_params(deserialize(mfile))
elif mfile != '':
logger.info('Unable to find saved model %s, starting over', mfile)
if self.live:
raise RuntimeError('Live inference requires a saved model')
if self.model.epochs_complete >= self.model.num_epochs:
return
if self.live:
return
self.model.fit(self.dataset)
if hasattr(self.model, 'serialized_path'):
if self.backend.rank() == 0:
serialize(self.model.get_params(), self.model.serialized_path)
def call_neon(params):
"""
runs the system call to neon and reads the result to give back to sm
"""
timestring = str(int(time.time()))
experiment_dir = os.path.realpath(os.environ['HYPEROPT_PATH'])
# Generate the yaml file
hyper_file = os.path.join(experiment_dir, 'hyperyaml.yaml')
yaml_file = os.path.join(experiment_dir, 'yamels',
'temp' + timestring + '.yaml')
try:
os.mkdir(os.path.join(experiment_dir, 'yamels'))
except OSError:
"Directory exists"
write_params(hyper_file, yaml_file, params)
# Initialize the neon experiment
logging.basicConfig(level=20)
experiment = deserialize(yaml_file)
backend = gen_backend(model=experiment.model) # , gpu='nervanagpu'
experiment.initialize(backend)
# ensure TOP1 error is calculated
if not hasattr(experiment, 'metrics'):
experiment.metrics = {'validation': [MisclassPercentage(error_rank=1)],
'test': [MisclassPercentage(error_rank=1)]}
for item in ['validation', 'test']:
if item not in experiment.metrics:
experiment.metrics[item] = [MisclassPercentage(error_rank=1)]
metriclist = [str(x) for x in experiment.metrics[item]]
if 'MisclassPercentage_TOP_1' not in metriclist:
experiment.metrics[item].append(MisclassPercentage(error_rank=1))
result = experiment.run()
# check if validation set is available
if experiment.dataset.has_set('validation'):
hyperopt_set = 'validation'
elif experiment.dataset.has_set('test'):
hyperopt_set = 'test'
print("Warning: No validation set found, performing hyperparameter "
"optimization on test set.")
else:
raise AttributeError("No error found.")
return result[hyperopt_set]['MisclassPercentage_TOP_1']
def load(self, backend=None, experiment=None):
'''
Imageset only supports nervanagpu based backends
'''
if not hasattr(self.backend, 'ng'):
raise DeprecationWarning("Only nervanagpu-based backends "
"supported. For using cudanet backend, "
"revert to neon 0.8.2 ")
bdir = os.path.expanduser(self.save_dir)
cachefile = os.path.join(bdir, 'dataset_cache.pkl')
if not os.path.exists(cachefile):
logger.error("Batch dir cache not found in %s:", cachefile)
response = raw_input("Press Y to create, otherwise exit: ")
if response == 'Y':
from neon.util.batch_writer import (BatchWriter,
BatchWriterImagenet)
if self.imageset.startswith('I1K'):
self.bw = BatchWriterImagenet(**self.__dict__)
else:
self.bw = BatchWriter(**self.__dict__)
self.bw.run()
logger.error('Done writing batches - please rerun to train.')
else:
logger.error('Exiting...')
sys.exit()
cstats = deserialize(cachefile, verbose=False)
if cstats['macro_size'] != self.macro_size:
raise NotImplementedError("Cached macro size %d different from "
"specified %d, delete save_dir %s "
"and try again.",
cstats['macro_size'],
self.macro_size,
self.save_dir)
# Set the max indexes of batches for each from the cache file
self.maxval = cstats['nval'] + cstats['val_start'] - 1
self.maxtrain = cstats['ntrain'] + cstats['train_start'] - 1
# Make sure only those properties not by yaml are updated
cstats.update(self.__dict__)
self.__dict__.update(cstats)
# Should also put (in addition to nclass), number of train/val images
req_param(self, ['ntrain', 'nval', 'train_start', 'val_start',
'train_mean', 'val_mean', 'labels_dict'])
def load(self, backend=None, experiment=None):
'''
Imageset only supports nervanagpu based backends
'''
if not hasattr(self.backend, 'ng'):
raise DeprecationWarning("Only nervanagpu-based backends "
"supported. For using cudanet backend, "
"revert to neon 0.8.2 ")
bdir = os.path.expanduser(self.save_dir)
cachefile = os.path.join(bdir, 'dataset_cache.pkl')
if not os.path.exists(cachefile):
logger.error("Batch dir cache not found in %s:", cachefile)
response = raw_input("Press Y to create, otherwise exit: ")
if response == 'Y':
from neon.util.batch_writer import (BatchWriter,
BatchWriterImagenet)
if self.imageset.startswith('I1K'):
self.bw = BatchWriterImagenet(**self.__dict__)
else:
self.bw = BatchWriter(**self.__dict__)
self.bw.run()
logger.error('Done writing batches - please rerun to train.')
