def eval_acts(self, inp):
""" Evaluate Tagger on given inp data
Here we are only evaluating the clean path."""
clean_path_kv = [[k, v] for k, v in self.clean.iteritems()]
givens = {}
givens.update({self.x: inp['features_labeled']})
givens.update({self.y: inp['targets_labeled']})
givens.update({self.x_only: inp['features_unlabeled']})
givens.update(
{self.masks_unlabeled: np.float32(inp['masks_unlabeled'])})
params, args = zip(*givens.iteritems())
# on_unused_input is set to ignore to omit some warning prints. It can be changed accordingly base on your use
# cases.
function = theano.function(params, [k[1] for k in clean_path_kv],
on_unused_input='ignore')
clean_path_v_outputs = function(*args)
acts = AttributeDict(clean=AttributeDict(
zip([k[0] for k in clean_path_kv], clean_path_v_outputs)))
return acts
def encoder(input_, path_name, input_noise_std=0, noise_std=[]):
h = input_
logger.info(' 0: noise %g' % input_noise_std)
if input_noise_std > 0.:
h = h + self.noise_like(h) * input_noise_std
d = AttributeDict()
d.unlabeled = self.new_activation_dict()
d.labeled = self.new_activation_dict()
d.labeled.z[0] = self.labeled(h)
d.unlabeled.z[0] = self.unlabeled(h)
prev_dim = input_dim
for i, (spec, _, act_f) in layers[1:]:
d.labeled.h[i - 1], d.unlabeled.h[i - 1] = self.split_lu(h)
noise = noise_std[i] if i < len(noise_std) else 0.
curr_dim, z, m, s, h = self.f(h,
prev_dim,
spec,
i,
act_f,
path_name=path_name,
noise_std=noise)
assert self.layer_dims.get(i) in (None, curr_dim)
self.layer_dims[i] = curr_dim
d.labeled.z[i], d.unlabeled.z[i] = self.split_lu(z)
d.unlabeled.s[i] = s
d.unlabeled.m[i] = m
prev_dim = curr_dim
d.labeled.h[i], d.unlabeled.h[i] = self.split_lu(h)
return d
def build_datasets(self, dataset_class, p):
train_split = ['train']
train_set = dataset_class(which_sets=train_split)
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
rng = numpy.random.RandomState(seed=p.seed)
rng.shuffle(all_ind)
valid_set = dataset_class(which_sets=["valid"])
valid_ind = numpy.arange(valid_set.num_examples)
trn_set_size = p.get('train_set_size', None)
train_ind = all_ind[:trn_set_size]
test_split = ['test']
test_set = dataset_class(which_sets=test_split)
test_ind = numpy.arange(test_set.num_examples)
trn = AttributeDict(set=train_set,
ind=train_ind,
batch_size=p.batch_size)
val = AttributeDict(set=valid_set,
ind=valid_ind,
batch_size=p.valid_batch_size)
tst = AttributeDict(set=test_set,
ind=test_ind,
batch_size=p.valid_batch_size)
return trn, val, tst
def encoder(input_, path_name, input_noise_std=0, noise_std=[]):
h = input_
logger.info(' 0: noise %g' % input_noise_std)
if input_noise_std > 0.:
h = h + self.noise_like(h) * input_noise_std
d = AttributeDict()
d.unlabeled = self.new_activation_dict()
d.labeled = self.new_activation_dict()
d.labeled.z[0] = self.labeled(h)
d.unlabeled.z[0] = self.unlabeled(h)
prev_dim = input_dim
for i, (spec, _, act_f) in layers[1:]:
d.labeled.h[i - 1], d.unlabeled.h[i - 1] = self.split_lu(h)
noise = noise_std[i] if i < len(noise_std) else 0.
curr_dim, z, m, s, h = self.f(h, prev_dim, spec, i, act_f,
path_name=path_name,
noise_std=noise)
assert self.layer_dims.get(i) in (None, curr_dim)
self.layer_dims[i] = curr_dim
d.labeled.z[i], d.unlabeled.z[i] = self.split_lu(z)
d.unlabeled.s[i] = s
d.unlabeled.m[i] = m
prev_dim = curr_dim
d.labeled.h[i], d.unlabeled.h[i] = self.split_lu(h)
return d
def load_and_log_params(cli_params):
cli_params = AttributeDict(cli_params)
if cli_params.get('load_from'):
p = load_df(cli_params.load_from, 'params').to_dict()[0]
p = AttributeDict(p)
for key in cli_params.keys():
if key not in p:
p[key] = None
new_params = cli_params
loaded = True
else:
p = cli_params
new_params = {}
loaded = False
# Make dseed seed unless specified explicitly
if p.get('dseed') is None and p.get('seed') is not None:
p['dseed'] = p['seed']
logger.info('== COMMAND LINE ==')
logger.info(' '.join(sys.argv))
logger.info('== PARAMETERS ==')
for k, v in p.items():
if new_params.get(k) is not None:
p[k] = new_params[k]
replace_str = "<- " + str(new_params.get(k))
else:
replace_str = ""
logger.info(" {!s:20s}: {!s:<20s} {}".format(k, v, replace_str))
return p, loaded
def decoder(self, clean, corr, batch_size):
get_unlabeled = lambda x: x[batch_size:] if x is not None else x
est = self.new_activation_dict()
costs = AttributeDict()
costs.denois = AttributeDict()
for i, ((_, spec), act_f) in self.layers[::-1]:
z_corr = get_unlabeled(corr.z[i])
z_clean = get_unlabeled(clean.z[i])
z_clean_s = get_unlabeled(clean.s.get(i))
z_clean_m = get_unlabeled(clean.m.get(i))
# It's the last layer
if i == len(self.layers) - 1:
fspec = (None, None)
ver = get_unlabeled(corr.h[i])
ver_dim = self.layer_dims[i]
top_g = True
else:
fspec = self.layers[i + 1][1][0]
ver = est.z.get(i + 1)
ver_dim = self.layer_dims.get(i + 1)
top_g = False
z_est = self.g(z_lat=z_corr,
z_ver=ver,
in_dims=ver_dim,
out_dims=self.layer_dims[i],
num=i,
fspec=fspec,
top_g=top_g)
# For semi-supervised version
if z_clean_s:
z_est_norm = (z_est - z_clean_m) / z_clean_s
else:
z_est_norm = z_est
z_est_norm = z_est
se = SquaredError('denois' + str(i))
costs.denois[i] = se.apply(z_est_norm.flatten(2),
z_clean.flatten(2)) \
/ np.prod(self.layer_dims[i], dtype=floatX)
costs.denois[i].name = 'denois' + str(i)
# Store references for later use
est.z[i] = z_est
est.h[i] = apply_act(z_est, act_f)
est.s[i] = None
est.m[i] = None
return est, costs
def decoder(self, clean, corr):
est = self.new_activation_dict()
costs = AttributeDict()
costs.denois = AttributeDict()
for i, ((_, spec), act_f) in self.layers[::-1]:
z_corr = corr.unlabeled.z[i]
z_clean = clean.unlabeled.z[i]
z_clean_s = clean.unlabeled.s.get(i)
z_clean_m = clean.unlabeled.m.get(i)
# It's the last layer
if i == len(self.layers) - 1:
fspec = (None, None)
ver = corr.unlabeled.h[i]
ver_dim = self.layer_dims[i]
top_g = True
else:
fspec = self.layers[i + 1][1][0]
ver = est.z.get(i + 1)
ver_dim = self.layer_dims.get(i + 1)
top_g = False
z_est = self.g(z_lat=z_corr,
z_ver=ver,
in_dims=ver_dim,
out_dims=self.layer_dims[i],
num=i,
fspec=fspec,
top_g=top_g)
# The first layer
if z_clean_s:
z_est_norm = (z_est - z_clean_m) / z_clean_s
else:
z_est_norm = z_est
se = SquaredError('denois' + str(i))
costs.denois[i] = se.apply(z_est_norm.flatten(2),
z_clean.flatten(2)) \
/ np.prod(self.layer_dims[i], dtype=floatX)
costs.denois[i].name = 'denois' + str(i)
# Store references for later use
est.z[i] = z_est
est.h[i] = apply_act(z_est, act_f)
est.s[i] = None
est.m[i] = None
return est, costs
def setup_data(p, test_set=False):
dataset_class, training_set_size = {
'cifar10': (CIFAR10, 40000),
'mnist': (MNIST, 50000),
'reddit': (SubredditTopPhotosFeatures22, 20000)
}[p.dataset]
# Allow overriding the default from command line
if p.get('unlabeled_samples') is not None:
training_set_size = p.unlabeled_samples
train_set = dataset_class(("train", ))
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
if p.get('dseed'):
rng = numpy.random.RandomState(seed=p.dseed)
rng.shuffle(all_ind)
d = AttributeDict()
# Choose the training set
d.train = train_set
d.train_ind = all_ind[:training_set_size]
# Then choose validation set from the remaining indices
d.valid = train_set
d.valid_ind = numpy.setdiff1d(all_ind, d.train_ind)[:p.valid_set_size]
logger.info('Using %d examples for validation' % len(d.valid_ind))
# Only touch test data if requested
if test_set:
d.test = dataset_class(("test", ))
d.test_ind = numpy.arange(d.test.num_examples)
# Setup optional whitening, only used for Cifar-10
in_dim = train_set.data_sources[train_set.sources.index(
'features')].shape[1:]
if len(in_dim) > 1 and p.whiten_zca > 0:
assert numpy.product(in_dim) == p.whiten_zca, \
'Need %d whitening dimensions, not %d' % (numpy.product(in_dim),
p.whiten_zca)
cnorm = ContrastNorm(p.contrast_norm) if p.contrast_norm != 0 else None
def get_data(d, i):
data = d.get_data(request=i)[d.sources.index('features')]
# Fuel provides Cifar in uint8, convert to float32
data = numpy.require(data, dtype=numpy.float32)
return data if cnorm is None else cnorm.apply(data)
if p.whiten_zca > 0:
logger.info('Whitening using %d ZCA components' % p.whiten_zca)
whiten = ZCA()
whiten.fit(p.whiten_zca, get_data(d.train, d.train_ind))
else:
whiten = None
return in_dim, d, whiten, cnorm
def mnist_load(train_size=50000, dseed=1):
