def __enter__(self):
function_hooks = chainer.get_function_hooks()
if self.name in function_hooks:
raise KeyError("hook %s already exists" % self.name)
function_hooks[self.name] = self
return self
def __enter__(self):
function_hooks = chainer.get_function_hooks()
if self.name in function_hooks:
raise KeyError('hook %s already exists' % self.name)
function_hooks[self.name] = self
self.added(None)
return self
def _backward_main(self, retain_grad, loss_scale):
self._node._check_old_style_gradient()
if self.creator_node is None:
return
initial_device = None
if cuda.available and isinstance(self.data, cuda.ndarray):
try:
initial_device = cuda.Device()
except cuda.cupy.cuda.runtime.CUDARuntimeError as e:
if e.status != 38: # cudaErrorNoDevice
raise
is_debug = chainer.is_debug()
cand_funcs = []
seen_set = set()
grads = _backprop_utils.GradTable(load_if_new=True)
# Initialize error by 1, if this is a loss variable
if self.data.size == 1 and self._grad_var is None:
if self.data.ndim != 0:
warnings.warn(
'Treating a scalar as a variable with only one element'
' in Variable.backward is deprecated. A scalar variable'
' must be a 0-dimensional array. Apply'
' chainer.functions.squeeze to obtain a scalar variable.'
' If the size of this variable accidentally becomes one,'
' set zero to grad.',
DeprecationWarning)
with cuda.get_device_from_array(self.data) as device:
if device is cuda.DummyDevice:
self.grad = numpy.ones_like(self.data)
else:
self.grad = cuda.cupy.ones_like(self.data)
if loss_scale is not None:
self.grad *= loss_scale
grads[self._node] = self._grad_var
def add_cand(cand):
if cand not in seen_set:
# Negate since heapq is min-heap
heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand))
seen_set.add(cand)
add_cand(self.creator_node)
leaf_nodes = set()
while cand_funcs:
_, _, func = heapq.heappop(cand_funcs)
inputs = func.inputs
target_input_indexes = tuple([
i for i, x in enumerate(inputs) if x.requires_grad
])
outputs = [y() for y in func.outputs] # access via weak ref
out_grad = tuple([grads.pop(y) for y in outputs])
if not target_input_indexes:
continue
in_data = tuple([x.data for x in inputs])
out_grad_data = tuple(
[None if g is None else g.data for g in out_grad])
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
hooks = hooks.values() # avoid six for performance
cuda.get_device_from_array(*(in_data + out_grad_data)).use()
for hook in hooks:
hook.backward_preprocess(func, in_data, out_grad_data)
# Collect the current input gradients.
target_inputs = [inputs[i] for i in target_input_indexes]
# Keep the order for the portability, rather than
# in_grad = {x: grads.get_as_list(x) for x in set(target_inputs)}
in_grad = collections.OrderedDict()
for x in target_inputs:
if x not in in_grad:
in_grad[x] = grads.get_as_list(x)
_backprop_utils.backprop_step(
func, target_input_indexes, out_grad, in_grad)
for hook in hooks:
hook.backward_postprocess(func, in_data, out_grad_data)
if is_debug:
# each grad is a list of variables
# iter_gxs expands it as a sequence of variables.
def iter_gxs(gxs):
for gx in gxs:
for gx_elem in gx:
yield gx_elem
for gx in iter_gxs(in_grad.values()):
gx_data = gx.data
if gx_data.dtype.kind == 'f':
cuda.get_device_from_array(gx_data).use()
if cuda.get_array_module(gx_data).isnan(gx_data).any():
raise RuntimeError(
'NaN is detected on backward computation of '
'{}'.format(func.label))
for y, gy in six.moves.zip(outputs, out_grad):
if y is not None and y is not self.node:
y_var = y.get_variable_or_none()
if y_var is not None:
y_var._grad_var = gy if retain_grad else None
for x, gx in in_grad.items():
if not gx: # gradient == None
continue
for gx_elem in gx:
_check_grad_type(func, x, gx_elem.data)
if x.creator_node is None: # leaf
leaf_nodes.add(x)
else:
add_cand(x.creator_node)
del in_grad # to reduce memory usage
if initial_device is not None:
initial_device.use()
for x in leaf_nodes:
x_var = x.get_variable_or_none()
gx = grads.pop(x)
if x_var is not None:
x_var._grad_var = gx
x_var._loss_scale = loss_scale
grads.assert_no_grads()
def __call__(self, *inputs):
"""Applies forward propagation with chaining backward references.
Basic behavior is expressed in documentation of :class:`Function`
class.
.. note::
If the :data:`~Variable.data` attribute of input variables exist on
GPU device, then, before it calls :meth:`forward` method, the
appropriate device is selected, so in most cases implementers do
not need to take care of device selection.
Args:
inputs: Tuple of input :class:`Variable` objects. The volatile
flags of all input variables must agree.
Returns:
One :class:`Variable` object or a tuple of multiple
:class:`Variable` objects.
"""
in_data = tuple([x.data for x in inputs])
if chainer.is_debug():
self._stack = traceback.extract_stack()
if self.type_check_enable:
self._check_data_type_forward(in_data)
hooks = collections.OrderedDict(chainer.get_function_hooks())
hooks.update(self.local_function_hooks)
for hook in six.itervalues(hooks):
hook.forward_preprocess(self, in_data)
# Forward prop
with cuda.get_device(*in_data):
outputs = self.forward(in_data)
assert type(outputs) == tuple
for hook in six.itervalues(hooks):
hook.forward_postprocess(self, in_data)
if chainer.is_debug():
if any(out.dtype.kind == "f" and cuda.get_array_module(out).isnan(out).any() for out in outputs):
msg = "NaN is detected on forward computation"
raise RuntimeError(msg)
out_v = flag.aggregate_flags([x.volatile for x in inputs])
ret = tuple([variable.Variable(y, volatile=out_v) for y in outputs])
if out_v != "on":
# Topological ordering
self.rank = max([x.rank for x in inputs]) if inputs else 0
# Backward edges
for y in ret:
y.set_creator(self)
self.inputs = inputs
# Forward edges (must be weak references)
self.outputs = tuple([weakref.ref(y) for y in ret])
if len(ret) == 1:
return ret[0]
else:
return ret
def __exit__(self, exc_type, exc_value, traceback):
del chainer.get_function_hooks()[self.name]
def backward(self, retain_grad=False):
"""Runs error backpropagation (a.k.a. backprop) from this variable.
On backprop, :meth:`Function.backward` is called on each
:class:`Function` object appearing in the backward graph starting from
this variable. The backward graph is represented by backward references
from variables to their creators, and from functions to their inputs.
The backprop stops at all root variables. Some functions set ``None``
as gradients of some inputs, where further backprop does not take place
at such input variables.
This method uses :data:`grad` as the initial error array. User can
manually set a gradient array before calling this method. If
:data:`data` contains only one element (i.e., it is scalar) and
:data:`grad` is None, then this method automatically complements 1.0 as
the initial error. This is useful on starting backprop from some scalar
loss value.
Args:
retain_grad (bool): If True, the gradient arrays of all
intermediate variables are kept. Otherwise, :data:`grad` of the
intermediate variables are set to ``None`` on appropriate
timing, which may reduce the maximum memory consumption.
In most cases of training some model, the purpose of backprop
is to compute gradients of parameters, not of variables, so it
is recommended to set this flag False.
