def check_patch_coverage(files):
rng = numpy.random.RandomState(1)
inputs = [(name, array.shape) for name, array in six.iteritems(files)]
shape = (5, 7, 7)
for fname, index in spatiotemporal_cubes(inputs, shape, 50000, rng):
cube = files[fname][index]
if len(files[fname].shape) == 3:
assert cube.shape == shape
else:
assert cube.shape[:3] == shape[:3]
cube[...] = True
for fname, array in six.iteritems(files):
assert array.all()
def check_patch_coverage(files):
rng = numpy.random.RandomState(1)
inputs = [(name, array.shape) for name, array in six.iteritems(files)]
shape = (5, 7, 7)
for fname, index in spatiotemporal_cubes(inputs, shape, 50000, rng):
cube = files[fname][index]
if len(files[fname].shape) == 3:
assert cube.shape == shape
else:
assert cube.shape[:3] == shape[:3]
cube[...] = True
for fname, array in six.iteritems(files):
assert array.all()
def get_updates(self, grads, lr_scalers=None):
"""
Provides the updates for learning with gradient descent + momentum.
Parameters
----------
learning_rate : float
Learning rate coefficient.
grads : dict
A dictionary mapping from the model's parameters to their
gradients.
lr_scalers : dict
A dictionary mapping from the model's parameters to a learning
rate multiplier.
"""
updates = OrderedDict()
for (param, grad) in six.iteritems(grads):
vel = theano.shared(param.get_value() * 0.)
assert param.dtype == vel.dtype
assert grad.dtype == param.dtype
if param.name is not None:
vel.name = 'vel_' + param.name
scaled_lr = self.learning_rate * lr_scalers.get(param, 1.)
updates[vel] = self.momentum * vel - scaled_lr * grad
inc = updates[vel]
if self.nesterov_momentum:
inc = self.momentum * inc - scaled_lr * grad
assert inc.dtype == vel.dtype
updates[param] = param + inc
return updates
def _get_lib_versions(self):
"""Get version of Python packages."""
repos = os.getenv('PYLEARN2_TRACK_MODULES', '')
default_repos = 'pylearn2:theano:numpy:scipy'
repos = default_repos + ":" + repos
repos = set(repos.split(':'))
for repo in repos:
try:
if repo == '':
continue
__import__(repo)
if hasattr(sys.modules[repo], '__version__'):
v = sys.modules[repo].__version__
if v != 'unknown':
self.versions[repo] = v
continue
self.versions[repo] = self._get_git_version(
self._get_module_parent_path(sys.modules[repo]))
except ImportError:
self.versions[repo] = None
known = copy.copy(self.versions)
# Put together all modules with unknown versions.
unknown = [k for k, w in known.items() if not w]
known = dict((k, w) for k, w in known.items() if w)
# Print versions.
self.str_versions = ' | '.join(
['%s:%s' % (k, w) for k, w in sorted(six.iteritems(known))] +
['%s:?' % ','.join(sorted(unknown))])
def _get_lib_versions(self):
"""Get version of Python packages."""
repos = os.getenv('PYLEARN2_TRACK_MODULES', '')
default_repos = 'pylearn2:theano:numpy:scipy'
repos = default_repos + ":" + repos
repos = set(repos.split(':'))
for repo in repos:
try:
if repo == '':
continue
__import__(repo)
if hasattr(sys.modules[repo], '__version__'):
v = sys.modules[repo].__version__
if v != 'unknown':
self.versions[repo] = v
continue
self.versions[repo] = self._get_git_version(
self._get_module_parent_path(sys.modules[repo]))
except ImportError:
self.versions[repo] = None
known = copy.copy(self.versions)
# Put together all modules with unknown versions.
unknown = [k for k, w in known.items() if not w]
known = dict((k, w) for k, w in known.items() if w)
# Print versions.
self.str_versions = ' | '.join(
['%s:%s' % (k, w) for k, w in sorted(six.iteritems(known))] +
['%s:?' % ','.join(sorted(unknown))])
def _str_index(self):
idx = self['index']
out = []
out += ['.. index:: %s' % idx.get('default','')]
for section, references in six.iteritems(idx):
if section == 'default':
continue
out += [' :%s: %s' % (section, ', '.join(references))]
return out
def _str_index(self):
idx = self['index']
out = []
out += ['.. index:: %s' % idx.get('default', '')]
for section, references in six.iteritems(idx):
if section == 'default':
continue
out += [' :%s: %s' % (section, ', '.join(references))]
return out
def docstring_errors(filename, global_dict=None):
"""
Run a Python file, parse the docstrings of all the classes
and functions it declares, and return them.