# borrowed from https://github.com/Lasagne/Lasagne/blob/master/examples/mnist.py
# We can now download and read the training and test set images and labels.
X_train = load_mnist_images('train-images-idx3-ubyte.gz')
y_train = load_mnist_labels('train-labels-idx1-ubyte.gz')
X_test = load_mnist_images('t10k-images-idx3-ubyte.gz')
y_test = load_mnist_labels('t10k-labels-idx1-ubyte.gz')
# We reserve the last 10000 training examples for validation.
rng = np.random.RandomState(dseed)
randix = rng.permutation(X_train.shape[0])
X_train, X_val = X_train[randix[:train_size]], X_train[randix[train_size:]]
y_train, y_val = y_train[randix[:train_size]], y_train[randix[train_size:]]
logger.debug('%d examples in training dataset' % X_train.shape[0])
logger.debug('%d examples in validation dataset' % X_val.shape[0])
logger.debug('%d examples in testing dataset' % X_test.shape[0])
return AttributeDict({
'X_train': X_train,
'y_train': y_train,
'X_val': X_val,
'y_val': y_val,
'X_test': X_test,
'y_test': y_test
})
def decoder(self, clean, corr, batch_size):
get_unlabeled = lambda x: x[batch_size:] if x is not None else x
est = self.new_activation_dict()
costs = AttributeDict()
costs.denois = AttributeDict()
for i, ((_, spec), act_f) in self.layers[::-1]:
z_corr = get_unlabeled(corr.z[i])
z_clean = get_unlabeled(clean.z[i])
z_clean_s = get_unlabeled(clean.s.get(i))
z_clean_m = get_unlabeled(clean.m.get(i))
# It's the last layer
if i == len(self.layers) - 1:
fspec = (None, None)
ver = get_unlabeled(corr.h[i])
ver_dim = self.layer_dims[i]
top_g = True
else:
fspec = self.layers[i + 1][1][0]
ver = est.z.get(i + 1)
ver_dim = self.layer_dims.get(i + 1)
top_g = False
z_est = self.g(
z_lat=z_corr, z_ver=ver, in_dims=ver_dim, out_dims=self.layer_dims[i], num=i, fspec=fspec, top_g=top_g
)
# For semi-supervised version
if z_clean_s:
z_est_norm = (z_est - z_clean_m) / z_clean_s
else:
z_est_norm = z_est
z_est_norm = z_est
se = SquaredError("denois" + str(i))
costs.denois[i] = se.apply(z_est_norm.flatten(2), z_clean.flatten(2)) / np.prod(
self.layer_dims[i], dtype=floatX
)
costs.denois[i].name = "denois" + str(i)
# Store references for later use
est.z[i] = z_est
est.h[i] = apply_act(z_est, act_f)
est.s[i] = None
est.m[i] = None
return est, costs
def load_and_log_params(cli_params):
cli_params = AttributeDict(cli_params)
if cli_params.get('load_from'):
p = load_df(cli_params.load_from, 'params').to_dict()[0]
p = AttributeDict(p)
for key in cli_params.iterkeys():
if key not in p:
p[key] = None
new_params = cli_params
loaded = True
else:
p = cli_params
new_params = {}
loaded = False
# Make dseed seed unless specified explicitly
if p.get('dseed') is None and p.get('seed') is not None:
p['dseed'] = p['seed']
logger.info('== COMMAND LINE ==')
logger.info(' '.join(sys.argv))
logger.info('== PARAMETERS ==')
for k, v in p.iteritems():
if new_params.get(k) is not None:
p[k] = new_params[k]
replace_str = "<- " + str(new_params.get(k))
else:
replace_str = ""
logger.info(" {:20}: {:<20} {}".format(k, v, replace_str))
return p, loaded
def get_mnist_data_dict(unlabeled_samples, valid_set_size, test_set=False):
train_set = MNIST(("train",))
# Make sure the MNIST data is in right format
train_set.data_sources = (
(train_set.data_sources[0] / 255.).astype(numpy.float32),
train_set.data_sources[1])
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
rng = numpy.random.RandomState(seed=1)
rng.shuffle(all_ind)
data = AttributeDict()
# Choose the training set
data.train = train_set
data.train_ind = all_ind[:unlabeled_samples]
# Then choose validation set from the remaining indices
data.valid = train_set
data.valid_ind = numpy.setdiff1d(all_ind, data.train_ind)[:valid_set_size]
logger.info('Using %d examples for validation' % len(data.valid_ind))
# Only touch test data if requested
if test_set:
data.test = MNIST(("test",))
data.test_ind = numpy.arange(data.test.num_examples)
return data
def __init__(self, params):
super().__init__()
if not isinstance(params, AttributeDict):
params = AttributeDict(params)
self.hparams = params
self.model = TransformerModel(params)
def setup_data(p, test_set=False):
dataset_class, training_set_size = {"cifar10": (CIFAR10, 40000), "mnist": (MNIST, 50000)}[p.dataset]
# Allow overriding the default from command line
if p.get("unlabeled_samples") is not None:
training_set_size = p.unlabeled_samples
train_set = dataset_class("train")
# Make sure the MNIST data is in right format
if p.dataset == "mnist":
d = train_set.data_sources[train_set.sources.index("features")]
assert numpy.all(d <= 1.0) and numpy.all(d >= 0.0), "Make sure data is in float format and in range 0 to 1"
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
if p.get("dseed"):
rng = numpy.random.RandomState(seed=p.dseed)
rng.shuffle(all_ind)
d = AttributeDict()
# Choose the training set
d.train = train_set
d.train_ind = all_ind[:training_set_size]
# Then choose validation set from the remaining indices
d.valid = train_set
d.valid_ind = numpy.setdiff1d(all_ind, d.train_ind)[: p.valid_set_size]
logger.info("Using %d examples for validation" % len(d.valid_ind))
# Only touch test data if requested
if test_set:
d.test = dataset_class("test")
d.test_ind = numpy.arange(d.test.num_examples)
# Setup optional whitening, only used for Cifar-10
in_dim = train_set.data_sources[train_set.sources.index("features")].shape[1:]
if len(in_dim) > 1 and p.whiten_zca > 0:
assert numpy.product(in_dim) == p.whiten_zca, "Need %d whitening dimensions, not %d" % (
numpy.product(in_dim),
p.whiten_zca,
)
cnorm = ContrastNorm(p.contrast_norm) if p.contrast_norm != 0 else None
def get_data(d, i):
data = d.get_data(request=i)[d.sources.index("features")]
# Fuel provides Cifar in uint8, convert to float32
data = numpy.require(data, dtype=numpy.float32)
return data if cnorm is None else cnorm.apply(data)
if p.whiten_zca > 0:
logger.info("Whitening using %d ZCA components" % p.whiten_zca)
whiten = ZCA()
whiten.fit(p.whiten_zca, get_data(d.train, d.train_ind))
else:
whiten = None
return in_dim, d, whiten, cnorm
def metrics(indexes=['data_objects', 'data_bundles']):
"""
return document counts
"""
return [
AttributeDict({
'name': index,
'count': len(store.get(index, []))
}) for index in indexes
]
def setup_data(p, test_set=False):
dataset_class, training_set_size = {
'cifar10': (CIFAR10, 40000),
'mnist': (MNIST, 50000),
'reddit': (SubredditTopPhotosFeatures22, 20000)
}[p.dataset]
# Allow overriding the default from command line
if p.get('unlabeled_samples') is not None:
training_set_size = p.unlabeled_samples
train_set = dataset_class(("train",))
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
if p.get('dseed'):
rng = numpy.random.RandomState(seed=p.dseed)
rng.shuffle(all_ind)
d = AttributeDict()
# Choose the training set
d.train = train_set
d.train_ind = all_ind[:training_set_size]
# Then choose validation set from the remaining indices
d.valid = train_set
d.valid_ind = numpy.setdiff1d(all_ind, d.train_ind)[:p.valid_set_size]
logger.info('Using %d examples for validation' % len(d.valid_ind))
# Only touch test data if requested
if test_set:
d.test = dataset_class(("test",))
d.test_ind = numpy.arange(d.test.num_examples)
# Setup optional whitening, only used for Cifar-10
in_dim = train_set.data_sources[train_set.sources.index('features')].shape[1:]
if len(in_dim) > 1 and p.whiten_zca > 0:
assert numpy.product(in_dim) == p.whiten_zca, \
'Need %d whitening dimensions, not %d' % (numpy.product(in_dim),
p.whiten_zca)
cnorm = ContrastNorm(p.contrast_norm) if p.contrast_norm != 0 else None
def get_data(d, i):
data = d.get_data(request=i)[d.sources.index('features')]
# Fuel provides Cifar in uint8, convert to float32
data = numpy.require(data, dtype=numpy.float32)
return data if cnorm is None else cnorm.apply(data)
if p.whiten_zca > 0:
logger.info('Whitening using %d ZCA components' % p.whiten_zca)
whiten = ZCA()
whiten.fit(p.whiten_zca, get_data(d.train, d.train_ind))
else:
whiten = None
return in_dim, d, whiten, cnorm
def setup_data(p, test_set=False):
dataset_class, training_set_size = {
'cifar10': (CIFAR10, 40000),
'mnist': (MNIST, 50000),
}[p.dataset]
# Allow overriding the default from command line
if p.get('unlabeled_samples') is not None:
training_set_size = p.unlabeled_samples
train_set = dataset_class("train")
# Make sure the MNIST data is in right format
if p.dataset == 'mnist':
d = train_set.data_sources[train_set.sources.index('features')]
assert numpy.all(d <= 1.0) and numpy.all(d >= 0.0), \
'Make sure data is in float format and in range 0 to 1'
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
if p.get('dseed'):
rng = numpy.random.RandomState(seed=p.dseed)
rng.shuffle(all_ind)
d = AttributeDict()
# Choose the training set
d.train = train_set
d.train_ind = all_ind[:training_set_size]
# Then choose validation set from the remaining indices
d.valid = train_set
d.valid_ind = numpy.setdiff1d(all_ind, d.train_ind)[:p.valid_set_size]
logger.info('Using %d examples for validation' % len(d.valid_ind))
# Only touch test data if requested
if test_set:
d.test = dataset_class("test")
d.test_ind = numpy.arange(d.test.num_examples)
in_dim = train_set.data_sources[train_set.sources.index(
'features')].shape[1:]
def get_data(d, i):
data = d.get_data(request=i)[d.sources.index('features')]
# Fuel provides Cifar in uint8, convert to float32
data = numpy.require(data, dtype=numpy.float32)
return data if cnorm is None else cnorm.apply(data)
return in_dim, d
def doPreprocessing(self):
results = AttributeDict()
results.dataset = []
for i in range(len(self.params.dataset)):
# shall we just load it?
filename = '%s/preprocessing-%s%s.mat' % (
self.params.dataset[i].savePath,
self.params.dataset[i].saveFile, self.params.saveSuffix)
if self.params.dataset[i].preprocessing.load and os.path.isfile(
filename):
r = loadmat(filename)
print('Loading file %s ...' % filename)
results.dataset[i].preprocessing = r.results_preprocessing
else:
# or shall we actually calculate it?
p = deepcopy(self.params)
p.dataset = self.params.dataset[i]
d = AttributeDict()
d.preprocessing = np.copy(SeqSLAM.preprocessing(p))
results.dataset.append(d)
if self.params.dataset[i].preprocessing.save:
results_preprocessing = results.dataset[i].preprocessing
savemat(filename,
{'results_preprocessing': results_preprocessing})
return results
def _load_extends_settings(self, section_name, store):
"""
Loads all settings from other template(s) specified by a section's
'extends' setting.