"""
if self.creator is None:
return
cand_funcs = []
seen_set = set()
seen_vars = set()
need_copy = set()
# Initialize error by 1, if this is a loss variable
if self.data.size == 1 and self.grad is None:
with cuda.get_device(self.data) as device:
if device is cuda.DummyDevice:
self.grad = numpy.ones_like(self.data)
else:
self.grad = cuda.cupy.ones_like(self.data)
def add_cand(cand):
if cand not in seen_set:
# Negate since heapq is min-heap
heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand))
seen_set.add(cand)
if self.creator is not None:
add_cand(self.creator)
while cand_funcs:
_, _, func = heapq.heappop(cand_funcs)
outputs = tuple(y() for y in func.outputs) # access via weak ref
in_data = tuple(x.data for x in func.inputs)
out_grad = tuple(None if y is None else y.grad for y in outputs)
hooks = collections.OrderedDict(chainer.get_function_hooks())
hooks.update(func.local_function_hooks)
for hook in six.itervalues(hooks):
hook.backward_preprocess(func, in_data, out_grad)
with cuda.get_device(*(in_data + out_grad)):
gxs = func.backward(in_data, out_grad)
assert len(gxs) == len(in_data)
for hook in six.itervalues(hooks):
hook.backward_postprocess(func, in_data, out_grad)
if chainer.is_debug():
if any(gx is not None
and cuda.get_array_module(gx).isnan(gx).any()
for gx in gxs):
msg = 'NaN is detected on backward computation'
raise RuntimeError(msg)
if not retain_grad:
for y in outputs:
if y is not None and y is not self:
y.grad = None
for x, gx in zip(func.inputs, gxs):
if gx is None:
continue
_check_grad_type(func, x, gx)
# Accumulate the gradient to x. It is a bit tricky to handle
# branches and parameter gradient accumulation correctly.
with cuda.get_device(gx):
id_x = id(x)
if x.creator is None: # leaf
if x._grad is None:
x.grad = gx
need_copy.add(id_x)
elif id_x in need_copy:
x.grad = x.grad + gx # copy
need_copy.remove(id_x)
else:
x._grad += gx
else: # not a leaf
add_cand(x.creator)
if id_x not in seen_vars: # 1st visit
x.grad = gx
seen_vars.add(id_x)
need_copy.add(id_x)
elif id_x in need_copy: # 2nd visit
x._grad = gx + x._grad # copied
need_copy.remove(id_x)
else: # 3rd or later visit
x._grad += gx
del gxs # to reduce memory usage
def __call__(self, *inputs):
"""Applies forward propagation with chaining backward references.
Basic behavior is expressed in documentation of :class:`Function`
class.
.. note::
If the :data:`~Variable.data` attribute of input variables exist on
GPU device, then, before it calls :meth:`forward` method, the
appropriate device is selected, so in most cases implementers do
not need to take care of device selection.
Args:
inputs: Tuple of input :class:`Variable`, :class:`numpy.ndarray` or
:class:`cupy.ndarray` objects. The volatile flags of all input
variables must agree. If the input is an :class:`numpy.ndarray`
or a :class:`cupy.ndarray`, it is automatically wrapped with
:class:`Variable`.
Returns:
One :class:`Variable` object or a tuple of multiple
:class:`Variable` objects.
"""
inputs = [
x if isinstance(x, chainer.Variable) else chainer.Variable(
x, volatile=flag.AUTO) for x in inputs
]
in_data = tuple([x.data for x in inputs])
if chainer.is_debug():
self._stack = traceback.extract_stack()
if self.type_check_enable:
self._check_data_type_forward(in_data)
hooks = collections.OrderedDict(chainer.get_function_hooks())
hooks.update(self.local_function_hooks)
for hook in six.itervalues(hooks):
hook.forward_preprocess(self, in_data)
# Forward prop
with cuda.get_device(*in_data):
outputs = self.forward(in_data)
assert type(outputs) == tuple
for hook in six.itervalues(hooks):
hook.forward_postprocess(self, in_data)
if chainer.is_debug():
if any(out.dtype.kind == 'f'
and cuda.get_array_module(out).isnan(out).any()
for out in outputs):
msg = 'NaN is detected on forward computation'
raise RuntimeError(msg)
out_v = flag.aggregate_flags([x.volatile for x in inputs])
ret = tuple([variable.Variable(y, volatile=out_v) for y in outputs])
if out_v != 'on':
# Topological ordering
self.rank = max([x.rank for x in inputs]) if inputs else 0
# Backward edges
for y in ret:
y.set_creator(self)
self.inputs = inputs
# Forward edges (must be weak references)
self.outputs = tuple([weakref.ref(y) for y in ret])
if len(ret) == 1:
return ret[0]
else:
return ret
def _backprop(outputs, inputs, grad_required, retain_grad, grads, loss_scale):
candidate_funcs, push_candidate, pop_candidate = _get_ordered_func_heap()
for y in outputs:
creator = y.creator_node
if creator is not None:
push_candidate(creator)
input_nodes = set(x.node for x in inputs)
ret_dict = {}
while candidate_funcs:
func = pop_candidate()
# Collect the gradients w.r.t. the outputs
ys = [y() for y in func.outputs] # access via weak ref
gys = tuple([grads.pop(y) for y in ys])
for node, gy in six.moves.zip(ys, gys):
if node is not None:
if node in input_nodes:
ret_dict[node] = gy
if retain_grad:
y = node.get_variable_or_none()
if y is not None:
y.grad_var = gy
y._loss_scale = loss_scale
# Collect the gradients w.r.t. the inputs
input_indexes = []
x_grads = collections.OrderedDict()
for i, x in enumerate(func.inputs):
if x not in grad_required:
continue
input_indexes.append(i)
if x not in x_grads:
x_grads[x] = grads.get_as_list(x)
if not input_indexes:
continue
input_indexes = tuple(input_indexes)
# Do backward
# Call pre-backward hooks
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
hooks = hooks.values() # avoid six for performance
in_data = tuple([x.data for x in func.inputs])
out_grad_data = tuple(
[None if g is None else g.data for g in gys])
with cuda.get_device_from_array(*in_data):
for hook in hooks:
hook.backward_preprocess(func, in_data, out_grad_data)
_backprop_utils.backprop_step(func, input_indexes, gys, x_grads)
# Call post-backward hooks
for hook in hooks:
hook.backward_postprocess(func, in_data, out_grad_data)
# Update grads
for node, g in x_grads.items():
if not g: # gradient == None
continue
creator = node.creator_node
if creator is not None:
push_candidate(creator)
for x in input_nodes:
if x not in ret_dict:
ret_dict[x] = grads.pop(x)
return ret_dict
def __exit__(self, *_):
chainer.get_function_hooks()[self.name].deleted(None)
del chainer.get_function_hooks()[self.name]
def apply(self, inputs):
"""Computes output variables and grows the computational graph.
Basic behavior is expressed in the documentation of
:class:`FunctionNode`.
.. note::
If the :data:`~Variable.data` attribute of input variables exist on
a GPU device, that device is made current before calling
:meth:`forward`, so implementors do not need to take care of device
selection in most cases.
Args:
inputs: Tuple of input variables. Each element can be either
:class:`~chainer.Variable`, :class:`numpy.ndarray`,
or :class:`cupy.ndarray`. If the element is an ndarray, it is
automatically wrapped with :class:`~chainer.Variable`.
Returns:
A tuple of output :class:`~chainer.Variable` objects.