Parameters
----------
filename : str
Filename of the module to run.
global_dict : dict, optional
Globals dictionary to pass along to `execfile()`.
Returns
-------
all_errors : list
Each entry of the list is a tuple, of length 2 or 3, with
format either
(func_or_class_name, docstring_error_description)
or
(class_name, method_name, docstring_error_description)
"""
if global_dict is None:
global_dict = {}
if '__file__' not in global_dict:
global_dict['__file__'] = filename
if '__doc__' not in global_dict:
global_dict['__doc__'] = None
try:
with open(filename) as f:
code = compile(f.read(), filename, 'exec')
exec(code, global_dict)
except SystemExit:
pass
except SkipTest:
raise AssertionError("Couldn't verify format of " + filename +
"due to SkipTest")
all_errors = []
for key, val in six.iteritems(global_dict):
if not key.startswith('_'):
module_name = ""
if hasattr(inspect.getmodule(val), '__name__'):
module_name = inspect.getmodule(val).__name__
if (inspect.isfunction(val) or inspect.isclass(val)) and\
(inspect.getmodule(val) is None
or module_name == '__builtin__'):
if inspect.isfunction(val):
all_errors.extend(handle_function(val, key))
elif inspect.isclass(val):
all_errors.extend(handle_class(val, key))
elif key == '__doc__':
all_errors.extend(handle_module(val, key))
if all_errors:
all_errors.insert(0, ("%s:"%filename,))
return all_errors
def docstring_errors(filename, global_dict=None):
"""
Run a Python file, parse the docstrings of all the classes
and functions it declares, and return them.
Parameters
----------
filename : str
Filename of the module to run.
global_dict : dict, optional
Globals dictionary to pass along to `execfile()`.
Returns
-------
all_errors : list
Each entry of the list is a tuple, of length 2 or 3, with
format either
(func_or_class_name, docstring_error_description)
or
(class_name, method_name, docstring_error_description)
"""
if global_dict is None:
global_dict = {}
if '__file__' not in global_dict:
global_dict['__file__'] = filename
if '__doc__' not in global_dict:
global_dict['__doc__'] = None
try:
with open(filename) as f:
code = compile(f.read(), filename, 'exec')
exec(code, global_dict)
except SystemExit:
pass
except SkipTest:
raise AssertionError("Couldn't verify format of " + filename +
"due to SkipTest")
all_errors = []
for key, val in six.iteritems(global_dict):
if not key.startswith('_'):
module_name = ""
if hasattr(inspect.getmodule(val), '__name__'):
module_name = inspect.getmodule(val).__name__
if (inspect.isfunction(val) or inspect.isclass(val)) and\
(inspect.getmodule(val) is None
or module_name == '__builtin__'):
if inspect.isfunction(val):
all_errors.extend(handle_function(val, key))
elif inspect.isclass(val):
all_errors.extend(handle_class(val, key))
elif key == '__doc__':
all_errors.extend(handle_module(val, key))
if all_errors:
all_errors.insert(0, ("%s:" % filename, ))
return all_errors
def get_gradients(self, model, data, **kwargs):
gradients, updates = self.cost.get_gradients(model, data, **kwargs)
norm = tensor.sqrt(
tensor.sum([
tensor.sum(param_gradient**2)
for param, param_gradient in six.iteritems(gradients)
if param.name not in self.exclude_params
]))
clipped_gradients = OrderedDict()
for param, param_gradient in six.iteritems(gradients):
if param.name not in self.exclude_params:
clipped_gradients[param] = tensor.switch(
tensor.ge(norm, self.clipping_value),
param_gradient / norm * self.clipping_value,
param_gradient)
gradients.update(clipped_gradients)
return gradients, updates
def get_gradients(self, model, data, **kwargs):
gradients, updates = self.cost.get_gradients(model, data, **kwargs)
norm = tensor.sqrt(tensor.sum(
[tensor.sum(param_gradient ** 2) for param, param_gradient
in six.iteritems(gradients)
if param.name not in self.exclude_params]
))
clipped_gradients = OrderedDict()
for param, param_gradient in six.iteritems(gradients):
if param.name not in self.exclude_params:
clipped_gradients[param] = tensor.switch(
tensor.ge(norm, self.clipping_value),
param_gradient / norm * self.clipping_value,
param_gradient
)
gradients.update(clipped_gradients)
return gradients, updates
def get_updates(self, gradients):
"""
Based on Pylearn2
(https://github.com/lisa-lab/pylearn2/blob/master/pylearn2/training_algorithms/learning_rule.py)
Implements momentum as described in Section 9 of
"A Practical Guide to Training Restricted Boltzmann Machines",
Geoffrey Hinton.