This method walks a dependency tree of sections from bottom up. Each
step is a group of settings for a section in the form of a dictionary.
A 'master' dictionary is updated with the settings at each step. This
causes the next group of settings to override the previous, and so on.
The 'section_name' settings are at the top of the dependency tree.
"""
section = store[section_name]
extends = section.get('extends')
if extends is None:
return
if DEBUG_CONFIG:
log.debug('%s extends %s' % (section_name, extends))
extensions = [section]
while True:
extends = section.get('extends', None)
if not extends:
break
try:
section = store[extends]
if section in extensions:
exts = ', '.join([self._get_section_name(x['__name__'])
for x in extensions])
raise exception.ConfigError(
"Cyclical dependency between sections %s. "
"Check your EXTENDS settings." % exts)
extensions.insert(0, section)
except KeyError:
raise exception.ConfigError(
"%s can't extend non-existent section %s" %
(section_name, extends))
transform = AttributeDict()
for extension in extensions:
transform.update(extension)
store[section_name] = transform
def __init__(self, p):
logger.debug(theano.config)
self.p = p
self.params = OrderedDict()
self.rstream = RandomStreams(seed=p.seed)
self.rng = np.random.RandomState(seed=p.seed)
self.in_dim = data_dim(p) # input dimensionality
input_type = T.type.TensorType('float32',
[False] * (len(self.in_dim) + 1))
input_plus_one_type = T.type.TensorType('float32', [False] *
(len(self.in_dim) + 2))
self.x_only = input_type('features_unlabeled')
self.x = input_type('features_labeled')
self.y = theano.tensor.lmatrix('targets_labeled')
self.masks_unlabeled = input_plus_one_type('masks_unlabeled')
# We noticed that continuous case becomes more stable if we
# have stable v, i.e. put Ladder's v through a sigmoid
if p.input_type == 'continuous':
decoder_spec = ('gauss_stable_v', )
else:
decoder_spec = ('gauss', )
# Ladder Network, a.k.a Parametric Mapping
ladder_p = AttributeDict({
'seed':
p.seed,
'encoder_layers':
p.encoder_proj,
'decoder_spec':
decoder_spec,
'denoising_cost_x': (0.0, ) * len(p.encoder_proj),
'act':
'relu',
# Ladder doesn't add noise to its layers. Tagger handles all corruption.
'f_local_noise_std':
0.,
'super_noise_std':
0.,
'zestbn':
'no',
'top_c':
True,
'lr':
0.,
})
self.ladder = LadderAE(ladder_p)
# disable logging from ladder
logging.getLogger('main.model').setLevel(logging.WARNING)
def merge_with_yaml(yaml_filename):
"""Load a yaml config file and merge it into the global config object"""
global _g_conf
with open(yaml_filename, 'r') as f:
yaml_file = yaml.load(f)
yaml_cfg = AttributeDict(yaml_file)
_merge_a_into_b(yaml_cfg, _g_conf)
path_parts = os.path.split(yaml_filename)
_g_conf.EXPERIMENT_BATCH_NAME = os.path.split(path_parts[-2])[-1]
_g_conf.EXPERIMENT_NAME = path_parts[-1].split('.')[-2]
def doPreprocessing(self):
results = AttributeDict()
results.dataset = []
for i in range(len(self.params.dataset)):
# shall we just load it?
filename = '%s/preprocessing-%s%s.mat' % (self.params.dataset[i].savePath, self.params.dataset[i].saveFile, self.params.saveSuffix)
if self.params.dataset[i].preprocessing.load and os.path.isfile(filename):
r = loadmat(filename)
print('Loading file %s ...' % filename)
results.dataset[i].preprocessing = r.results_preprocessing
else:
# or shall we actually calculate it?
p = deepcopy(self.params)
p.dataset = self.params.dataset[i]
d = AttributeDict()
d.preprocessing = np.copy(SeqSLAM.preprocessing(p))
results.dataset.append(d)
if self.params.dataset[i].preprocessing.save:
results_preprocessing = results.dataset[i].preprocessing
savemat(filename, {'results_preprocessing': results_preprocessing})
return results
def __init__(self, params):
super().__init__()
if not isinstance(params, AttributeDict):
params = AttributeDict(params)
self.hparams = params
self.model = VQVCModel(params)
self.spec_augmenter = SpecAugmentation(time_drop_width=3,
time_stripes_num=2,
freq_drop_width=3,
freq_stripes_num=2)
def encoder(self, input_, path_name, input_noise_std, noise_std):
h = input_
h = h + (self.rstream.normal(size=h.shape).astype(floatX) *
input_noise_std)
d = AttributeDict()
d.unlabeled = self.new_activation_dict()
d.labeled = self.new_activation_dict()
d.labeled.z[0], d.unlabeled.z[0] = self.split_lu(h)
prev_dim = self.input_dim
for i, (spec, act_f) in self.layers[1:]:
d.labeled.h[i - 1], d.unlabeled.h[i - 1] = self.split_lu(h)
noise = noise_std[i] if i < len(noise_std) else 0.
curr_dim, z, m, s, h = self.f(h, prev_dim, spec, i, act_f,
path_name=path_name,
noise_std=noise)
self.layer_dims[i] = curr_dim
d.labeled.z[i], d.unlabeled.z[i] = self.split_lu(z)
d.unlabeled.s[i] = s
d.unlabeled.m[i] = m
prev_dim = curr_dim
d.labeled.h[i], d.unlabeled.h[i] = self.split_lu(h)
return d
def get_dataset(root: str = DATASETS_PATH,
train_size: float = .8,
shuffle=True,
seed=None) -> AttributeDict:
dtype_dict = {
x: np.float_ if x in CATEGORICAL_FEATURES else np.float_
for x in ALL_FEATURES
}
dataset = AttributeDict(
train=AttributeDict(X=pd.read_csv(os.path.join(root, 'X1.csv'),
dtype=dtype_dict),
y=pd.read_csv(os.path.join(root, 'Y1.csv'),
header=None,
dtype={"shares": np.float_},
names=["shares"])),
validation=AttributeDict(X=None, y=None),
test=AttributeDict(X=pd.read_csv(os.path.join(root, 'X2.csv'),
dtype=dtype_dict),
y=None))
if train_size < 1:
dataset.train.X, dataset.validation.X, dataset.train.y, dataset.validation.y = train_test_split(
dataset.train.X,
dataset.train.y,
train_size=train_size,
shuffle=shuffle,
random_state=seed)
else:
dataset.validation.X = pd.DataFrame(
{c: []
for c in dataset.train.X.columns.values})
dataset.validation.y = pd.DataFrame(
{c: []
for c in dataset.train.y.columns.values})
return dataset
def setup_data(p, test_set=False):
dataset_class, training_set_size = {
'cifar10': (CIFAR10, 40000),
'mnist': (MNIST, 50000),
}[p.dataset]
# Allow overriding the default from command line
if p.get('unlabeled_samples') is not None:
training_set_size = p.unlabeled_samples
train_set = dataset_class("train")
# Make sure the MNIST data is in right format
if p.dataset == 'mnist':
d = train_set.data_sources[train_set.sources.index('features')]
assert numpy.all(d <= 1.0) and numpy.all(d >= 0.0), \
'Make sure data is in float format and in range 0 to 1'
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
if p.get('dseed'):
rng = numpy.random.RandomState(seed=p.dseed)
rng.shuffle(all_ind)
d = AttributeDict()
# Choose the training set
d.train = train_set
d.train_ind = all_ind[:training_set_size]
# Then choose validation set from the remaining indices
d.valid = train_set
d.valid_ind = numpy.setdiff1d(all_ind, d.train_ind)[:p.valid_set_size]
logger.info('Using %d examples for validation' % len(d.valid_ind))
# Only touch test data if requested
if test_set:
d.test = dataset_class("test")
d.test_ind = numpy.arange(d.test.num_examples)
in_dim = train_set.data_sources[train_set.sources.index('features')].shape[1:]
def get_data(d, i):
data = d.get_data(request=i)[d.sources.index('features')]
# Fuel provides Cifar in uint8, convert to float32
data = numpy.require(data, dtype=numpy.float32)
return data if cnorm is None else cnorm.apply(data)
return in_dim, d
def metrics(indexes=['data_objects', 'data_bundles']):
"""
return document counts
"""
def _count(index):
try:
s = Search(using=client, index=index, doc_type=index[:-1])
return s.count()
except elasticsearch.exceptions.NotFoundError as no_found:
log.info('NotFoundError {} (expected on empty db)'.format(index))
return 0
except Exception as e:
log.error('error getting count of documents in {}'.format(index))
log.exception(e)
raise e
return [
AttributeDict(
{'name': index, 'count': _count(index)}
) for index in indexes
]
def merge_with_yaml(yaml_filename):
"""Load a yaml config file and merge it into the global config object"""
global _g_conf
with open(yaml_filename, 'r') as f:
yaml_file = yaml.load(f)
yaml_cfg = AttributeDict(yaml_file)
print("yaml here", yaml_cfg)
print("batch size ", yaml_cfg.BATCH_SIZE)
_merge_a_into_b(yaml_cfg, _g_conf)
#TODO: Merging is missing
path_parts = os.path.split(yaml_filename)
_g_conf.EXPERIMENT_BATCH_NAME = os.path.split(path_parts[-2])[-1]
_g_conf.EXPERIMENT_NAME = path_parts[-1].split('.')[-2]
def load_and_log_params(cli_params):
cli_params = AttributeDict(cli_params)
# 如果有load_from参数，说明参数是有文件里面读取的，文件格式为hdf
if cli_params.get('load_from'):
# load_from值 + params组成完整地址
# string => dict
p = load_df(cli_params.load_from, 'params').to_dict()[0]
# dict => AttributeDict
p = AttributeDict(p)
for key in cli_params.iterkeys():
if key not in p:
p[key] = None
new_params = cli_params
loaded = True
# 如果没有load from参数，直接封装一下cli_params
else:
p = cli_params
new_params = {}
loaded = False
# Make dseed seed unless specified explicitly
# dseed为空而seed不为空时，dseed复制为seed
if p.get('dseed') is None and p.get('seed') is not None:
p['dseed'] = p['seed']
# log相关
logger.info('== COMMAND LINE ==')
logger.info(' '.join(sys.argv))
logger.info('== PARAMETERS ==')
for k, v in p.iteritems():
if new_params.get(k) is not None:
p[k] = new_params[k]
replace_str = "<- " + str(new_params.get(k))
else:
replace_str = ""
logger.info(" {:20}: {:<20} {}".format(k, v, replace_str))
return p, loaded
def update(_id, doc, index='data_objects'):
"""
partial update using the contructed _id
"""
def merge(a, b, path=None):
"merges b into a"
if path is None:
path = []
for key in b:
if key in a:
if isinstance(a[key], dict) and isinstance(b[key], dict):
merge(a[key], b[key], path + [str(key)])
elif a[key] == b[key]:
pass # same leaf value
else:
a[key] = b[key] # update with new
else:
a[key] = b[key]
return a
store[index].append(
AttributeDict(merge(_get(_id, index=index, current=True),
doc))) # noqa
def manage_review_view(request, review_id=None):
if not request.user.id:
return HttpResponseRedirect('/login')
if request.method in CREATE_OR_UPDATE_METHODS:
data = request.POST
if data.get('action') == 'delete':
Review.objects.filter(user=request.user, id=review_id).delete()
next_url = '/'
elif data.get('action') == 'autofill':
fill_missing_review_data(review_id)
next_url = request.path
else:
review, created = Review.create_or_update(request.user,
data,
id=review_id)
if created:
fill_missing_review_data(review.id)
next_url = '/?submittedReview={}'.format(review.id)
return HttpResponseRedirect(next_url)
else:
if review_id:
review = Review.objects.get(user=request.user, id=review_id)
else:
# pre-populate fields from url parameters.
review = AttributeDict({
k: v
for k, v in request.GET.items()
if k in {'url', 'rating', 'text'}
})
return render_with_globals(request, 'manage_review.html', {
'review': review,
'review_id': review_id
})
def save(doc, index='data_objects'):
"""
save the body in the index, ensure version and id set
"""
doc = AttributeDict(doc)
version = doc.get('version', None)
if not version:
doc['version'] = now()
if not doc.get('id', None):
temp_id = str(uuid.uuid4())
doc['id'] = temp_id
if _is_duplicate(doc, index):
raise Exception("duplicate document")
if index not in store:
store[index] = []
doc.meta = AttributeDict({'id': doc.id})
store[index].append(doc)
log.info(doc.id)
return doc
def setup_data(p, use_unlabeled=True, use_labeled=True):
assert use_unlabeled or use_labeled, 'Cannot train without cost'
dataset_class = DATASETS[p.dataset]
dataset = dataset_class(p)
train_ind = dataset.trn.ind
if 'labeled_samples' not in p or p.labeled_samples == 0:
n_labeled = len(train_ind)
else:
n_labeled = p.labeled_samples
if 'unlabeled_samples' not in p:
n_unlabeled = len(train_ind)
else:
n_unlabeled = p.unlabeled_samples
assert p.batch_size <= n_labeled, "batch size too large"
assert len(train_ind) >= n_labeled
assert len(train_ind) >= n_unlabeled, "not enough training samples"
assert n_labeled <= n_unlabeled, \
"at least as many unlabeled samples as number of labeled samples"
# If not using all labels, let's balance classes
balance_classes = n_labeled < len(train_ind)
if balance_classes and use_labeled:
# Ensure each label is equally represented
y = dataset.get_train_labels()
n_classes = numpy.max(y) + 1
n_from_each_class = n_labeled / n_classes
logger.info('n_sample_from_each_class {0}'.format(n_from_each_class))
assert n_labeled % n_classes == 0
i_labeled = []
for c in xrange(n_classes):
i = (train_ind[y[:, 0] == c])[:n_from_each_class]
if len(i) < n_from_each_class:
logger.warning('Class {0} : only got {1}'.format(c, len(i)))
i_labeled += list(i)
else:
i_labeled = train_ind[:n_labeled]
def make_unlabeled_set(train_ind, i_labeled, n_unlabeled):
""" i_unused_labeled: the labels that are not used in i_labeled.
n_unlabeled_needed: the number of need for i_unlabeled beyond len(i_labeled)
"""
i_unused_labeled = list(set(train_ind) - set(i_labeled))
n_unlabeled_needed = n_unlabeled - len(i_labeled)
i_unlabeled = i_unused_labeled[:n_unlabeled_needed]
i_unlabeled.extend(i_labeled)
return i_unlabeled
i_unlabeled = make_unlabeled_set(train_ind, i_labeled, n_unlabeled)
logger.info('Creating data set with %d labeled and %d total samples' %
(len(i_labeled), len(i_unlabeled)))
streams = AttributeDict()
def make(kind, ind_labeled, ind_unlabeled):
ds_labeled, ds_unlabeled = None, None
if use_labeled:
ds_labeled = dataset.get_datastream(kind, ind_labeled)
if use_unlabeled:
ds_unlabeled = dataset.get_datastream(kind, ind_unlabeled)
return combine_datastreams(ds_labeled, ds_unlabeled)
streams.train = make('trn', i_labeled, i_unlabeled)
streams.valid = make('val', None, None) # use all indices
streams.test = make('tst', None, None) # use all indices
return streams
def apply(self, input_labeled, target_labeled, input_unlabeled):
self.layer_counter = 0
input_dim = self.p.encoder_layers[0]
# Store the dimension tuples in the same order as layers.
layers = self.layers
self.layer_dims = {0: input_dim}
self.lr = self.shared(self.default_lr, 'learning_rate', role=None)
self.costs = costs = AttributeDict()
self.costs.denois = AttributeDict()
self.act = AttributeDict()
self.error = AttributeDict()
self.oos = AttributeDict()
top = len(layers) - 1
N = input_labeled.shape[0]
self.join = lambda l, u: T.concatenate([l, u], axis=0)
self.labeled = lambda x: x[:N] if x is not None else x
self.unlabeled = lambda x: x[N:] if x is not None else x
self.split_lu = lambda x: (self.labeled(x), self.unlabeled(x))
input_concat = self.join(input_labeled, input_unlabeled)
def encoder(input_, path_name, input_noise_std=0, noise_std=[]):
h = input_
logger.info(' 0: noise %g' % input_noise_std)
if input_noise_std > 0.:
h = h + self.noise_like(h) * input_noise_std
d = AttributeDict()
d.unlabeled = self.new_activation_dict()
d.labeled = self.new_activation_dict()
d.labeled.z[0] = self.labeled(h)
d.unlabeled.z[0] = self.unlabeled(h)
prev_dim = input_dim
for i, (spec, _, act_f) in layers[1:]:
d.labeled.h[i - 1], d.unlabeled.h[i - 1] = self.split_lu(h)
noise = noise_std[i] if i < len(noise_std) else 0.
curr_dim, z, m, s, h = self.f(h,
prev_dim,
spec,
i,
act_f,
path_name=path_name,
noise_std=noise)
assert self.layer_dims.get(i) in (None, curr_dim)
self.layer_dims[i] = curr_dim
d.labeled.z[i], d.unlabeled.z[i] = self.split_lu(z)
d.unlabeled.s[i] = s
d.unlabeled.m[i] = m
prev_dim = curr_dim
d.labeled.h[i], d.unlabeled.h[i] = self.split_lu(h)
return d
# Clean, supervised
logger.info('Encoder: clean, labeled')
clean = self.act.clean = encoder(input_concat, 'clean')
# Corrupted, supervised
logger.info('Encoder: corr, labeled')
corr = self.act.corr = encoder(input_concat,
'corr',
input_noise_std=self.p.super_noise_std,
noise_std=self.p.f_local_noise_std)
est = self.act.est = self.new_activation_dict()
# Decoder path in opposite order
logger.info('Decoder: z_corr -> z_est')
for i, ((_, spec), l_type, act_f) in layers[::-1]:
z_corr = corr.unlabeled.z[i]
z_clean = clean.unlabeled.z[i]
z_clean_s = clean.unlabeled.s.get(i)
z_clean_m = clean.unlabeled.m.get(i)
fspec = layers[i + 1][1][0] if len(layers) > i + 1 else (None,
None)
if i == top:
ver = corr.unlabeled.h[i]
ver_dim = self.layer_dims[i]
top_g = True
else:
ver = est.z.get(i + 1)
ver_dim = self.layer_dims.get(i + 1)
top_g = False
z_est = self.g(z_lat=z_corr,
z_ver=ver,
in_dims=ver_dim,
out_dims=self.layer_dims[i],
l_type=l_type,
num=i,
fspec=fspec,
top_g=top_g)
if z_est is not None:
# Denoising cost
if z_clean_s and self.p.zestbn == 'bugfix':
z_est_norm = (z_est - z_clean_m
) / T.sqrt(z_clean_s + np.float32(1e-10))
elif z_clean_s is None or self.p.zestbn == 'no':
z_est_norm = z_est
else:
assert False, 'Not supported path'
se = SquaredError('denois' + str(i))
costs.denois[i] = se.apply(z_est_norm.flatten(2),
z_clean.flatten(2)) \
/ np.prod(self.layer_dims[i], dtype=floatX)
costs.denois[i].name = 'denois' + str(i)
denois_print = 'denois %.2f' % self.p.denoising_cost_x[i]
else:
denois_print = ''
# Store references for later use
est.h[i] = self.apply_act(z_est, act_f)
est.z[i] = z_est
est.s[i] = None
est.m[i] = None
logger.info(' g%d: %10s, %s, dim %s -> %s' %
(i, l_type, denois_print, self.layer_dims.get(i + 1),
self.layer_dims.get(i)))
# Costs
y = target_labeled.flatten()
Q = int(self.layer_dims[top][0]) - 1
logger.info('Q=%d' % Q)
costs.class_clean = CategoricalCrossEntropyIV(
Q=Q,
alpha=self.p.alpha,
beta=self.p.beta,
dbeta=self.p.dbeta,
gamma=self.p.gamma,
gamma1=self.p.gamma1).apply(y, clean.labeled.h[top])
costs.class_clean.name = 'cost_class_clean'
costs.class_corr = CategoricalCrossEntropyIV(
Q=Q,
alpha=self.p.alpha,
beta=self.p.beta,
dbeta=self.p.dbeta,
gamma=self.p.gamma,
gamma1=self.p.gamma1,
).apply(y, corr.labeled.h[top])
costs.class_corr.name = 'cost_class_corr'
# This will be used for training
costs.total = costs.class_corr * 1.0
for i in range(top + 1):
if costs.denois.get(i) and self.p.denoising_cost_x[i] > 0:
costs.total += costs.denois[i] * self.p.denoising_cost_x[i]
if self.p.alpha_clean:
y_true = y
eps = np.float32(1e-6)
# scale preds so that the class probas of each sample sum to 1
y_pred = clean.labeled.h[top] + eps
y_pred /= y_pred.sum(axis=-1, keepdims=True)
y0 = T.or_(T.eq(y_true, 0), T.gt(y_true,
Q)) # out-of-set or unlabeled
y0sum = y0.sum() + eps # number of oos
cost1 = T.nnet.categorical_crossentropy(y_pred, y_pred)
cost1 = T.dot(y0,
cost1) / y0sum # average cost per labeled example
costs.total += self.p.alpha_clean * cost1
costs.total.name = 'cost_total'
# Classification error
mr = MisclassificationRateIV(oos_thr=self.p.oos_thr)
self.error.clean = mr.apply(y, clean.labeled.h[top]) * np.float32(100.)