"""
chainerx_in_data = None
chainerx_device = None
is_chainerx, in_data = _extract_apply_in_data(inputs)
if is_chainerx:
# Try ChainerX C++ implementation.
# If it's supported, the output arrays are wrapped with Variables
# and returned.
# If not supported, FunctionNode.forward_chainerx should return
# Fallback.
# In that case the input arrays are converted to numpy.ndarray
# or cupy.ndarray (depending on the ChainerX backend) and
# forward computation falls back to the conventional
# FunctionNode.forward() implementaion.
outputs = self.forward_chainerx(in_data)
if outputs is not chainer.Fallback:
# Supported. Wrap with variables and return
assert isinstance(outputs, tuple)
return tuple([
variable.Variable(
y, requires_grad=y.is_backprop_required())
for y in outputs])
# Fall back to FunctionNode.forward()
chainerx_in_data, in_data, chainerx_device = (
self._chainerx_apply_fallback_preprocess(in_data, inputs))
self._is_chainex_fallback_mode = True
self.chainerx_device = chainerx_device
utils._check_arrays_forward_compatible(in_data, self.label)
is_debug = chainer.is_debug()
if is_debug:
# Keep stack trace for debug
self.stack = traceback.extract_stack()
if configuration.config.type_check:
self._check_data_type_forward(in_data)
hooks = chainer.get_function_hooks()
if self._n_local_function_hooks > 0:
hooks = collections.OrderedDict(hooks)
hooks.update(self.local_function_hooks)
hooks = hooks.values() # avoid six for performance
for hook in hooks:
hook.forward_preprocess(self, in_data)
# Forward propagation
with cuda.get_device_from_array(*in_data):
self._input_indexes_to_retain = None
self._output_indexes_to_retain = None
if chainer.config.schedule_func is not None:
outputs = static_forward_optimizations(self, in_data)
elif self._is_chainex_fallback_mode:
# In ChainerX fallback, __class__ is temporarily replaced with
# the fabricated one with automatic attirbute fallback.
with _chainerx_attribute_fallback(self, chainerx_device):
outputs = self.forward(in_data)
else:
# In normal case, simply run the forward method.
outputs = self.forward(in_data)
# Check for output array types
if not isinstance(outputs, tuple):
raise TypeError(
'forward output must be a tuple ({})\n'
'Actual: {}'.format(self.label, type(outputs)))
if not chainer.is_arrays_compatible(outputs):
raise TypeError(
'incompatible array types are mixed in the forward output '
'({}).\n'
'Actual: {}'.format(
self.label,
', '.join(str(type(x)) for x in outputs)))
for hook in hooks:
hook.forward_postprocess(self, in_data)
# NaN check of output values
if is_debug:
if any(chainer.backend._contains_nan(out)
for out in outputs):
msg = ('NaN is detected on forward computation of '
'{}'.format(self.label))
raise RuntimeError(msg)
self._output_count = len(outputs)
if self._is_chainex_fallback_mode:
ret = self._chainerx_apply_fallback_postprocess(
chainerx_in_data, inputs, outputs)
else:
input_vars = [chainer.as_variable(x) for x in inputs]
requires_grad = any([x.requires_grad for x in input_vars])
ret = tuple(
[variable.Variable(y, requires_grad=requires_grad)
for y in outputs])
if configuration.config.enable_backprop:
# Topological ordering
self.rank = max(
[x.rank for x in input_vars]) if input_vars else 0
# Add backward edges
for y in ret:
y.creator_node = self
self.inputs = tuple([x.node for x in input_vars])
# Add forward edges (must be weak references)
self.outputs = tuple([weakref.ref(y.node) for y in ret])
if self._input_indexes_to_retain is not None:
for index in self._input_indexes_to_retain:
input_vars[index].retain_data()
if self._output_indexes_to_retain is not None:
retained_data = []
for index in self._output_indexes_to_retain:
ret[index].retain_data()
retained_data.append(outputs[index])
self._retained_output_data = tuple(retained_data)
self.lazy_grad_sum = configuration.config.lazy_grad_sum
return ret
def _backprop(outputs, inputs, grad_required, retain_grad, grads, loss_scale):
candidate_funcs, push_candidate, pop_candidate = _get_ordered_func_heap()
for y in outputs:
creator = y.creator_node
if creator is not None:
push_candidate(creator)
input_nodes = set(x.node for x in inputs)
while candidate_funcs:
func = pop_candidate()
# Collect the gradients w.r.t. the outputs
gys = []
for y_ref in func.outputs:
y = y_ref()
if y is None:
# output is not a part of the selected subgraph and has already
# been released.
gys.append(None)
continue
gys.append(grads.get(y, None))
gys = tuple(gys)
# Collect the gradients w.r.t. the inputs
#
# Note (Tokui): when the same variable is passed multiple times as
# inputs in the same function (e.g. an expression like f(x, x)), the
# current implementation passes None as the current gradient w.r.t.
# such an input except for the first one (i.e., it builds gxs like
# (gx, None) where gx is the current gradient w.r.t. x).
gxs = []
input_indexes = []
selected_inputs = set()
for i, x in enumerate(func.inputs):
if x not in grad_required:
continue
input_indexes.append(i)
if x in selected_inputs:
gxs.append(None)
else:
gxs.append(grads.get(x, None))
selected_inputs.add(x)
gxs = tuple(gxs)
input_indexes = tuple(input_indexes)
if not input_indexes:
continue
# Do backward
gys = tuple([
gy if not isinstance(gy, tuple) else chainer.functions.add(*gy)
for gy in gys
])
# Call pre-backward hooks
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
hooks = hooks.values() # avoid six for performance
in_data = tuple([x.data for x in func.inputs])
out_grad_data = tuple([None if g is None else g.data for g in gys])
cuda.get_device_from_array(*in_data).use()
for hook in hooks:
hook.backward_preprocess(func, in_data, out_grad_data)
new_gxs = func.backward_accumulate(input_indexes, gys, gxs)
# Call post-backward hooks
for hook in hooks:
hook.backward_postprocess(func, in_data, out_grad_data)
# Delete output gradients that are not required to return
for y_ref in func.outputs:
y = y_ref()
if y is not None and y in grads and y not in input_nodes:
del grads[y]
# Update grads
selected_inputs = set()
for i, g in zip(input_indexes, new_gxs):
if g is None:
continue
node = func.inputs[i]
if node in selected_inputs:
# Accumulate the duplicated gradients here
cur_gx = grads.get(node, None)
if cur_gx is not None:
if func.lazy_grad_sum:
if x.creator is None:
g = _backprop_utils.add(g, cur_gx)
else:
g = _backprop_utils.concat_variable(g, cur_gx)
# cur_gx can't be tuple, the lazy_grad_sum can't
# be enabled in its sibling node.
else:
g = g + cur_gx
else:
selected_inputs.add(node)
grads[node] = g
if retain_grad:
v = node.get_variable_or_none()
if v is not None:
v.grad_var = g
v._loss_scale = loss_scale
creator = node.creator_node
if creator is not None:
push_candidate(creator)
def apply(self, inputs):
"""Computes output variables and grows the computational graph.
Basic behavior is expressed in the documentation of
:class:`FunctionNode`.
.. note::
If the :data:`~Variable.data` attribute of input variables exist on
a GPU device, that device is made current before calling
:meth:`forward`, so implementors do not need to take care of device
selection in most cases.
Args:
inputs: Tuple of input variables. Each element can be either
:class:`~chainer.Variable` or :ref:`ndarray`. If the element
is an ndarray, it is automatically wrapped with
:class:`~chainer.Variable`.
Returns:
A tuple of output :class:`~chainer.Variable` objects.