else:
logger.error('Exiting...')
sys.exit()
cstats = deserialize(cachefile, verbose=False)
if cstats['macro_size'] != self.macro_size:
raise NotImplementedError("Cached macro size %d different from "
"specified %d, delete save_dir %s "
"and try again.",
cstats['macro_size'],
self.macro_size,
self.save_dir)
# Set the max indexes of batches for each from the cache file
self.maxval = cstats['nval'] + cstats['val_start'] - 1
self.maxtrain = cstats['ntrain'] + cstats['train_start'] - 1
# Make sure only those properties not by yaml are updated
cstats.update(self.__dict__)
self.__dict__.update(cstats)
# Should also put (in addition to nclass), number of train/val images
req_param(self, ['ntrain', 'nval', 'train_start', 'val_start',
'train_mean', 'val_mean', 'labels_dict'])
def train():
save_file = sys.argv[1]
if len(sys.argv) > 2:
model = deserialize(sys.argv[2])
else:
layers = get_parameters(n_in=FEATURE_LENGTH, n_hidden_units=[100, 50, NUM_CLASSES])
# define model
model = MLP(num_epochs=1, batch_size=MINIBATCH_SIZE,
layers=layers, epochs_complete=0)
model.link()
#be.configure(model, datapar=False, modelpar=False)
model.initialize(be)
model.data_layer = model.layers[0]
model.cost_layer = model.layers[-1]
dataset = Fly(backend=be,
repo_path=os.path.expanduser('~/flyvfly/'))
# par related init
be.actual_batch_size = model.batch_size
be.mpi_size = 1
be.mpi_rank = 0
be.par = NoPar()
be.par.backend = be
max_macro_epochs = 1000
min_err = sys.maxint
for i in range(max_macro_epochs):
model.epochs_complete = 0
dataset.use_set = "train"
model.fit(dataset)
#scores, targets = model.predict_fullset(dataset, "validation")
val_err = get_validation(model, dataset)
logger.info('epoch: %d, valid error: %0.6f', i, val_err)
if val_err < min_err:
serialize(model, save_file)
min_err = val_err
def run_experiment_in_steps(self,
config_file,
n,
k,
be,
init_config=None,
**be_args):
# run an experiment for N epochs in k stepe with n/k
# epochs per step. Last step will be enough epochs
# to reach n total epochs. Between each step the
# model will be serialized and saved, then reloaded
# from that saved file at the next step if init_config
# is not None, then that file will be used # as the
# initial deserialize file - this is used for handing off
# models run on cpu.gpu backends to running on gpu/cpu backends
stepsize = n / k
last_saved_state = None
for ind in range(1, k + 1):
# run the same learning with N/k epochs k times
# each time saving and reloading the serialized model state
if ind == k:
# in case N/k is a fraction
# last step will end at epoch N
end_epoch = n
else:
end_epoch = ind * stepsize
# load up base experiment config
experiment = deserialize(config_file)
# run for N/k steps
experiment.model.num_epochs = end_epoch
if ind > 1:
# after step 1 need to load initial config
# from last runs serialized file
experiment.model.deserialized_path = last_saved_state
elif init_config is not None:
# used the given pickle file to initialize the mdoel
experiment.model.deserialized_path = init_config
# save the model to this file
last_saved_state = os.path.join(self.model_path,
'%d_step_%d.prm' % (k, end_epoch))
print(last_saved_state)
if os.path.exists(last_saved_state):
print('removing %s' % last_saved_state)
os.remove(last_saved_state)
experiment.model.serialized_path = last_saved_state
experiment.model.serialize_schedule = k
if k == 1:
# keep copies of all checkpoint files for cp tests
experiment.model.save_checkpoints = n
backend = gen_backend(model=experiment.model, **be_args)
experiment.initialize(backend)
if ind == 1:
# save the initial weights for check with other runs
intial_weights = {}
for ind, layer in enumerate(experiment.model.layers):
if hasattr(layer, 'weights'): # only checking weights
# ensure unique layer names
ln = '%s_%d' % (layer.name, ind)
intial_weights[ln] = np.copy(
layer.weights.asnumpyarray())
experiment.run()
return (last_saved_state, intial_weights)
def get_macro_batch(self):
self.macro_idx = (self.macro_idx + 1 - self.startb) \
% self.nmacros + self.startb