Parameters are updated by the formula:
inc := momentum * inc - learning_rate * d cost / d param
param := param + inc
Also has the option to implement Nesterov momentum (accelerated momentum), which works better in a lot of cases.
Parameters
----------
gradients : dict
A dictionary mapping from the model's parameters to their
gradients.
Returns
-------
updates : OrderdDict
A dictionary mapping from the old model parameters, to their new
values after a single iteration of the learning rule.
"""
log.debug(
'Setting up Stochastic Gradient Descent with momentum for optimizer...'
)
updates = OrderedDict()
for (param, gradient) in six.iteritems(gradients):
velocity = sharedX(param.get_value() * 0.)
assert param.dtype == velocity.dtype
assert gradient.dtype == param.dtype
if param.name is not None:
velocity.name = 'vel_' + param.name
scaled_lr = self.learning_rate * self.lr_scalers.get(param, 1.)
updates[velocity] = self.momentum * velocity - scaled_lr * gradient
inc = updates[velocity]
if self.nesterov_momentum:
log.debug('Using Nesterov momentum for parameter %s',
str(param))
inc = self.momentum * inc - scaled_lr * gradient
assert inc.dtype == velocity.dtype
updates[param] = param + inc
return updates
def get_updates(self, gradients):
"""
This returns the parameter updates to use during training. It defaults to only using (annealed) learning rate.
:param gradients: (parameter, gradient) tuples representing the parameters to update and their gradients
:type gradients: list(tuple)
:return: the updates
:rtype: updates
"""
log.debug('Setting up Stochastic Gradient Descent for optimizer...')
updates = OrderedDict()
for (param, gradient) in six.iteritems(gradients):
scaled_lr = self.learning_rate * self.lr_scalers.get(param, 1.)
updates[param] = param - scaled_lr * gradient
return updates
def inverse_vocabulary(self):
"""
The inverse vocabulary, a dictionary from
integers to strings. If it does not exist,
it is created from the vocabulary if possible.
"""
if hasattr(self, '_inverse_vocabulary'):
return self._inverse_vocabulary
elif hasattr(self, '_vocabulary'):
self._inverse_vocabulary = dict((index, word) for word, index
in six.iteritems(self._vocabulary))
return self._inverse_vocabulary
else:
raise NotImplementedError
def get_updates(self, learning_rate, grads, lr_scalers=None):
"""
Provides the updates for learning with gradient descent + momentum.
Parameters
----------
learning_rate : float
Learning rate coefficient.
grads : dict
A dictionary mapping from the model's parameters to their
gradients.
lr_scalers : dict
A dictionary mapping from the model's parameters to a learning
rate multiplier.
"""
updates = OrderedDict()
for (param, grad) in six.iteritems(grads):
vel = sharedX(param.get_value() * 0.)
assert param.dtype == vel.dtype
assert grad.dtype == param.dtype
if param.name is not None:
vel.name = 'vel_' + param.name
scaled_lr = learning_rate * lr_scalers.get(param, 1.)