"""
input_vars = [chainer.as_variable(x) for x in inputs]
in_data = tuple([x.data for x in input_vars])
requires_grad = any([x.requires_grad for x in input_vars])
# Check for input array types
if not chainer.is_arrays_compatible(in_data):
raise TypeError(
'incompatible array types are mixed in the forward input '
'({}).\n'
'Actual: {}'.format(self.label,
', '.join(str(type(x)) for x in in_data)))
is_debug = chainer.is_debug()
if is_debug:
# Keep stack trace for debug
self.stack = traceback.extract_stack()
if configuration.config.type_check:
self._check_data_type_forward(in_data)
hooks = chainer.get_function_hooks()
if self._n_local_function_hooks > 0:
hooks = collections.OrderedDict(hooks)
hooks.update(self.local_function_hooks)
hooks = hooks.values() # avoid six for performance
for hook in hooks:
hook.forward_preprocess(self, in_data)
# Forward propagation
with cuda.get_device_from_array(*in_data):
self._input_indexes_to_retain = None
self._output_indexes_to_retain = None
if chainer.config.schedule_func is not None:
outputs = static_forward_optimizations(self, in_data)
else:
outputs = self.forward(in_data)
# Check for output array types
if not isinstance(outputs, tuple):
raise TypeError('forward output must be a tuple ({})\n'
'Actual: {}'.format(self.label, type(outputs)))
if not chainer.is_arrays_compatible(outputs):
raise TypeError(
'incompatible array types are mixed in the forward output '
'({}).\n'
'Actual: {}'.format(self.label,
', '.join(str(type(x)) for x in outputs)))
for hook in hooks:
hook.forward_postprocess(self, in_data)
# NaN check of output values
if is_debug:
if any(chainer.backend._contains_nan(out) for out in outputs):
msg = ('NaN is detected on forward computation of '
'{}'.format(self.label))
raise RuntimeError(msg)
ret = tuple([
variable.Variable(y, requires_grad=requires_grad) for y in outputs
])
if configuration.config.enable_backprop:
# Topological ordering
self.rank = max([x.rank for x in input_vars]) if input_vars else 0
# Add backward edges
for y in ret:
y.creator_node = self
self.inputs = tuple([x.node for x in input_vars])
# Add forward edges (must be weak references)
self.outputs = tuple([weakref.ref(y.node) for y in ret])
if self._input_indexes_to_retain is not None:
for index in self._input_indexes_to_retain:
input_vars[index].retain_data()
if self._output_indexes_to_retain is not None:
retained_data = []
for index in self._output_indexes_to_retain:
ret[index].retain_data()
retained_data.append(outputs[index])
self._retained_output_data = tuple(retained_data)
self.lazy_grad_sum = configuration.config.lazy_grad_sum
return ret
def _backprop_to_all(outputs, retain_grad, loss_scale):
"""Backprop to all input variables
Args:
outputs (list of tuple): each tuple is (y_node, y_grad_var).
y_grad_var should not be None.
retain_grad (bool): see docstring of Variable.backward
loss_scale (float): see docstring of Variable.backward
"""
OrderedDict = chainer.utils._collections.OrderedDict # fix py2 memory leak
cand_funcs = []
seen_set = set()
def add_cand(cand):
if cand not in seen_set:
# Negate since heapq is min-heap
heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand))
seen_set.add(cand)
grads = _backprop_utils.GradTable(accumulate_grad_inputs=True)
leaf_nodes = set()
for y, gy in outputs:
grads.accumulate(y, gy)
func = y.creator_node
if func is None: # leaf
leaf_nodes.add(y)
else:
add_cand(func)
# Fix F812 (Python 2)
y = None
del y
is_debug = chainer.is_debug()
base_hooks = chainer.get_function_hooks().values()
while cand_funcs:
_, _, func = heapq.heappop(cand_funcs)
inputs = func.inputs
target_input_indexes = tuple([
i for i, x in enumerate(inputs) if x.requires_grad
])
outputs = [y() for y in func.outputs] # access via weak ref
out_grad = tuple([grads.pop(y)
if y is not None and y.creator_node is not None
else None
for y in outputs])
if not target_input_indexes:
continue
in_data = [x.data for x in inputs]
out_grad_array = [None if g is None else g.raw_array for g in out_grad]
if func._n_local_function_hooks != 0:
local_hooks = collections.OrderedDict(chainer.get_function_hooks())
local_hooks.update(func.local_function_hooks)
hooks = local_hooks.values() # avoid six for performance
else:
hooks = base_hooks
with chainer.using_device(
backend.get_device_from_array(*(in_data + out_grad_array))):
for hook in hooks:
hook.backward_preprocess(
func, tuple(in_data), tuple(out_grad_array))
# Collect the current input gradients.
target_inputs = [inputs[i] for i in target_input_indexes]
# Keep the order for the portability, rather than
# in_grad = {x: grads.get_as_list(x)
# for x in set(target_inputs)}
in_grad = OrderedDict()
for x in target_inputs:
if x not in in_grad:
in_grad[x] = grads.get_as_list(x)
_backprop_utils.backprop_step(
func, target_input_indexes, out_grad, in_grad, is_debug)
for hook in hooks:
hook.backward_postprocess(
func, tuple(in_data), tuple(out_grad_array))
if retain_grad:
# The gradients of the outputs of `func` are final. Store them if
# retain_grad=True.
for y, gy in six.moves.zip(outputs, out_grad):
if y is not None:
y._set_grad_var_if_available(gy)
del gy # to reduce memory usage
del out_grad # to reduce memory usage
for x, gx in in_grad.items():
if not gx: # gradient == None
continue
for gx_elem in gx:
if gx_elem is not None:
chainer.variable._check_grad_type(
func, x, True, gx_elem.raw_array)
del gx_elem # to reduce memory usage
if x.creator_node is None: # leaf
leaf_nodes.add(x)
else:
add_cand(x.creator_node)
del gx, in_grad # to reduce memory usage
for x in leaf_nodes:
x_var = x.get_variable_or_none()
gx = grads.pop(x)
if x_var is not None:
x_var._set_grad_var_without_check(gx)
x_var._loss_scale = loss_scale
grads.assert_no_grads()
def backward(self, retain_grad=False):
"""Runs error backpropagation (a.k.a. backprop) from this variable.
On backprop, :meth:`FunctionNode.backward` is called on each
:class:`FunctionNode` object appearing in the backward graph starting
from this variable. The backward graph is represented by backward
references from variable nodes to their creators, and from function
nodes to their input variable nodes. The backprop stops at all root
nodes. Some function nodes set ``None`` as gradients of some inputs,
where further backprop does not take place at such inputs.
This method uses :data:`grad` as the initial error array. User can
manually set a gradient array before calling this method. If
:data:`data` contains only one element (i.e., it is scalar) and
:data:`grad` is ``None``, then this method automatically complements
1.0 as the initial error. This is useful on starting backprop from
some scalar loss value.
Note that this method does not support *differentiable backprop*. Use
:func:`grad` to compute the gradient of gradients.
Args:
retain_grad (bool): If ``True``, the gradient arrays of all
intermediate variables are kept. Otherwise, :data:`grad` of the
intermediate variables are set to ``None`` on appropriate
timing, which may reduce the maximum memory consumption.
In most cases of training some models, the purpose of backprop
is to compute gradients of parameters, not of all variables,
and therefore it is recommended to set this flag ``False``.