self.error.clean.name = 'error_rate_clean'
oosr = OOSRateIV()
self.oos.clean = oosr.apply(y, clean.labeled.h[top]) * np.float32(100.)
self.oos.clean.name = 'oos_rate_clean'
# -*- encoding: utf-8 -*-
# customize/override for your backend
import logging
import os
import uuid
from utils import AttributeDict, now, add_created_timestamps, \
add_updated_timestamps
log = logging.getLogger(__name__)
DEFAULT_PAGE_SIZE = 100
store = AttributeDict({})
def save(doc, index='data_objects'):
"""
save the body in the index, ensure version and id set
"""
doc = AttributeDict(doc)
version = doc.get('version', None)
if not version:
doc['version'] = now()
if not doc.get('id', None):
temp_id = str(uuid.uuid4())
doc['id'] = temp_id
if _is_duplicate(doc, index):
def defaultParameters():
params = AttributeDict()
# switches
params.DO_PREPROCESSING = 1
params.DO_RESIZE = 0
params.DO_GRAYLEVEL = 1
params.DO_PATCHNORMALIZATION = 1 #!!!! 1
params.DO_SAVE_PREPROCESSED_IMG = 0
params.DO_DIFF_MATRIX = 1
params.DO_CONTRAST_ENHANCEMENT = 1
params.DO_FIND_MATCHES = 1
# parameters for preprocessing
params.downsample = AttributeDict()
params.downsample.size = [32, 64] # height, width
try:
params.downsample.method = Image.LANCZOS
except:
params.downsample.method = Image.ANTIALIAS
params.normalization = AttributeDict()
params.normalization.sideLength = 8
params.normalization.mode = 1
# parameters regarding the matching between images
params.matching = AttributeDict()
params.matching.ds = 10
params.matching.Rrecent=5
params.matching.vmin = 0.8
params.matching.vskip = 0.1
params.matching.vmax = 1.2
params.matching.Rwindow = 10
params.matching.save = 1
params.matching.load = 0 #1
# parameters for contrast enhancement on difference matrix
params.contrastEnhancement = AttributeDict()
params.contrastEnhancement.R = 10
# load old results or re-calculate? save results?
params.differenceMatrix = AttributeDict()
params.differenceMatrix.save = 1
params.differenceMatrix.load = 0 #1
params.contrastEnhanced = AttributeDict()
params.contrastEnhanced.save = 1
params.contrastEnhanced.load = 0 #1
# suffix appended on files containing the results
params.saveSuffix=''
return params
def new_activation_dict(self):
return AttributeDict({'z': {}, 'h': {}, 's': {}, 'm': {}})
def demo():
# set the parameters
# start with default parameters
params = defaultParameters()
# Nordland spring dataset
ds = AttributeDict()
ds.name = 'spring'
path = os.environ['DATASET_1_PATH']
ds.imagePath = path
ds.prefix = 'images-'
ds.extension = '.png'
ds.suffix = ''
ds.imageSkip = 1 # use every n-nth image
ds.imageIndices = range(1, 26, ds.imageSkip)
ds.savePath = 'results'
ds.saveFile = '%s-%d-%d-%d' % (ds.name, ds.imageIndices[0], ds.imageSkip,
ds.imageIndices[-1])
ds.preprocessing = AttributeDict()
ds.preprocessing.save = 1
ds.preprocessing.load = 0 #1
ds.crop = []
spring = ds
ds2 = deepcopy(ds)
# Nordland winter dataset
ds2.name = 'winter'
#ds.imagePath = '../datasets/nordland/64x32-grayscale-1fps/winter'
path = os.environ['DATASET_2_PATH']
ds2.saveFile = '%s-%d-%d-%d' % (ds2.name, ds2.imageIndices[0],
ds2.imageSkip, ds2.imageIndices[-1])
ds2.crop = []
winter = ds2
params.dataset = [spring, winter]
# load old results or re-calculate?
params.differenceMatrix.load = 0
params.contrastEnhanced.load = 0
params.matching.load = 0
# where to save / load the results
params.savePath = 'results'
## now process the dataset
ss = SeqSLAM(params)
t1 = time.time()
results = ss.run()
t2 = time.time()
print "time taken: " + str(t2 - t1)
## show some results
if len(results.matches) > 0:
m = results.matches[:,
0] # The LARGER the score, the WEAKER the match.
thresh = 0.90 # you can calculate a precision-recall plot by varying this threshold
m[results.matches[:,
1] > thresh] = np.nan # remove the weakest matches
plt.plot(m, '.') # ideally, this would only be the diagonal
plt.title('Matchings')
plt.show()
else:
print "Zero matches"
from __future__ import unicode_literals
from ast import literal_eval
from utils import AttributeDict
import copy
import numpy as np
import os
import os.path as osp
import yaml
from configs.namer import generate_name
from logger.coil_logger import create_log, add_message
# TODO: NAMing conventions ?
_g_conf = AttributeDict()
"""#### GENERAL CONFIGURATION PARAMETERS ####"""
_g_conf.NUMBER_OF_LOADING_WORKERS = 12
_g_conf.SENSORS = {'rgb': (3, 128, 128)}
_g_conf.MEASUREMENTS = {'targets': (31)}
_g_conf.TARGETS = ['steer', 'throttle', 'brake']
_g_conf.INPUTS = ['speed_module']
_g_conf.BALANCE_DATA = True
_g_conf.STEERING_DIVISION = [0.05, 0.05, 0.1, 0.3, 0.3, 0.1, 0.05, 0.05]
#_g_conf.STEERING_DIVISION = [0.01, 0.02, 0.07, 0.4, 0.4, 0.07, 0.02, 0.01] # Forcing curves alot
_g_conf.LABELS_DIVISION = [[0, 2, 5], [3], [4]]
_g_conf.BATCH_SIZE = 120
#_g_conf.AUGMENTATION_SUITE = [iag.ToGPU()]#, iag.Add((0, 0)), iag.Dropout(0, 0), iag.Multiply((1, 1.04)),
# #iag.GaussianBlur(sigma=(0.0, 3.0)),
# iag.ContrastNormalization((0.5, 1.5))
def setup_data(p, use_unlabeled=True, use_labeled=True):
assert use_unlabeled or use_labeled, 'Cannot train without cost'
dataset_class = DATASETS[p.dataset]
dataset = dataset_class(p)
train_ind = dataset.trn.ind
if 'labeled_samples' not in p or p.labeled_samples == 0:
n_labeled = len(train_ind)
else:
n_labeled = p.labeled_samples
if 'unlabeled_samples' not in p:
n_unlabeled = len(train_ind)
else:
n_unlabeled = p.unlabeled_samples
assert p.batch_size <= n_labeled, "batch size too large"
assert len(train_ind) >= n_labeled
assert len(train_ind) >= n_unlabeled, "not enough training samples"
assert n_labeled <= n_unlabeled, \
"at least as many unlabeled samples as number of labeled samples"
# If not using all labels, let's balance classes
balance_classes = n_labeled < len(train_ind)
if balance_classes and use_labeled:
# Ensure each label is equally represented
y = dataset.get_train_labels()
n_classes = numpy.max(y) + 1
n_from_each_class = n_labeled / n_classes
logger.info('n_sample_from_each_class {0}'.format(n_from_each_class))
assert n_labeled % n_classes == 0
i_labeled = []
for c in xrange(n_classes):
i = (train_ind[y[:, 0] == c])[:n_from_each_class]
if len(i) < n_from_each_class:
logger.warning('Class {0} : only got {1}'.format(c, len(i)))
i_labeled += list(i)
else:
i_labeled = train_ind[:n_labeled]
def make_unlabeled_set(train_ind, i_labeled, n_unlabeled):
""" i_unused_labeled: the labels that are not used in i_labeled.
n_unlabeled_needed: the number of need for i_unlabeled beyond len(i_labeled)
"""
i_unused_labeled = list(set(train_ind) - set(i_labeled))
n_unlabeled_needed = n_unlabeled - len(i_labeled)
i_unlabeled = i_unused_labeled[:n_unlabeled_needed]
i_unlabeled.extend(i_labeled)
return i_unlabeled
i_unlabeled = make_unlabeled_set(train_ind, i_labeled, n_unlabeled)
logger.info('Creating data set with %d labeled and %d total samples' %
(len(i_labeled), len(i_unlabeled)))
streams = AttributeDict()
def make(kind, ind_labeled, ind_unlabeled):
ds_labeled, ds_unlabeled = None, None
if use_labeled:
ds_labeled = dataset.get_datastream(kind, ind_labeled)
if use_unlabeled:
ds_unlabeled = dataset.get_datastream(kind, ind_unlabeled)
return combine_datastreams(ds_labeled, ds_unlabeled)
streams.train = make('trn', i_labeled, i_unlabeled)
streams.valid = make('val', None, None) # use all indices
streams.test = make('tst', None, None) # use all indices
return streams
def demo():
# set the parameters
# start with default parameters
params = defaultParameters()
# Nordland spring dataset
ds = AttributeDict()
ds.name = 'spring'
try:
path = os.environ['DATASET_1_PATH']
except:
path = '../datasets/nordland/64x32-grayscale-1fps/spring'
print "Warning: Environment variable DATASET_1_PATH not found! Trying '"+path+"'"
ds.imagePath = path
ds.prefix='images-'
ds.extension='.png'
ds.suffix=''
ds.imageSkip = 100 # use every n-nth image
ds.imageIndices = range(1, 35700, ds.imageSkip)
ds.savePath = 'results'
ds.saveFile = '%s-%d-%d-%d' % (ds.name, ds.imageIndices[0], ds.imageSkip, ds.imageIndices[-1])
ds.preprocessing = AttributeDict()
ds.preprocessing.save = 1
ds.preprocessing.load = 0 #1
#ds.crop=[1 1 60 32] # x0 y0 x1 y1 cropping will be done AFTER resizing!