updates[vel] = self.momentum * vel - scaled_lr * grad
inc = updates[vel]
if self.nesterov_momentum:
inc = self.momentum * inc - scaled_lr * grad
assert inc.dtype == vel.dtype
print 'param', param.name, '------>> min:', param.get_value().min(), ', max:', param.get_value().max()
if(self.rr==0) or vel.name[-1]=='b':
updates[param] = param + inc
else:
scale = np.ones(param.get_value().shape)
temp = np.ones(param.get_value().shape)
for i in xrange(temp.shape[0]):
scale[i] = max(abs(param.get_value()[i].min()), param.get_value()[i].max())
# print 'param', param.name, 'scale[0]', scale[0], 'scale', scale, \
# 'scale.shape', scale.shape, 'param.get_value().shape', param.get_value().shape
updates[param] = self.rrf(param + inc, scale)
return updates
def get_updates(self, gradients):
"""
Based on Pylearn2
(https://github.com/lisa-lab/pylearn2/blob/master/pylearn2/training_algorithms/learning_rule.py)
Implements momentum as described in Section 9 of
"A Practical Guide to Training Restricted Boltzmann Machines",
Geoffrey Hinton.
Parameters are updated by the formula:
inc := momentum * inc - learning_rate * d cost / d param
param := param + inc
Also has the option to implement Nesterov momentum (accelerated momentum), which works better in a lot of cases.
Parameters
----------
gradients : dict
A dictionary mapping from the model's parameters to their
gradients.
Returns
-------
updates : OrderdDict
A dictionary mapping from the old model parameters, to their new
values after a single iteration of the learning rule.
"""
log.debug('Setting up Stochastic Gradient Descent with momentum for optimizer...')
updates = OrderedDict()
for (param, gradient) in six.iteritems(gradients):
velocity = sharedX(param.get_value() * 0.)
assert param.dtype == velocity.dtype
assert gradient.dtype == param.dtype
if param.name is not None:
velocity.name = 'vel_' + param.name
scaled_lr = self.learning_rate * self.lr_scalers.get(param, 1.)
updates[velocity] = self.momentum * velocity - scaled_lr * gradient
inc = updates[velocity]
if self.nesterov_momentum:
log.debug('Using Nesterov momentum for parameter %s', str(param))
inc = self.momentum * inc - scaled_lr * gradient
assert inc.dtype == velocity.dtype
updates[param] = param + inc
return updates
def inverse_vocabulary(self):
"""
The inverse vocabulary, a dictionary from
integers to strings. If it does not exist,
it is created from the vocabulary if possible.
"""
if hasattr(self, '_inverse_vocabulary'):
return self._inverse_vocabulary
elif hasattr(self, '_vocabulary'):
self._inverse_vocabulary = dict(
(index, word)
for word, index in six.iteritems(self._vocabulary))
return self._inverse_vocabulary
else:
raise NotImplementedError
def __str__(self):
args_dict = dict(self)
builder = args_dict.pop("__builder__", "")
ret_list = []
if builder:
ret_list.append("!obj:%s {" % builder)
else:
ret_list.append("{")
for key, val in six.iteritems(args_dict):
# This will call str() on keys and values, not repr(), so unicode
# objects will have the form 'blah', not "u'blah'".
ret_list.append("%s: %s," % (key, val))
ret_list.append("}")
return "\n".join(ret_list)
def __str__(self):
args_dict = dict(self)
builder = args_dict.pop('__builder__', '')
ret_list = []
if builder:
ret_list.append('!obj:%s {' % builder)
else:
ret_list.append('{')
for key, val in six.iteritems(args_dict):
# This will call str() on keys and values, not repr(), so unicode
# objects will have the form 'blah', not "u'blah'".
ret_list.append('%s: %s,' % (key, val))
ret_list.append('}')
return '\n'.join(ret_list)
def __str__(self):
args_dict = dict(self)
builder = args_dict.pop('__builder__', '')
ret_list = []
if builder:
ret_list.append('!obj:%s {' % builder)
else:
ret_list.append('{')
for key, val in six.iteritems(args_dict):