"""
self._node._check_old_style_gradient()
if self.creator_node is None:
return
initial_device = None
if cuda.available and isinstance(self.data, cuda.cupy.ndarray):
try:
initial_device = cuda.Device()
except cuda.cupy.cuda.runtime.CUDARuntimeError as e:
if e.status != 38: # cudaErrorNoDevice
raise
is_debug = chainer.is_debug()
cand_funcs = []
seen_set = set()
grads = {}
# Initialize error by 1, if this is a loss variable
if self.data.size == 1 and self._grad_var is None:
with cuda.get_device_from_array(self.data) as device:
if device is cuda.DummyDevice:
self.grad = numpy.ones_like(self.data)
else:
self.grad = cuda.cupy.ones_like(self.data)
grads[self._node] = self._grad_var
def add_cand(cand):
if cand not in seen_set:
# Negate since heapq is min-heap
heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand))
seen_set.add(cand)
add_cand(self.creator_node)
def get_grad(node):
if node is None:
return None
if node in grads:
return grads[node]
return node.grad_var
while cand_funcs:
_, _, func = heapq.heappop(cand_funcs)
inputs = func.inputs
outputs = [y() for y in func.outputs] # access via weak ref
in_data = tuple([x.data for x in inputs])
out_grad = tuple([get_grad(y) for y in outputs])
out_grad_data = tuple(
[None if g is None else g.data for g in out_grad])
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
hooks = hooks.values() # avoid six for performance
cuda.get_device_from_array(*in_data).use()
for hook in hooks:
hook.backward_preprocess(func, in_data, out_grad_data)
# Collect the current input gradients.
#
# Note (Tokui): When the same variable is passed to multiple input
# slots (e.g. an expression like ``f(x, x)``), it makes the
# gradient accumulation complicated since the back-propagated
# gradients w.r.t. the first and second argument should be
# accumulated to the current gradient w.r.t. the same variable.
# In this case, the current implementation passes the current
# gradient only to the first occurrence of the variable in the
# input tuple and passes ``None`` to the rest of the occurrences.
# For example, when the input variables are ``(x, x)``, the
# input gradient passed to the ``backward_accumulate`` method is
# ``(gx, None)`` where ``gx`` is the current gradient of ``x``.
# See also the docstring of ``FunctionNode.backward_accumulate``.
target_input_indexes = [
i for i, x in enumerate(inputs) if x.requires_grad
]
target_inputs = [inputs[i] for i in target_input_indexes]
in_grad = []
for i, index_i in enumerate(target_input_indexes):
x = inputs[index_i]
if x in target_inputs[:i]:
# Pass ``None`` for duplicated input variables except for
# the first occurrence (see the comment above).
gx = None
elif x in grads:
gx = grads[x]
elif x.creator_node is None:
x._check_old_style_gradient()
# accumulate the gradient only if the node is a leaf
gx = x.grad_var
else:
gx = None
in_grad.append(gx)
gxs = func.backward_accumulate(target_input_indexes, out_grad,
in_grad)
assert len(gxs) == len(in_grad)
for hook in hooks:
hook.backward_postprocess(func, in_data, out_grad_data)
if is_debug:
for gx in gxs:
if gx is None:
continue
gx_data = gx.data
cuda.get_device_from_array(gx_data).use()
if cuda.get_array_module(gx_data).isnan(gx_data).any():
msg = 'NaN is detected on backward computation'
raise RuntimeError(msg)
if not retain_grad:
for y in outputs:
if y is not None and y is not self.node:
grads[y] = None
y_var = y.get_variable()
if y_var is not None:
y_var._grad_var = None
for i, gx in enumerate(gxs):
if gx is None:
continue
x = target_inputs[i]
if not x.requires_grad:
continue
_check_grad_type(func, x, gx.data)
if x in target_inputs[:i]:
# Accumulate the duplicated gradients here. See the comment
# above the code that builds ``in_grad``.
cur_gx = grads[x]
grads[x] = gx if cur_gx is None else gx + cur_gx
else:
grads[x] = gx
x_var = x.get_variable()
if x_var is not None:
x_var._grad_var = grads[x]
if x.creator_node is not None:
add_cand(x.creator_node)
del gxs # to reduce memory usage
if initial_device is not None:
initial_device.use()
def __exit__(self, exc_type, exc_value, traceback):
del chainer.get_function_hooks()[self.name]
def apply(self, inputs):
"""Computes output variables and grows the computational graph.
Basic behavior is expressed in the documentation of
:class:`FunctionNode`.
.. note::
If the :data:`~Variable.data` attribute of input variables exist on
a GPU device, that device is made current before calling
:meth:`forward`, so implementors do not need to take care of device
selection in most cases.
Args:
inputs: Tuple of input variables. Each element can be either
:class:`~chainer.Variable`, :class:`numpy.ndarray`,
or :class:`cupy.ndarray`. If the element is an ndarray, it is
automatically wrapped with :class:`~chainer.Variable`.
Returns:
A tuple of output :class:`~chainer.Variable` objects.
"""
input_vars = [chainer.as_variable(x) for x in inputs]
in_data = tuple([x.data for x in input_vars])
requires_grad = any([x.requires_grad for x in input_vars])
# Check for input array types
if not chainer.is_arrays_compatible(in_data):
raise TypeError(
'incompatible array types are mixed in the forward input '
'({}).\n'
'Actual: {}'.format(
self.label,
', '.join(str(type(x)) for x in in_data)))
is_debug = chainer.is_debug()
if is_debug:
# Keep stack trace for debug
self.stack = traceback.extract_stack()
if configuration.config.type_check:
self._check_data_type_forward(in_data)
hooks = chainer.get_function_hooks()
if self._n_local_function_hooks > 0:
hooks = collections.OrderedDict(hooks)
hooks.update(self.local_function_hooks)
hooks = hooks.values() # avoid six for performance
for hook in hooks:
hook.forward_preprocess(self, in_data)
# Forward propagation
with cuda.get_device_from_array(*in_data):
self._input_indexes_to_retain = None
self._output_indexes_to_retain = None
if chainer.config.schedule_func is not None:
outputs = static_forward_optimizations(self, in_data)
else:
outputs = self.forward(in_data)
# Check for output array types
if not isinstance(outputs, tuple):
raise TypeError(
'forward output must be a tuple ({})\n'
'Actual: {}'.format(self.label, type(outputs)))
if not chainer.is_arrays_compatible(outputs):
raise TypeError(
'incompatible array types are mixed in the forward output '
'({}).\n'
'Actual: {}'.format(
self.label,
', '.join(str(type(x)) for x in outputs)))
for hook in hooks:
hook.forward_postprocess(self, in_data)
# NaN check of output values
if is_debug:
if any(chainer.backend._contains_nan(out)
for out in outputs):
msg = ('NaN is detected on forward computation of '
'{}'.format(self.label))
raise RuntimeError(msg)
ret = tuple([variable.Variable(y, requires_grad=requires_grad)
for y in outputs])
if configuration.config.enable_backprop:
# Topological ordering
self.rank = max([x.rank for x in input_vars]) if input_vars else 0
# Add backward edges
for y in ret:
y.creator_node = self
self.inputs = tuple([x.node for x in input_vars])
# Add forward edges (must be weak references)
self.outputs = tuple([weakref.ref(y.node) for y in ret])
if self._input_indexes_to_retain is not None:
for index in self._input_indexes_to_retain:
input_vars[index].retain_data()
if self._output_indexes_to_retain is not None:
retained_data = []
for index in self._output_indexes_to_retain:
ret[index].retain_data()
retained_data.append(outputs[index])
self._retained_output_data = tuple(retained_data)
self.lazy_grad_sum = configuration.config.lazy_grad_sum
return ret
def backward(self, retain_grad=False):
"""Runs error backpropagation (a.k.a. backprop) from this variable.
On backprop, :meth:`Function.backward` is called on each
:class:`Function` object appearing in the backward graph starting from
this variable. The backward graph is represented by backward references
from variables to their creators, and from functions to their inputs.
The backprop stops at all root variables. Some functions set ``None``
as gradients of some inputs, where further backprop does not take place
at such input variables.