ds.crop=[]
spring=ds
ds2 = deepcopy(ds)
# Nordland winter dataset
ds2.name = 'winter'
#ds.imagePath = '../datasets/nordland/64x32-grayscale-1fps/winter'
try:
path = os.environ['DATASET_2_PATH']
except:
path = '../datasets/nordland/64x32-grayscale-1fps/winter'
print "Warning: Environment variable DATASET_2_PATH not found! Trying '"+path+"'"
ds2.saveFile = '%s-%d-%d-%d' % (ds2.name, ds2.imageIndices[0], ds2.imageSkip, ds2.imageIndices[-1])
# ds.crop=[5 1 64 32]
ds2.crop=[]
winter=ds2
params.dataset = [spring, winter]
# load old results or re-calculate?
params.differenceMatrix.load = 0
params.contrastEnhanced.load = 0
params.matching.load = 0
# where to save / load the results
params.savePath='results'
## now process the dataset
ss = SeqSLAM(params)
t1=time.time()
results = ss.run()
t2=time.time()
print "time taken: "+str(t2-t1)
## show some results
if len(results.matches) > 0:
m = results.matches[:,0] # The LARGER the score, the WEAKER the match.
thresh=0.9 # you can calculate a precision-recall plot by varying this threshold
m[results.matches[:,1]>thresh] = np.nan # remove the weakest matches
plt.plot(m,'.') # ideally, this would only be the diagonal
plt.title('Matchings')
plt.show()
else:
print "Zero matches"
def defaultParameters():
params = AttributeDict()
# switches
params.DO_PREPROCESSING = 1
params.DO_RESIZE = 1
params.DO_GRAYLEVEL = 1
params.DO_PATCHNORMALIZATION = 0 #!!!! 1
params.DO_SAVE_PREPROCESSED_IMG = 1
params.DO_DIFF_MATRIX = 1
params.DO_CONTRAST_ENHANCEMENT = 1
params.DO_FIND_MATCHES = 1
# parameters for preprocessing
params.downsample = AttributeDict()
params.downsample.size = [48, 64] # height, width
try:
params.downsample.method = Image.LANCZOS
except:
params.downsample.method = Image.ANTIALIAS
params.normalization = AttributeDict()
params.normalization.sideLength = 8
params.normalization.mode = 1
# parameters regarding the matching between images
params.matching = AttributeDict()
params.matching.ds = 10
params.matching.Rrecent = 5
params.matching.vmin = 0.01
params.matching.vskip = 0.1
params.matching.vmax = 1.2
params.matching.Rwindow = 8
params.matching.save = 1
params.matching.load = 0 #1
# parameters for contrast enhancement on difference matrix
params.contrastEnhancement = AttributeDict()
params.contrastEnhancement.R = 10
# load old results or re-calculate? save results?
params.differenceMatrix = AttributeDict()
params.differenceMatrix.save = 1
params.differenceMatrix.load = 0 #1
params.contrastEnhanced = AttributeDict()
params.contrastEnhanced.save = 1
params.contrastEnhanced.load = 0 #1
# suffix appended on files containing the results
params.saveSuffix = ''
return params
def apply_tagger(self, x, apply_noise, y=None):
""" Build one path of Tagger """
mb_size = x.shape[1]
input_shape = (self.p.n_groups, mb_size) + self.in_dim
in_dim = np.prod(self.in_dim)
# Add noise
x_corr = self.corrupt(x) if apply_noise else x
# Repeat input
x_corr = T.repeat(x_corr, self.p.n_groups, 0)
# Compute v
if self.p.input_type == 'binary':
v = None
elif self.p.input_type == 'continuous':
v = self.weight(1., 'v')
v = v * T.alloc(1., *input_shape)
# Cap to positive range
v = nn.exp_inv_sinh(v)
d = AttributeDict()
if y:
d.pred = []
d.class_error, d.class_cost = [], []
# here we have the book-keeping of z and m for the visualizations.
d.z = []
d.m = []
else:
d.denoising_cost, d.ami_score, d.ami_score_per_sample = [], [], []
assert self.p.n_iterations >= 1
# z_hat is the value for the next iteration of tagger.
# z is the current iteration tagger input
# m is the current iteration mask input
# m_hat is the value for the next iteration of tagger.
# m_lh is the mask likelihood.
# z_delta is the gradient of z, which depends on x, z and m.
for step in xrange(self.p.n_iterations):
# Encoder
# =======
# Compute m, z and z_hat_pre_bin
if step == 0:
# No values from previous iteration, so let's make them up
m, z = self.init_m_z(input_shape)
z_hat_pre_bin = None
# let's keep in the bookkeeping for the visualizations.
if y:
d.z.append(z)
d.m.append(m)
else:
# Feed in the previous iteration's estimates
z = z_hat
m = m_hat
# Compute m_lh
m_lh = self.m_lh(x_corr, z, v)
z_delta = self.f_z_deriv(x_corr, z, m)
z_tilde = z_hat_pre_bin if z_hat_pre_bin is not None else z
# Concatenate all inputs
inputs = [z_tilde, z_delta, m, m_lh]
inputs = T.concatenate(inputs, axis=2)
# Projection, batch-normalization and activation to a hidden layer
z = self.proj(inputs, in_dim * 4, self.p.encoder_proj[0])
z -= z.mean((0, 1), keepdims=True)
z /= T.sqrt(z.var((0, 1), keepdims=True) + np.float32(1e-10))
z += self.bias(0.0 * np.ones(self.p.encoder_proj[0]), 'b')
h = self.apply_act(z, 'relu')
# The first dimension is the group. Let's flatten together with
# minibatch in order to have parametric mapping compute all groups
# in parallel
h, undo_flatten = flatten_first_two_dims(h)
# Parametric Mapping
# ==================
self.ladder.apply(None, self.y, h)
ladder_encoder_output = undo_flatten(
self.ladder.act.corr.unlabeled.h[len(self.p.encoder_proj) - 1])
ladder_decoder_output = undo_flatten(self.ladder.act.est.z[0])
# Decoder
# =======
# compute z_hat
z_u = self.proj(ladder_decoder_output,
self.p.encoder_proj[0],
in_dim,
scope='z_u')
z_u -= z_u.mean((0, 1), keepdims=True)
z_u /= T.sqrt(z_u.var((0, 1), keepdims=True) + np.float32(1e-10))
z_hat = self.weight(np.ones(in_dim), 'c1') * z_u + self.bias(
np.zeros(in_dim), 'b1')
z_hat = z_hat.reshape(input_shape)
# compute m_hat
m_u = self.proj(ladder_decoder_output,
self.p.encoder_proj[0],
in_dim,
scope='m_u')
m_u -= m_u.mean((0, 1), keepdims=True)
m_u /= T.sqrt(m_u.var((0, 1), keepdims=True) + np.float32(1e-10))
c = self.weight(np.float32(1), 'c2')
m_hat = nn.softmax_n(m_u * c, axis=0)
m_hat = m_hat.reshape(input_shape)
# Apply sigmoid activation if input_type is binary
if self.p.input_type == 'binary':
z_hat_pre_bin = z_hat
z_hat = self.apply_act(z_hat, 'sigmoid')
# Collapse layer
# ==============
# Remove the last dim, which is assumed to be class 'None'
pred = ladder_encoder_output[:, :, :-1]
# Normalize
pred /= T.sum(T.sum(pred, axis=2, keepdims=True),
axis=0,
keepdims=True)
# Denoising and Classification costs
# ==================================
if y:
class_cost, class_error = self.compute_classification_cost_and_error(
pred, y)
d.pred.append(pred)
d.class_cost.append(class_cost)
d.class_error.append(class_error)
d.m.append(m_hat)
d.z.append(z_hat)
else:
d.denoising_cost.append(self.denoising_cost(
z_hat, m_hat, x, v))
ami_score, ami_score_per_sample = self.mask_accuracy(
self.masks_unlabeled, m_hat)
d.ami_score.append(ami_score)
d.ami_score_per_sample.append(ami_score_per_sample)
# stack the list of tensors into one
d = AttributeDict(
{key: T.stacklists(val)
for key, val in d.iteritems()})
return d
def setup_data(p, test_set=False):
dataset_class, training_set_size = {
'cifar10': (CIFAR10, 40000),
'mnist': (MNIST, 50000),
'conll': (EMBOOT_CONLL, 13900),
'ontonotes': (EMBOOT_ONTO, 67000)
}[p.dataset]
print("p.dataset = ", p.dataset)
# Allow overriding the default from command line
if p.get('unlabeled_samples') is not None:
training_set_size = p.unlabeled_samples
print("Training set size : ", training_set_size)
train_set = dataset_class(["train"])
print("train_set.num_examples : ", train_set.num_examples)
# Make sure the MNIST data is in right format
if p.dataset == 'mnist':
d = train_set.data_sources[train_set.sources.index('features')]
assert numpy.all(d <= 1.0) and numpy.all(d >= 0.0), \
'Make sure data is in float format and in range 0 to 1'
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
if p.get('dseed'):
rng = numpy.random.RandomState(seed=p.dseed)
rng.shuffle(all_ind)
d = AttributeDict()
# Choose the training set
d.train = train_set
d.train_ind = all_ind[:training_set_size]
# Then choose validation set from the remaining indices
d.valid = train_set
d.valid_ind = numpy.setdiff1d(all_ind, d.train_ind)[:p.valid_set_size]
logger.info('Using %d examples for validation' % len(d.valid_ind))
# Only touch test data if requested
if test_set:
d.test = dataset_class(["test"])
d.test_ind = numpy.arange(d.test.num_examples)
print("d.test.num_examples = ", d.test.num_examples)
in_dim = train_set.data_sources[train_set.sources.index(
'features')].shape[1:]
if p.dataset == 'conll' or p.dataset == 'ontonotes':
whiten = None
cnorm = None
# Setup optional whitening, only used for Cifar-10
elif p.dataset == 'cifar10':
if len(in_dim) > 1 and p.whiten_zca > 0:
assert numpy.product(in_dim) == p.whiten_zca, \
'Need %d whitening dimensions, not %d' % (numpy.product(in_dim),
p.whiten_zca)
cnorm = ContrastNorm(p.contrast_norm) if p.contrast_norm != 0 else None
def get_data(d, i):
data = d.get_data(request=list(i))[d.sources.index('features')]