# This will call str() on keys and values, not repr(), so unicode
# objects will have the form 'blah', not "u'blah'".
ret_list.append('%s: %s,' % (key, val))
ret_list.append('}')
return '\n'.join(ret_list)
def test_impl(norb):
label_to_value_maps = (
# category
{
0: 'animal',
1: 'human',
2: 'airplane',
3: 'truck',
4: 'car',
5: 'blank'
},
# instance
dict(safe_zip(range(10), range(10))),
# elevation
dict(safe_zip(range(9),
numpy.arange(9) * 5 + 30)),
# azimuth
dict(safe_zip(range(0, 36, 2), numpy.arange(0, 360, 20))),
# lighting
dict(safe_zip(range(5), range(5))),
# horizontal shift
dict(safe_zip(range(-5, 6), range(-5, 6))),
# vertical shift
dict(safe_zip(range(-5, 6), range(-5, 6))),
# lumination change
dict(safe_zip(range(-19, 20), range(-19, 20))),
# contrast change
dict(safe_zip(range(2), (0.8, 1.3))))
# Use of zip rather than safe_zip intentional;
# norb.label_to_value_funcs will be shorter than
# label_to_value_maps if norb is small NORB.
for (label_to_value_map,
label_to_value_func) in zip(label_to_value_maps,
norb.label_to_value_funcs):
for label, expected_value in six.iteritems(label_to_value_map):
actual_value = label_to_value_func(label)
assert expected_value == actual_value
def safe_update(dict_to, dict_from):
"""
Like dict_to.update(dict_from), except don't overwrite any keys.
Parameters
----------
dict_to : WRITEME
dict_from : WRITEME
Returns
-------
WRITEME
"""
for key, val in six.iteritems(dict_from):
if key in dict_to:
raise KeyError(key)
dict_to[key] = val
return dict_to
def safe_update(dict_to, dict_from):
"""
Like dict_to.update(dict_from), except don't overwrite any keys.
Parameters
----------
dict_to : WRITEME
dict_from : WRITEME
Returns
-------
WRITEME
"""
for key, val in six.iteritems(dict_from):
if key in dict_to:
raise KeyError(key)
dict_to[key] = val
return dict_to
def test_impl(norb):
label_to_value_maps = (
# category
{0: 'animal',
1: 'human',
2: 'airplane',
3: 'truck',
4: 'car',
5: 'blank'},
# instance
dict(safe_zip(range(10), range(10))),
# elevation
dict(safe_zip(range(9), numpy.arange(9) * 5 + 30)),
# azimuth
dict(safe_zip(range(0, 36, 2), numpy.arange(0, 360, 20))),
# lighting
dict(safe_zip(range(5), range(5))),
# horizontal shift
dict(safe_zip(range(-5, 6), range(-5, 6))),
# vertical shift
dict(safe_zip(range(-5, 6), range(-5, 6))),
# lumination change
dict(safe_zip(range(-19, 20), range(-19, 20))),
# contrast change
dict(safe_zip(range(2), (0.8, 1.3))))
# Use of zip rather than safe_zip intentional;
# norb.label_to_value_funcs will be shorter than
# label_to_value_maps if norb is small NORB.
for (label_to_value_map,
label_to_value_func) in zip(label_to_value_maps,
norb.label_to_value_funcs):
for label, expected_value in six.iteritems(label_to_value_map):
actual_value = label_to_value_func(label)
assert expected_value == actual_value
def handle_class(val, class_name):
cls_errors = []
docstring = inspect.getdoc(val)
if docstring is None:
cls_errors.append((class_name, '**missing** class-level docstring'))
else:
cls_errors = [(e, ) for e in NumpyClassDocString(
docstring, class_name, val).get_errors()]
# Get public methods and parse their docstrings
methods = dict(((name, func) for name, func in inspect.getmembers(val)
if not name.startswith('_') and callable(func)
and type(func) is not type))
for m_name, method in six.iteritems(methods):
# skip error check if the method was inherited
# from a parent class (which means it wasn't
# defined in this source file)
if inspect.getmodule(method) is not None:
continue
cls_errors.extend(handle_method(method, m_name, class_name))
return cls_errors
def get_updates(self, grads):
"""
From Pylearn2 (https://github.com/lisa-lab/pylearn2/blob/master/pylearn2/training_algorithms/learning_rule.py)
Implements momentum as described in Section 9 of
"A Practical Guide to Training Restricted Boltzmann Machines",
Geoffrey Hinton.
Parameters are updated by the formula:
inc := momentum * inc - learning_rate * d cost / d param
param := param + inc
Also has the option to implement Nesterov momentum (accelerated momentum), which works better in a lot of cases.