This method uses :data:`grad` as the initial error array. User can
manually set a gradient array before calling this method. If
:data:`data` contains only one element (i.e., it is scalar) and
:data:`grad` is ``None``, then this method automatically complements
1.0 as the initial error. This is useful on starting backprop from
some scalar loss value.
Args:
retain_grad (bool): If ``True``, the gradient arrays of all
intermediate variables are kept. Otherwise, :data:`grad` of the
intermediate variables are set to ``None`` on appropriate
timing, which may reduce the maximum memory consumption.
In most cases of training some models, the purpose of backprop
is to compute gradients of parameters, not of variables, so it
is recommended to set this flag ``False``.
"""
if self.creator is None:
return
initial_device = None
if cuda.available:
try:
initial_device = cuda.Device()
except cuda.cupy.cuda.runtime.CUDARuntimeError as e:
if e.status != 38: # cudaErrorNoDevice
raise
is_debug = chainer.is_debug()
cand_funcs = []
seen_set = set()
seen_vars = set()
need_copy = set()
# Initialize error by 1, if this is a loss variable
if self.data.size == 1 and self.grad is None:
with cuda.get_device(self.data) as device:
if device is cuda.DummyDevice:
self.grad = numpy.ones_like(self.data)
else:
self.grad = cuda.cupy.ones_like(self.data)
def add_cand(cand):
if cand not in seen_set:
# Negate since heapq is min-heap
heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand))
seen_set.add(cand)
add_cand(self.creator)
while cand_funcs:
_, _, func = heapq.heappop(cand_funcs)
outputs = [y() for y in func.outputs] # access via weak ref
in_data = tuple([x.data for x in func.inputs])
out_grad = ()
# if enable grad accumulate
if mkld.enable_acc_gradF((in_data,)) and in_data[0].ndim == 4 and all(isinstance(xi, numpy.ndarray) for xi in in_data):
out_grad_tmp = tuple([None if y is None else y.grad for y in outputs])
acc_grad_tuple = tuple([None if y is None else y.acc_grad for y in outputs])
for grad_tmp, acc_grad in zip(out_grad_tmp, acc_grad_tuple):
if len(acc_grad) == 0:
# no need accumulate, just return grad
out_grad += (grad_tmp,)
else:
"""
acc_grad's length is not 0, means need to do grad accumulate
call native MKLDNN sum primitive
"""
y = numpy.empty((grad_tmp.shape), dtype=grad_tmp.dtype)
acc_grad += (grad_tmp,)
mkldnn_sum = mkldnn.Sum_F32()
mkldnn_sum.sum4d_gx(acc_grad, y)
out_grad += (y,)
else:
out_grad = tuple([None if y is None else y.grad for y in outputs])
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
cuda.get_device(*(in_data + out_grad)).use()
for hook in six.itervalues(hooks):
hook.backward_preprocess(func, in_data, out_grad)
if isinstance(func, chainer.functions.connection.convolution_2d.Convolution2DFunction):
_x = func.inputs[0]
if _x.creator is None and func.in_chain is True:
func.mkldnn_opt = True
cosim_output = func.backward_cpu_cosim(in_data, out_grad)
gxs = func.backward(in_data, out_grad)
assert len(gxs) == len(in_data)
func.cpu_cosim_verify_result(gxs, cosim_output)
for hook in six.itervalues(hooks):
hook.backward_postprocess(func, in_data, out_grad)
if is_debug:
for gx in gxs:
if gx is None:
continue
cuda.get_device(gx).use()
if cuda.get_array_module(gx).isnan(gx).any():
msg = 'NaN is detected on backward computation'
raise RuntimeError(msg)
if not retain_grad:
for y in outputs:
if y is not None and y is not self:
y.grad = None
for x, gx in zip(func.inputs, gxs):
if gx is None:
continue
_check_grad_type(func, x, gx)
# Accumulate the gradient to x. It is a bit tricky to handle
# branches and parameter gradient accumulation correctly.
id_x = id(x)
if x.creator is None: # leaf
if x._grad is None: # 1st visit
x.grad = gx
need_copy.add(id_x)
else:
cuda.get_device(gx).use()
if id_x in need_copy: # 2nd visit
if mkld.enable_acc_gradF((in_data,)) and in_data[0].ndim == 4 and all(isinstance(xi, numpy.ndarray) for xi in in_data):
# if enable_acc_grad,will deply to do grad accumulate,only record grad
x.acc_grad += (gx,)
else:
x.grad = utils.force_array(x.grad + gx) # copy
need_copy.remove(id_x) # remove from list in 2nd visit
else:
if mkld.enable_acc_gradF((in_data,)) and in_data[0].ndim == 4 and all(isinstance(xi, numpy.ndarray) for xi in in_data):
# if enable_acc_grad, will deply to do grad accumulate, only record grad
if len(x.acc_grad) > 0: # means 3rd or later visit for variable x
x.acc_grad += (gx,)
else: # means this variable is W or b
x._grad += gx
else:
x._grad += gx # 3rd or later visit
else: # not a leaf
add_cand(x.creator)
if id_x not in seen_vars: # 1st visit
x.grad = gx
seen_vars.add(id_x)
need_copy.add(id_x)
else:
cuda.get_device(gx).use()
if id_x in need_copy: # 2nd visit
if mkld.enable_acc_gradF((in_data,)) and in_data[0].ndim == 4 and all(isinstance(xi, numpy.ndarray) for xi in in_data):
# if enable_acc_grad, will deply to do grad accumulate, only record grad
x.acc_grad += (gx,)
else:
x._grad = utils.force_array(gx + x._grad) # copied
need_copy.remove(id_x)
else: # 3rd or later visit
if mkld.enable_acc_gradF((in_data,)) and in_data[0].ndim == 4 and all(isinstance(xi, numpy.ndarray) for xi in in_data):
# if enable_acc_grad, will deply to do grad accumulate, only record grad
x.acc_grad += (gx,)
else:
x._grad += gx
del gxs # to reduce memory usage
if initial_device is not None:
initial_device.use()
def __exit__(self, *_):
del chainer.get_function_hooks()[self.name]
def _backprop(outputs, inputs, grad_required, retain_grad, grads, loss_scale):
candidate_funcs, push_candidate, pop_candidate = _get_ordered_func_heap()
for y in outputs:
creator = y.creator_node
if creator is not None:
push_candidate(creator)
input_nodes = set(x.node for x in inputs)
while candidate_funcs:
func = pop_candidate()
# Collect the gradients w.r.t. the outputs
gys = []
for y_ref in func.outputs:
y = y_ref()
if y is None:
# output is not a part of the selected subgraph and has already
# been released.
gys.append(None)
continue
gys.append(grads.get(y, None))
gys = tuple(gys)