# Fuel provides Cifar in uint8, convert to float32
data = numpy.require(data, dtype=numpy.float32)
return data if cnorm is None else cnorm.apply(data)
if p.whiten_zca > 0:
logger.info('Whitening using %d ZCA components' % p.whiten_zca)
whiten = ZCA()
whiten.fit(p.whiten_zca, get_data(d.train, d.train_ind))
else:
whiten = None
return in_dim, d, whiten, cnorm
def setup_data(p, test_set=False):
# CIFAR10与MNIST都是封装过后的HDF5数据集
# p.dataset为命令行传入的参数，在cifar10与mnist之间选择其一
dataset_class, training_set_size = {
'cifar10': (CIFAR10, 40000),
'mnist': (MNIST, 50000),
}[p.dataset]
# 可以通过命令行指定为标注样本的大小
# Allow overriding the default from command line
if p.get('unlabeled_samples') is not None:
training_set_size = p.unlabeled_samples
# 选出mnist数据集里面的train子集
train_set = dataset_class("train")
# Make sure the MNIST data is in right format
# 对minst进行数据检查，查看是否所有值都在0-1之间且都为float
if p.dataset == 'mnist':
# features大小为60000＊1＊28＊28，num_examples*channel*height*weight，minst为灰度图片所以channel=1
d = train_set.data_sources[train_set.sources.index('features')]
assert numpy.all(d <= 1.0) and numpy.all(d >= 0.0), \
'Make sure data is in float format and in range 0 to 1'
# 随机打乱样本顺序
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
if p.get('dseed'):
# 通过dseed制作一个随机器，用于打乱样本编号
rng = numpy.random.RandomState(seed=p.dseed)
rng.shuffle(all_ind)
d = AttributeDict()
# Choose the training set
d.train = train_set
# 此时index应该都被打乱
# 取出前training_set_size个数的样本做为训练集（的index）
d.train_ind = all_ind[:training_set_size]
# 选出一部分数据作为验证集
# Then choose validation set from the remaining indices
d.valid = train_set
# 全部的数据集中去掉训练用的样本，剩下的作为验证集
d.valid_ind = numpy.setdiff1d(all_ind, d.train_ind)[:p.valid_set_size]
logger.info('Using %d examples for validation' % len(d.valid_ind))
# 如果有测试数据的话，生成测试数据的index
# Only touch test data if requested
if test_set:
d.test = dataset_class("test")
d.test_ind = numpy.arange(d.test.num_examples)
# Setup optional whitening, only used for Cifar-10
# 计算特征值的维度，shape[1:]：获取第一个样本的维度
in_dim = train_set.data_sources[train_set.sources.index('features')].shape[1:]
if len(in_dim) > 1 and p.whiten_zca > 0:
assert numpy.product(in_dim) == p.whiten_zca, \
'Need %d whitening dimensions, not %d' % (numpy.product(in_dim),
p.whiten_zca)
# 归一化参数如果不为空，创建归一化类
cnorm = ContrastNorm(p.contrast_norm) if p.contrast_norm != 0 else None
def get_data(d, i):
data = d.get_data(request=i)[d.sources.index('features')]
# Fuel provides Cifar in uint8, convert to float32
# 检查data集合中的item是否符合float32类型
data = numpy.require(data, dtype=numpy.float32)
# TODO ContrastNorm.apply
return data if cnorm is None else cnorm.apply(data)
if p.whiten_zca > 0:
logger.info('Whitening using %d ZCA components' % p.whiten_zca)
# TODO ZCA
whiten = ZCA()
whiten.fit(p.whiten_zca, get_data(d.train, d.train_ind))
else:
whiten = None
return in_dim, d, whiten, cnorm
def setup_data(p, test_set=False):
dataset_class = {
'cifar10': (CIFAR10),
'jos' : (JOS),
'mnist': (MNIST),
}[p.dataset]
training_set_size = p.unlabeled_samples
# Allow overriding the default from command line
if p.get('unlabeled_samples') is not None:
training_set_size = p.unlabeled_samples
train_set = dataset_class(["train"])
# Make sure the MNIST data is in right format
if p.dataset == 'mnist':
d = train_set.data_sources[train_set.sources.index('features')]
assert numpy.all(d <= 1.0) and numpy.all(d >= 0.0), \
'Make sure data is in float format and in range 0 to 1'
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
if p.get('dseed'):
rng = numpy.random.RandomState(seed=p.dseed)
rng.shuffle(all_ind)
d = AttributeDict()
# Choose the training set
d.train = train_set
d.train_ind = all_ind[:training_set_size]
# Then choose validation set from the remaining indices
d.valid = train_set
d.valid_ind = numpy.setdiff1d(all_ind, d.train_ind)[:p.valid_set_size]
logger.info('Using %d examples for validation' % len(d.valid_ind))
# Only touch test data if requested
if test_set:
d.test = dataset_class(["test"])
d.test_ind = numpy.arange(d.test.num_examples)
# Setup optional whitening, only used for Cifar-10
fn = find_in_data_path(train_set.filename)
#iprint(fn)
s1 = H5PYDataset(fn, ("train",))
handle = s1.open()
in_dim = s1.get_data(handle,slice(0,1))[0].shape[1:]
s1.close(handle)
#in_dim = train_set.data_sources[train_set.sources.index('features')].shape[1:]
if len(in_dim) > 1 and p.whiten_zca > 0:
assert numpy.product(in_dim) == p.whiten_zca, \
'Need %d whitening dimensions, not %d' % (numpy.product(in_dim),
p.whiten_zca)
cnorm = ContrastNorm(p.contrast_norm) if p.contrast_norm != 0 else None
def get_data(d, i):
data = d.get_data(request=list(i))[d.sources.index('features')]
# Fuel provides Cifar in uint8, convert to float32
data = numpy.require(data, dtype=numpy.float32)
return data if cnorm is None else cnorm.apply(data)
if p.whiten_zca > 0:
logger.info('Whitening using %d ZCA components' % p.whiten_zca)
whiten = ZCA()
whiten.fit(p.whiten_zca, get_data(d.train, d.train_ind))
else:
whiten = None
return in_dim, d, whiten, cnorm
def setup_data(p, test_set=False):
if p.dataset in ['cifar10','mnist']:
dataset_class, training_set_size = {
'cifar10': (CIFAR10, 40000),
'mnist': (MNIST, 50000),
}[p.dataset]
else:
from fuel.datasets import H5PYDataset
from fuel.utils import find_in_data_path
from functools import partial
fn=p.dataset
fn=os.path.join(fn, fn + '.hdf5')
def dataset_class(which_sets):
return H5PYDataset(file_or_path=find_in_data_path(fn),
which_sets=which_sets,
load_in_memory=True)
training_set_size = None
train_set = dataset_class(["train"])
# Allow overriding the default from command line
if p.get('unlabeled_samples') is not None and p.unlabeled_samples >= 0:
training_set_size = p.unlabeled_samples
elif training_set_size is None:
training_set_size = train_set.num_examples
# Make sure the MNIST data is in right format
if p.dataset == 'mnist':
d = train_set.data_sources[train_set.sources.index('features')]
assert numpy.all(d <= 1.0) and numpy.all(d >= 0.0), \
'Make sure data is in float format and in range 0 to 1'
# Take all indices and permutate them
all_ind = numpy.arange(train_set.num_examples)
if p.get('dseed'):
rng = numpy.random.RandomState(seed=p.dseed)
rng.shuffle(all_ind)
d = AttributeDict()
# Choose the training set
d.train = train_set
d.train_ind = all_ind[:training_set_size]
# Then choose validation set from the remaining indices
d.valid = train_set
d.valid_ind = numpy.setdiff1d(all_ind, d.train_ind)[:p.valid_set_size]
logger.info('Using %d examples for validation' % len(d.valid_ind))
# Only touch test data if requested
if test_set:
d.test = dataset_class(["test"])
d.test_ind = numpy.arange(d.test.num_examples)
# Setup optional whitening, only used for Cifar-10
in_dim = train_set.data_sources[train_set.sources.index('features')].shape[1:]
if len(in_dim) > 1 and p.whiten_zca > 0:
assert numpy.product(in_dim) == p.whiten_zca, \
'Need %d whitening dimensions, not %d' % (numpy.product(in_dim),
p.whiten_zca)
cnorm = ContrastNorm(p.contrast_norm) if p.contrast_norm != 0 else None
def get_data(d, i):
data = d.get_data(request=i)[d.sources.index('features')]
# Fuel provides Cifar in uint8, convert to float32
data = numpy.require(data, dtype=numpy.float32)
return data if cnorm is None else cnorm.apply(data)
if p.whiten_zca > 0:
logger.info('Whitening using %d ZCA components' % p.whiten_zca)
whiten = ZCA()
whiten.fit(p.whiten_zca, get_data(d.train, d.train_ind))
else:
whiten = None
return in_dim, d, whiten, cnorm
def apply(self, input_labeled, target_labeled, input_unlabeled):
self.layer_counter = 0
input_dim = self.p.encoder_layers[0]
# Store the dimension tuples in the same order as layers.
layers = self.layers
self.layer_dims = {0: input_dim}
self.lr = self.shared(self.default_lr, 'learning_rate', role=None)
self.costs = costs = AttributeDict()
self.costs.denois = AttributeDict()
self.act = AttributeDict()
self.error = AttributeDict()
top = len(layers) - 1
N = input_labeled.shape[0]
self.join = lambda l, u: T.concatenate([l, u], axis=0)
self.labeled = lambda x: x[:N] if x is not None else x
self.unlabeled = lambda x: x[N:] if x is not None else x
self.split_lu = lambda x: (self.labeled(x), self.unlabeled(x))
input_concat = self.join(input_labeled, input_unlabeled)
def encoder(input_, path_name, input_noise_std=0, noise_std=[]):
h = input_
logger.info(' 0: noise %g' % input_noise_std)
if input_noise_std > 0.:
h = h + self.noise_like(h) * input_noise_std
d = AttributeDict()
d.unlabeled = self.new_activation_dict()
d.labeled = self.new_activation_dict()
d.labeled.z[0] = self.labeled(h)
d.unlabeled.z[0] = self.unlabeled(h)
prev_dim = input_dim
for i, (spec, _, act_f) in layers[1:]:
d.labeled.h[i - 1], d.unlabeled.h[i - 1] = self.split_lu(h)
noise = noise_std[i] if i < len(noise_std) else 0.