:param grads: OrderedDict
An OrderedDict of (parameter, gradient) for the model's gradients
:return: OrderedDict
Updates at each training step
"""
log.debug(
'Setting up Stochastic Gradient Descent with momentum for optimizer...'
)
updates = OrderedDict()
for (param, gradient) in six.iteritems(grads):
vel = sharedX(param.get_value() * 0.)
assert param.dtype == vel.dtype
assert gradient.dtype == param.dtype
if param.name is not None:
vel.name = 'vel_' + param.name
scaled_lr = self.learning_rate * self.lr_scalers.get(param, 1.)
updates[vel] = self.momentum * vel - scaled_lr * gradient
inc = updates[vel]
if self.nesterov_momentum:
log.debug('Using Nesterov momentum')
inc = self.momentum * inc - scaled_lr * gradient
assert inc.dtype == vel.dtype
updates[param] = param + inc
return updates
def get_updates(self, gradients):
"""
This returns the parameter updates to use during training. It defaults to only using (annealed) learning rate.
Parameters
----------
gradients : dict
A dictionary mapping from the model's parameters to their gradients.
Returns
-------
updates : OrderdDict
A dictionary mapping from the old model parameters, to their new
values after a single iteration of the learning rule.
"""
log.debug('Setting up Stochastic Gradient Descent for optimizer...')
updates = OrderedDict()
for (param, gradient) in six.iteritems(gradients):
scaled_lr = self.learning_rate * self.lr_scalers.get(param, 1.)
updates[param] = param - scaled_lr * gradient
return updates
def get_updates(self, gradients):
"""
This returns the parameter updates to use during training. It defaults to only using (annealed) learning rate.
Parameters
----------
gradients : dict
A dictionary mapping from the model's parameters to their gradients.
Returns
-------
updates : OrderdDict
A dictionary mapping from the old model parameters, to their new
values after a single iteration of the learning rule.
"""
log.debug('Setting up Stochastic Gradient Descent for optimizer...')
updates = OrderedDict()
for (param, gradient) in six.iteritems(gradients):
scaled_lr = self.learning_rate * self.lr_scalers.get(param, 1.)
updates[param] = param - scaled_lr * gradient
return updates
def handle_class(val, class_name):
cls_errors = []
docstring = inspect.getdoc(val)
if docstring is None:
cls_errors.append((class_name,
'**missing** class-level docstring'))
else:
cls_errors = [
(e,) for e in
NumpyClassDocString(docstring, class_name, val).get_errors()
]
# Get public methods and parse their docstrings
methods = dict(((name, func) for name, func in inspect.getmembers(val)
if not name.startswith('_') and callable(func) and type(func) is not type))
for m_name, method in six.iteritems(methods):
# skip error check if the method was inherited
# from a parent class (which means it wasn't
# defined in this source file)
if inspect.getmodule(method) is not None:
continue
cls_errors.extend(handle_method(method, m_name, class_name))
return cls_errors
def _instantiate_proxy_tuple(proxy, bindings=None):
"""
Helper function for `_instantiate` that handles objects of the `Proxy`
class.
Parameters
----------
proxy : Proxy object
A `Proxy` object that.
bindings : dict, opitonal
A dictionary mapping previously instantiated `Proxy` objects
to their instantiated values.
Returns
-------
obj : object
The result object from recursively instantiating the object DAG.
"""
if proxy in bindings:
return bindings[proxy]
else:
# Respect do_not_recurse by just un-packing it (same as calling).
if proxy.callable == do_not_recurse:
obj = proxy.keywords['value']
else:
# TODO: add (requested) support for positionals (needs to be added
# to checked_call also).
if len(proxy.positionals) > 0:
raise NotImplementedError('positional arguments not yet '
'supported in proxy instantiation')
kwargs = dict((k, _instantiate(v, bindings))
for k, v in six.iteritems(proxy.keywords))
obj = checked_call(proxy.callable, kwargs)
try:
obj.yaml_src = proxy.yaml_src
except AttributeError: # Some classes won't allow this.
pass
bindings[proxy] = obj
return bindings[proxy]
def _instantiate(proxy, bindings=None):
"""
Instantiate a (hierarchy of) Proxy object(s).
Parameters
----------
proxy : object
A `Proxy` object or list/dict/literal. Strings are run through
`preprocess`.
bindings : dict, opitonal
A dictionary mapping previously instantiated `Proxy` objects
to their instantiated values.