# Collect the gradients w.r.t. the inputs
#
# Note (Tokui): when the same variable is passed multiple times as
# inputs in the same function (e.g. an expression like f(x, x)), the
# current implementation passes None as the current gradient w.r.t.
# such an input except for the first one (i.e., it builds gxs like
# (gx, None) where gx is the current gradient w.r.t. x).
gxs = []
input_indexes = []
selected_inputs = set()
for i, x in enumerate(func.inputs):
if x not in grad_required:
continue
input_indexes.append(i)
if x in selected_inputs:
gxs.append(None)
else:
gxs.append(grads.get(x, None))
selected_inputs.add(x)
gxs = tuple(gxs)
input_indexes = tuple(input_indexes)
if not input_indexes:
continue
# Do backward
gys = tuple([gy if not isinstance(gy, tuple) else
chainer.functions.add(*gy)
for gy in gys])
# Call pre-backward hooks
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
hooks = hooks.values() # avoid six for performance
in_data = tuple([x.data for x in func.inputs])
out_grad_data = tuple(
[None if g is None else g.data for g in gys])
cuda.get_device_from_array(*in_data).use()
for hook in hooks:
hook.backward_preprocess(func, in_data, out_grad_data)
new_gxs = func.backward_accumulate(input_indexes, gys, gxs)
# Call post-backward hooks
for hook in hooks:
hook.backward_postprocess(func, in_data, out_grad_data)
# Delete output gradients that are not required to return
for y_ref in func.outputs:
y = y_ref()
if y is not None and y in grads and y not in input_nodes:
del grads[y]
# Update grads
selected_inputs = set()
for i, g in zip(input_indexes, new_gxs):
if g is None:
continue
node = func.inputs[i]
if node in selected_inputs:
# Accumulate the duplicated gradients here
cur_gx = grads.get(node, None)
if cur_gx is not None:
if func.lazy_grad_sum:
if x.creator is None:
g = _backprop_utils.add(g, cur_gx)
else:
g = _backprop_utils.concat_variable(g, cur_gx)
# cur_gx can't be tuple, the lazy_grad_sum can't
# be enabled in its sibling node.
else:
g = g + cur_gx
else:
selected_inputs.add(node)
grads[node] = g
if retain_grad:
v = node.get_variable_or_none()
if v is not None:
v.grad_var = g
v._loss_scale = loss_scale
creator = node.creator_node
if creator is not None:
push_candidate(creator)
def _backward_main(self, retain_grad):
self._node._check_old_style_gradient()
if self.creator_node is None:
return
initial_device = None
if cuda.available and isinstance(self.data, cuda.cupy.ndarray):
try:
initial_device = cuda.Device()
except cuda.cupy.cuda.runtime.CUDARuntimeError as e:
if e.status != 38: # cudaErrorNoDevice
raise
is_debug = chainer.is_debug()
cand_funcs = []
seen_set = set()
grads = {}
# Initialize error by 1, if this is a loss variable
if self.data.size == 1 and self._grad_var is None:
with cuda.get_device_from_array(self.data) as device:
if device is cuda.DummyDevice:
self.grad = numpy.ones_like(self.data)
else:
self.grad = cuda.cupy.ones_like(self.data)
grads[self._node] = self._grad_var
def add_cand(cand):
if cand not in seen_set:
# Negate since heapq is min-heap
heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand))
seen_set.add(cand)
add_cand(self.creator_node)
def get_grad(node):
if node is None:
return None
if node in grads:
return grads[node]
return node.grad_var
while cand_funcs:
_, _, func = heapq.heappop(cand_funcs)
inputs = func.inputs
target_input_indexes = [
i for i, x in enumerate(inputs) if x.requires_grad
]
if not target_input_indexes:
continue
outputs = [y() for y in func.outputs] # access via weak ref
in_data = tuple([x.data for x in inputs])
out_grad = tuple([get_grad(y) for y in outputs])
out_grad_data = tuple(
[None if g is None else g.data for g in out_grad])
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
hooks = hooks.values() # avoid six for performance
cuda.get_device_from_array(*in_data).use()
for hook in hooks:
hook.backward_preprocess(func, in_data, out_grad_data)
# Collect the current input gradients.
#
# Note (Tokui): When the same variable is passed to multiple input
# slots (e.g. an expression like ``f(x, x)``), it makes the
# gradient accumulation complicated since the back-propagated
# gradients w.r.t. the first and second argument should be
# accumulated to the current gradient w.r.t. the same variable.
# In this case, the current implementation passes the current
# gradient only to the first occurrence of the variable in the
# input tuple and passes ``None`` to the rest of the occurrences.
# For example, when the input variables are ``(x, x)``, the
# input gradient passed to the ``backward_accumulate`` method is
# ``(gx, None)`` where ``gx`` is the current gradient of ``x``.
# See also the docstring of ``FunctionNode.backward_accumulate``.
target_inputs = [inputs[i] for i in target_input_indexes]
in_grad = []
for i, index_i in enumerate(target_input_indexes):
x = inputs[index_i]
if x in target_inputs[:i]:
# Pass ``None`` for duplicated input variables except for
# the first occurrence (see the comment above).
gx = None
elif x in grads:
gx = grads[x]
elif x.creator_node is None:
x._check_old_style_gradient()
# accumulate the gradient only if the node is a leaf
gx = x.grad_var
else:
gx = None
in_grad.append(gx)
gxs = func.backward_accumulate(target_input_indexes, out_grad,
in_grad)
assert len(gxs) == len(in_grad)
for hook in hooks:
hook.backward_postprocess(func, in_data, out_grad_data)
if is_debug:
for gx in gxs:
if gx is None:
continue
gx_data = gx.data
if gx_data.dtype.kind == 'f':
cuda.get_device_from_array(gx_data).use()
if cuda.get_array_module(gx_data).isnan(gx_data).any():
raise RuntimeError(
'NaN is detected on backward computation of '
'{}'.format(func.label))
if not retain_grad:
for y in outputs:
if y is not None and y is not self.node:
grads[y] = None
y_var = y.get_variable_or_none()
if y_var is not None:
y_var._grad_var = None
for i, gx in enumerate(gxs):
if gx is None:
continue
x = target_inputs[i]
if not x.requires_grad:
continue
_check_grad_type(func, x, gx.data)
if x in target_inputs[:i]:
# Accumulate the duplicated gradients here. See the comment
# above the code that builds ``in_grad``.
cur_gx = grads[x]
grads[x] = gx if cur_gx is None else gx + cur_gx
else:
grads[x] = gx
x_var = x.get_variable_or_none()
if x_var is not None:
x_var._grad_var = grads[x]
if x.creator_node is not None:
add_cand(x.creator_node)
del gxs # to reduce memory usage
if initial_device is not None:
initial_device.use()
def _backward_main(self, retain_grad, loss_scale):
self._node._check_old_style_gradient()
if self.creator_node is None:
return
cand_funcs = []
seen_set = set()
grads = _backprop_utils.GradTable(load_if_new=True)
# Initialize error by 1, if this is a loss variable
if self.data.size == 1 and self._grad_var is None:
if self.data.ndim != 0:
warnings.warn(
'Treating a scalar as a variable with only one element'
' in Variable.backward is deprecated. A scalar variable'
' must be a 0-dimensional array. Apply'
' chainer.functions.squeeze to obtain a scalar variable.'
' If the size of this variable accidentally becomes one,'
' set zero to grad.',
DeprecationWarning)
with cuda.get_device_from_array(self.data) as device:
if device is cuda.DummyDevice:
self.grad = numpy.ones_like(self.data)
else:
self.grad = cuda.cupy.ones_like(self.data)
if loss_scale is not None:
self.grad *= loss_scale
grads[self._node] = self._grad_var
def add_cand(cand):
if cand not in seen_set:
# Negate since heapq is min-heap
heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand))
seen_set.add(cand)
add_cand(self.creator_node)
leaf_nodes = set()
while cand_funcs:
_, _, func = heapq.heappop(cand_funcs)
inputs = func.inputs
target_input_indexes = tuple([
i for i, x in enumerate(inputs) if x.requires_grad
])
outputs = [y() for y in func.outputs] # access via weak ref
out_grad = tuple([grads.pop(y) for y in outputs])
if not target_input_indexes:
continue
in_data = tuple([x.data for x in inputs])
out_grad_data = tuple(
[None if g is None else g.data for g in out_grad])
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
hooks = hooks.values() # avoid six for performance
with cuda.get_device_from_array(*(in_data + out_grad_data)):
for hook in hooks:
hook.backward_preprocess(func, in_data, out_grad_data)