curr_dim, z, m, s, h = self.f(h,
prev_dim,
spec,
i,
act_f,
path_name=path_name,
noise_std=noise)
assert self.layer_dims.get(i) in (None, curr_dim)
self.layer_dims[i] = curr_dim
d.labeled.z[i], d.unlabeled.z[i] = self.split_lu(z)
d.unlabeled.s[i] = s
d.unlabeled.m[i] = m
prev_dim = curr_dim
d.labeled.h[i], d.unlabeled.h[i] = self.split_lu(h)
return d
# Clean, supervised
logger.info('Encoder: clean, labeled')
clean = self.act.clean = encoder(input_concat, 'clean')
# Corrupted, supervised
logger.info('Encoder: corr, labeled')
corr = self.act.corr = encoder(input_concat,
'corr',
input_noise_std=self.p.super_noise_std,
noise_std=self.p.f_local_noise_std)
est = self.act.est = self.new_activation_dict()
# Decoder path in opposite order
logger.info('Decoder: z_corr -> z_est')
for i, ((_, spec), l_type, act_f) in layers[::-1]:
z_corr = corr.unlabeled.z[i]
z_clean = clean.unlabeled.z[i]
z_clean_s = clean.unlabeled.s.get(i)
z_clean_m = clean.unlabeled.m.get(i)
fspec = layers[i + 1][1][0] if len(layers) > i + 1 else (None,
None)
if i == top:
ver = corr.unlabeled.h[i]
ver_dim = self.layer_dims[i]
top_g = True
else:
ver = est.z.get(i + 1)
ver_dim = self.layer_dims.get(i + 1)
top_g = False
z_est = self.g(z_lat=z_corr,
z_ver=ver,
in_dims=ver_dim,
out_dims=self.layer_dims[i],
l_type=l_type,
num=i,
fspec=fspec,
top_g=top_g)
if z_est is not None:
# Denoising cost
if z_clean_s:
z_est_norm = (z_est - z_clean_m) / z_clean_s
else:
z_est_norm = z_est
se = SquaredError('denois' + str(i))
costs.denois[i] = se.apply(z_est_norm.flatten(2),
z_clean.flatten(2)) \
/ np.prod(self.layer_dims[i], dtype=floatX)
costs.denois[i].name = 'denois' + str(i)
denois_print = 'denois %.2f' % self.p.denoising_cost_x[i]
else:
denois_print = ''
# Store references for later use
est.h[i] = self.apply_act(z_est, act_f)
est.z[i] = z_est
est.s[i] = None
est.m[i] = None
logger.info(' g%d: %10s, %s, dim %s -> %s' %
(i, l_type, denois_print, self.layer_dims.get(i + 1),
self.layer_dims.get(i)))
# Costs
y = target_labeled.flatten()
costs.class_clean = CategoricalCrossEntropy().apply(
y, clean.labeled.h[top])
costs.class_clean.name = 'cost_class_clean'
costs.class_corr = CategoricalCrossEntropy().apply(
y, corr.labeled.h[top])
costs.class_corr.name = 'cost_class_corr'
# This will be used for training
costs.total = costs.class_corr * 1.0
for i in range(top + 1):
if costs.denois.get(i) and self.p.denoising_cost_x[i] > 0:
costs.total += costs.denois[i] * self.p.denoising_cost_x[i]
costs.total.name = 'cost_total'
# Classification error
mr = MisclassificationRate()
self.error.clean = mr.apply(y, clean.labeled.h[top]) * np.float32(100.)
self.error.clean.name = 'error_rate_clean'
import sys
from utils import AttributeDict
from tagger_exp import TaggerExperiment
p = AttributeDict()
p.encoder_proj = (3000, 2000, 1000)
p.input_noise = 0.2
p.class_cost_x = 0.
p.zhat_init_value = 0.5
p.n_iterations = 3
p.n_groups = 4
p.lr = 0.001
p.labeled_samples = 1000
p.save_freq = 50
p.seed = 1
p.num_epochs = 150
p.batch_size = 100
p.valid_batch_size = 100
p.objects_per_sample = 2
p.dataset = 'freq20-2mnist'
p.input_type = 'continuous'
if __name__ == '__main__':
if len(sys.argv) == 2 and sys.argv[1] == '--pretrain':
p.save_to = 'freq20-2mnist-pretraining'
experiment = TaggerExperiment(p)
experiment.train()
elif len(sys.argv) == 3 and sys.argv[1] == '--continue':
def __init__(self, *args, **kwargs):
AttributeDict.__init__(self)
self.env = Environment.getInstance()
self.resourceType = self.__class__.__name__
self.isUpdated = False
seen = set()
if not hasattr(self, '_schema'):
raise Fail("Resource failed to define a valid _schema")
# union global schema with local schema
schema = copy.deepcopy(self._schema)
for key in self.s_globalSchema:
if not key in schema:
schema[key] = self.s_globalSchema[key]
resolvedArgs = { }
keys = schema.keys()
keysLen = len(keys)
index = 0
# resolve unnamed arguments with names corresponding to the order
# they were passed to Resource's ctor and their relative definitions
# in the subclass' ResourceArgumentSchema (which is an OrderedDict,
# so as to retain this ordering information).
for arg in args:
if index < keysLen:
key = keys[index]
resolvedArgs[keys[index]] = arg
else:
raise InvalidArgument("Invalid unnamed argument %s provided to resource %s" % (arg, str(self)))
index += 1
for arg in kwargs:
if arg in resolvedArgs:
raise InvalidArgument("Invalid mixture of named and unnamed arguments provided to resource %s, possibly around argument %s" % (str(self), arg))
else:
resolvedArgs[arg] = kwargs[arg]
utils.log("Initializing resource '%s' with args: %s" % (self.resourceType, resolvedArgs))
# validate resource arguments
output = schema.validate(resolvedArgs)
for key in output:
self[key] = output[key]
self.subscriptions = {
'immediate' : set(),
'delayed' : set()
}
for sub in self.subscribes:
if len(sub) == 2:
action, resource = sub
immediate = False
else:
action, resource, immediate = sub
resource.subscribe(action, self, immediate)
for sub in self.notifies:
self.subscribe(*sub)
self._validate()
self._register()
utils.log("Added new resource '%s'" % (str(self), ))
def apply_tagger(self, x, apply_noise, y=None):
""" Build one path of Tagger """
mb_size = x.shape[1]
input_shape = (self.p.n_groups, mb_size) + self.in_dim
in_dim = np.prod(self.in_dim)
# Add noise
x_corr = self.corrupt(x) if apply_noise else x
# Repeat input
x_corr = T.repeat(x_corr, self.p.n_groups, 0)
# Compute v
if self.p.input_type == 'binary':
v = None
elif self.p.input_type == 'continuous':
v = self.weight(1., 'v')
v = v * T.alloc(1., *input_shape)
# Cap to positive range
v = nn.exp_inv_sinh(v)
d = AttributeDict()
if y:
d.pred = []
d.class_error, d.class_cost = [], []
# here we have the book-keeping of z and m for the visualizations.
d.z = []
d.m = []
else:
d.denoising_cost, d.ami_score, d.ami_score_per_sample = [], [], []
assert self.p.n_iterations >= 1
# z_hat is the value for the next iteration of tagger.
# z is the current iteration tagger input
# m is the current iteration mask input
# m_hat is the value for the next iteration of tagger.
# m_lh is the mask likelihood.
# z_delta is the gradient of z, which depends on x, z and m.
for step in xrange(self.p.n_iterations):
# Encoder
# =======
# Compute m, z and z_hat_pre_bin
if step == 0:
# No values from previous iteration, so let's make them up
m, z = self.init_m_z(input_shape)
z_hat_pre_bin = None
# let's keep in the bookkeeping for the visualizations.
if y:
d.z.append(z)
d.m.append(m)
else:
# Feed in the previous iteration's estimates
z = z_hat
m = m_hat
# Compute m_lh
m_lh = self.m_lh(x_corr, z, v)
z_delta = self.f_z_deriv(x_corr, z, m)
z_tilde = z_hat_pre_bin if z_hat_pre_bin is not None else z
# Concatenate all inputs
inputs = [z_tilde, z_delta, m, m_lh]
inputs = T.concatenate(inputs, axis=2)
# Projection, batch-normalization and activation to a hidden layer
z = self.proj(inputs, in_dim * 4, self.p.encoder_proj[0])
z -= z.mean((0, 1), keepdims=True)
z /= T.sqrt(z.var((0, 1), keepdims=True) + np.float32(1e-10))
z += self.bias(0.0 * np.ones(self.p.encoder_proj[0]), 'b')
h = self.apply_act(z, 'relu')
# The first dimension is the group. Let's flatten together with
# minibatch in order to have parametric mapping compute all groups
# in parallel
h, undo_flatten = flatten_first_two_dims(h)
# Parametric Mapping
# ==================
self.ladder.apply(None, self.y, h)
ladder_encoder_output = undo_flatten(self.ladder.act.corr.unlabeled.h[len(self.p.encoder_proj) - 1])
ladder_decoder_output = undo_flatten(self.ladder.act.est.z[0])
# Decoder
# =======
# compute z_hat
z_u = self.proj(ladder_decoder_output, self.p.encoder_proj[0], in_dim, scope='z_u')
z_u -= z_u.mean((0, 1), keepdims=True)
z_u /= T.sqrt(z_u.var((0, 1), keepdims=True) + np.float32(1e-10))
z_hat = self.weight(np.ones(in_dim), 'c1') * z_u + self.bias(np.zeros(in_dim), 'b1')
z_hat = z_hat.reshape(input_shape)
# compute m_hat
m_u = self.proj(ladder_decoder_output, self.p.encoder_proj[0], in_dim, scope='m_u')
m_u -= m_u.mean((0, 1), keepdims=True)
m_u /= T.sqrt(m_u.var((0, 1), keepdims=True) + np.float32(1e-10))
c = self.weight(np.float32(1), 'c2')
m_hat = nn.softmax_n(m_u * c, axis=0)
m_hat = m_hat.reshape(input_shape)
# Apply sigmoid activation if input_type is binary
if self.p.input_type == 'binary':
z_hat_pre_bin = z_hat
z_hat = self.apply_act(z_hat, 'sigmoid')
# Collapse layer
# ==============
# Remove the last dim, which is assumed to be class 'None'
pred = ladder_encoder_output[:, :, :-1]
# Normalize
pred /= T.sum(T.sum(pred, axis=2, keepdims=True), axis=0, keepdims=True)
# Denoising and Classification costs
# ==================================
if y:
class_cost, class_error = self.compute_classification_cost_and_error(pred, y)
d.pred.append(pred)
d.class_cost.append(class_cost)
d.class_error.append(class_error)
d.m.append(m_hat)
d.z.append(z_hat)
else:
d.denoising_cost.append(self.denoising_cost(z_hat, m_hat, x, v))
ami_score, ami_score_per_sample = self.mask_accuracy(self.masks_unlabeled, m_hat)
d.ami_score.append(ami_score)
d.ami_score_per_sample.append(ami_score_per_sample)
# stack the list of tensors into one
d = AttributeDict({key: T.stacklists(val) for key, val in d.iteritems()})
return d
from utils import AttributeDict
from tagger_exp import TaggerExperiment
p = AttributeDict()
p.encoder_proj = (2000, 1000, 500)
p.input_noise = 0.2
p.class_cost_x = 0
p.zhat_init_value = 0.26 # mean of the input data.
p.n_iterations = 3
p.n_groups = 4
p.lr = 0.0004
p.seed = 10
p.num_epochs = 100
p.batch_size = 100
p.valid_batch_size = 100
p.dataset = 'shapes50k20x20'
p.input_type = 'binary'
p.save_to = 'shapes50k20x20'
if __name__ == '__main__':
experiment = TaggerExperiment(p)
experiment.train()