Returns
-------
obj : object
The result object from recursively instantiating the object DAG.
Notes
-----
This should not be considered part of the stable, public API.
"""
if bindings is None:
bindings = {}
if isinstance(proxy, Proxy):
return _instantiate_proxy_tuple(proxy, bindings)
elif isinstance(proxy, dict):
# Recurse on the keys too, for backward compatibility.
# Is the key instantiation feature ever actually used, by anyone?
return dict((_instantiate(k, bindings), _instantiate(v, bindings))
for k, v in six.iteritems(proxy))
elif isinstance(proxy, list):
return [_instantiate(v, bindings) for v in proxy]
# In the future it might be good to consider a dict argument that provides
# a type->callable mapping for arbitrary transformations like this.
elif isinstance(proxy, six.string_types):
return preprocess(proxy)
else:
return proxy
def _instantiate(proxy, bindings=None):
"""
Instantiate a (hierarchy of) Proxy object(s).
Parameters
----------
proxy : object
A `Proxy` object or list/dict/literal. Strings are run through
`preprocess`.
bindings : dict, opitonal
A dictionary mapping previously instantiated `Proxy` objects
to their instantiated values.
Returns
-------
obj : object
The result object from recursively instantiating the object DAG.
Notes
-----
This should not be considered part of the stable, public API.
"""
if bindings is None:
bindings = {}
if isinstance(proxy, Proxy):
return _instantiate_proxy_tuple(proxy, bindings)
elif isinstance(proxy, dict):
# Recurse on the keys too, for backward compatibility.
# Is the key instantiation feature ever actually used, by anyone?
return dict((_instantiate(k, bindings), _instantiate(v, bindings))
for k, v in six.iteritems(proxy))
elif isinstance(proxy, list):
return [_instantiate(v, bindings) for v in proxy]
# In the future it might be good to consider a dict argument that provides
# a type->callable mapping for arbitrary transformations like this.
elif isinstance(proxy, six.string_types):
return preprocess(proxy)
else:
return proxy
def _instantiate_proxy_tuple(proxy, bindings=None):
"""
Helper function for `_instantiate` that handles objects of the `Proxy`
class.
Parameters
----------
proxy : Proxy object
A `Proxy` object that.
bindings : dict, opitonal
A dictionary mapping previously instantiated `Proxy` objects
to their instantiated values.
Returns
-------
obj : object
The result object from recursively instantiating the object DAG.
"""
if proxy in bindings:
return bindings[proxy]
else:
# Respect do_not_recurse by just un-packing it (same as calling).
if proxy.callable == do_not_recurse:
obj = proxy.keywords['value']
else:
# TODO: add (requested) support for positionals (needs to be added
# to checked_call also).
if len(proxy.positionals) > 0:
raise NotImplementedError('positional arguments not yet '
'supported in proxy instantiation')
kwargs = dict((k, _instantiate(v, bindings))
for k, v in six.iteritems(proxy.keywords))
obj = checked_call(proxy.callable, kwargs)
try:
obj.yaml_src = proxy.yaml_src
except AttributeError: # Some classes won't allow this.
pass
bindings[proxy] = obj
return bindings[proxy]
def get_updates(self, learning_rate, grads, lr_scalers=None):
"""
Provides the updates for learning with gradient descent + momentum.
Parameters
----------
learning_rate : float
Learning rate coefficient.
grads : dict
A dictionary mapping from the model's parameters to their
gradients.
lr_scalers : dict
A dictionary mapping from the model's parameters to a learning
rate multiplier.
"""
updates = OrderedDict()
for (param, grad) in six.iteritems(grads):
vel = sharedX(param.get_value() * 0.)
assert param.dtype == vel.dtype
assert grad.dtype == param.dtype
if param.name is not None:
vel.name = 'vel_' + param.name
scaled_lr = learning_rate * lr_scalers.get(param, 1.)
updates[vel] = self.momentum * vel - scaled_lr * grad
inc = updates[vel]
if self.nesterov_momentum:
inc = self.momentum * inc - scaled_lr * grad
assert inc.dtype == vel.dtype
updates[param] = param + inc
return updates