fname = os.path.join(self.save_dir,
'data_batch_{:d}'.format(self.macro_idx))
return deserialize(os.path.expanduser(fname), verbose=False)
def get_macro_batch(self):
self.macro_idx = (self.macro_idx + 1 - self.startb) \
% self.nmacros + self.startb
fname = os.path.join(self.save_dir,
'data_batch_{:d}'.format(self.macro_idx))
return deserialize(os.path.expanduser(fname), verbose=False)
model.data_layer.use_set('test', predict=True)
dataset.use_set = 'test'
scores, targets = model.predict_fullset(dataset, "test")
scores_ts = np.transpose(scores.asnumpyarray())
targets_ts = np.transpose(targets.asnumpyarray())
model_no = sys.argv[1].split(".")[0][-2:]
find_no_class(targets_ts[:, 5], scores_ts[:, 5], targets_tr[:, 5], scores_tr[:, 5], model_no)
#prc_curve(targets_ts, scores_ts, targets_tr, scores_tr, model_no)
def visualize():
weights = model.layers[-2].weights.asnumpyarray()
np.savetxt("mcmodel12weights3.txt", weights)
np.savetxt("mcmodel12weights2.txt", model.layers[-3].weights.asnumpyarray())
np.savetxt("mcmodel12weights1.txt", model.layers[-4].weights.asnumpyarray())
plt.subplot(1, 2, 1)
plt.imshow(np.transpose(np.sort(abs(model.layers[-2].weights.asnumpyarray()))), cmap = cm.Greys_r)
plt.subplot(1, 2, 2)
plt.imshow(np.sort(np.transpose(abs(model.layers[-3].weights.asnumpyarray()))), cmap = cm.Greys_r)
plt.show()
weights_sort = weights.argsort()
max_weights = weights[0, weights_sort[0, -5:]]
min_weights = weights[0, weights_sort[0, 0:5]]
print min_weights
if __name__ == '__main__':
with open(sys.argv[1], 'r') as f:
model = deserialize(f)
model.print_layers()
visualize()
#test()
def serialize_check_alexnet(conf_file, result, **be_args):
experiment = deserialize(os.path.join(dir, conf_file))
backend = gen_backend(model=experiment.model, **be_args)
experiment.initialize(backend)
res = experiment.run()
print float(res['validation']['MisclassPercentage_TOP_1']), result
def run_experiment_in_steps(self, config_file, n, k, be,
init_config=None, **be_args):
# run an experiment for N epochs in k stepe with n/k
# epochs per step. Last step will be enough epochs
# to reach n total epochs. Between each step the
# model will be serialized and saved, then reloaded
# from that saved file at the next step if init_config
# is not None, then that file will be used # as the
# initial deserialize file - this is used for handing off
# models run on cpu.gpu backends to running on gpu/cpu backends
stepsize = n/k
last_saved_state = None
for ind in range(1, k+1):
# run the same learning with N/k epochs k times
# each time saving and reloading the serialized model state
if ind == k:
# in case N/k is a fraction
# last step will end at epoch N
end_epoch = n
else:
end_epoch = ind*stepsize
# load up base experiment config
experiment = deserialize(config_file)
# run for N/k steps
experiment.model.num_epochs = end_epoch
if ind > 1:
# after step 1 need to load initial config
# from last runs serialized file
experiment.model.deserialized_path = last_saved_state
elif init_config is not None:
# used the given pickle file to initialize the mdoel
experiment.model.deserialized_path = init_config
# save the model to this file
last_saved_state = os.path.join(self.model_path,
'%d_step_%d.prm' % (k, end_epoch))
print(last_saved_state)
if os.path.exists(last_saved_state):
print('removing %s' % last_saved_state)
os.remove(last_saved_state)
experiment.model.serialized_path = last_saved_state
experiment.model.serialize_schedule = k
if k == 1:
# keep copies of all checkpoint files for cp tests
experiment.model.save_checkpoints = n
backend = gen_backend(model=experiment.model, **be_args)
experiment.initialize(backend)
if ind == 1:
# save the initial weights for check with other runs
intial_weights = {}
for ind, layer in enumerate(experiment.model.layers):
if hasattr(layer, 'weights'): # only checking weights
# ensure unique layer names
ln = '%s_%d' % (layer.name, ind)
intial_weights[ln] = np.copy(
layer.weights.asnumpyarray()
)
experiment.run()
return (last_saved_state, intial_weights)