# Collect the current input gradients.
target_inputs = [inputs[i] for i in target_input_indexes]
# Keep the order for the portability, rather than
# in_grad = {x: grads.get_as_list(x)
# for x in set(target_inputs)}
in_grad = collections.OrderedDict()
for x in target_inputs:
if x not in in_grad:
in_grad[x] = grads.get_as_list(x)
# to reduce memory usage
x._set_grad_var_if_available(None)
_backprop_utils.backprop_step(
func, target_input_indexes, out_grad, in_grad)
for hook in hooks:
hook.backward_postprocess(func, in_data, out_grad_data)
for y, gy in six.moves.zip(outputs, out_grad):
if y is not None and y is not self.node:
y._set_grad_var_if_available(
gy if retain_grad else None)
del gy, out_grad # to reduce memory usage
for x, gx in in_grad.items():
if not gx: # gradient == None
continue
for gx_elem in gx:
_check_grad_type(func, x, gx_elem.data)
del gx_elem # to reduce memory usage
if x.creator_node is None: # leaf
leaf_nodes.add(x)
else:
add_cand(x.creator_node)
del gx, in_grad # to reduce memory usage
for x in leaf_nodes:
x_var = x.get_variable_or_none()
gx = grads.pop(x)
if x_var is not None:
x_var._grad_var = gx
x_var._loss_scale = loss_scale
grads.assert_no_grads()
def __exit__(self, *_):
chainer.get_function_hooks()[self.name].deleted()
del chainer.get_function_hooks()[self.name]
def __call__(self, *inputs):
"""Applies forward propagation with chaining backward references.
Basic behavior is expressed in documentation of :class:`Function`
class.
.. note::
If the :data:`~Variable.data` attribute of input variables exist on
GPU device, then, before it calls :meth:`forward` method, the
appropriate device is selected, so in most cases implementers do
not need to take care of device selection.
Args:
inputs: Tuple of input :class:`Variable`, :class:`numpy.ndarray` or
:class:`cupy.ndarray` objects.
If the input is an :class:`numpy.ndarray` or a
:class:`cupy.ndarray`, it is automatically wrapped with
:class:`Variable`.
Returns:
One :class:`Variable` object or a tuple of multiple
:class:`Variable` objects.
"""
inputs = [
x if isinstance(x, variable.Variable) else variable.Variable(
x, requires_grad=False) for x in inputs
]
in_data = tuple([x.data for x in inputs])
requires_grad = any([x.requires_grad for x in inputs])
if chainer.is_debug():
self._stack = traceback.extract_stack()
if configuration.config.type_check:
self._check_data_type_forward(in_data)
hooks = chainer.get_function_hooks()
if self._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(self.local_function_hooks)
for hook in six.itervalues(hooks):
hook.forward_preprocess(self, in_data)
# Forward prop
with cuda.get_device_from_array(*in_data):
self._input_indexes_to_retain = None
self._output_indexes_to_retain = None
outputs = self.forward(in_data)
assert type(outputs) == tuple
for hook in six.itervalues(hooks):
hook.forward_postprocess(self, in_data)
if chainer.is_debug():
if any(out.dtype.kind == 'f'
and cuda.get_array_module(out).isnan(out).any()
for out in outputs):
msg = 'NaN is detected on forward computation'
raise RuntimeError(msg)
ret = tuple([
variable.Variable(y, requires_grad=requires_grad) for y in outputs
])
if configuration.config.enable_backprop:
# Topological ordering
self.rank = max([x.rank for x in inputs]) if inputs else 0
# Backward edges
for y in ret:
y.set_creator(self)
self.inputs = tuple([x.node for x in inputs])
# Forward edges (must be weak references)
self.outputs = tuple([weakref.ref(y.node) for y in ret])
input_indexes_to_retain = self._input_indexes_to_retain
if input_indexes_to_retain is None:
# input arrays are retained by default
input_indexes_to_retain = six.moves.range(len(inputs))
for index in input_indexes_to_retain:
inputs[index].retain_data()
del self._input_indexes_to_retain
output_indexes_to_retain = self._output_indexes_to_retain
if output_indexes_to_retain is not None:
for index in output_indexes_to_retain:
ret[index].retain_data()
del self._output_indexes_to_retain
if len(ret) == 1:
return ret[0]
else:
return ret
def _backward_main(self, retain_grad):
self._node._check_old_style_gradient()
if self.creator_node is None:
return
initial_device = None
if cuda.available and isinstance(self.data, cuda.cupy.ndarray):
try:
initial_device = cuda.Device()
except cuda.cupy.cuda.runtime.CUDARuntimeError as e:
if e.status != 38: # cudaErrorNoDevice
raise
is_debug = chainer.is_debug()
cand_funcs = []
seen_set = set()
grads = {}
# Initialize error by 1, if this is a loss variable
if self.data.size == 1 and self._grad_var is None:
with cuda.get_device_from_array(self.data) as device:
if device is cuda.DummyDevice:
self.grad = numpy.ones_like(self.data)
else:
self.grad = cuda.cupy.ones_like(self.data)
grads[self._node] = self._grad_var
def add_cand(cand):
if cand not in seen_set:
# Negate since heapq is min-heap
heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand))
seen_set.add(cand)
add_cand(self.creator_node)
def get_grad(node):
if node is None:
return None
if node in grads:
return grads[node]
return node.grad_var
while cand_funcs:
_, _, func = heapq.heappop(cand_funcs)
inputs = func.inputs
target_input_indexes = [
i for i, x in enumerate(inputs) if x.requires_grad
]
if not target_input_indexes:
continue
outputs = [y() for y in func.outputs] # access via weak ref
in_data = tuple([x.data for x in inputs])
out_grad = tuple([get_grad(y) for y in outputs])
out_grad_data = tuple(
[None if g is None else g.data for g in out_grad])
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
hooks = hooks.values() # avoid six for performance
cuda.get_device_from_array(*in_data).use()
for hook in hooks:
hook.backward_preprocess(func, in_data, out_grad_data)
# Collect the current input gradients.
#
# Note (Tokui): When the same variable is passed to multiple input
# slots (e.g. an expression like ``f(x, x)``), it makes the
# gradient accumulation complicated since the back-propagated
# gradients w.r.t. the first and second argument should be
# accumulated to the current gradient w.r.t. the same variable.
# In this case, the current implementation passes the current
# gradient only to the first occurrence of the variable in the
# input tuple and passes ``None`` to the rest of the occurrences.
# For example, when the input variables are ``(x, x)``, the
# input gradient passed to the ``backward_accumulate`` method is
# ``(gx, None)`` where ``gx`` is the current gradient of ``x``.
# See also the docstring of ``FunctionNode.backward_accumulate``.
target_inputs = [inputs[i] for i in target_input_indexes]
in_grad = []
for i, index_i in enumerate(target_input_indexes):
x = inputs[index_i]
if x in target_inputs[:i]:
# Pass ``None`` for duplicated input variables except for
# the first occurrence (see the comment above).
gx = None
elif x in grads:
gx = grads[x]
elif x.creator_node is None:
x._check_old_style_gradient()
# accumulate the gradient only if the node is a leaf
gx = x.grad_var
else:
gx = None
in_grad.append(gx)
gxs = func.backward_accumulate(
target_input_indexes, out_grad, in_grad)
assert len(gxs) == len(in_grad)
for hook in hooks:
hook.backward_postprocess(func, in_data, out_grad_data)
if is_debug:
for gx in gxs:
if gx is None:
continue
gx_data = gx.data
if gx_data.dtype.kind == 'f':
cuda.get_device_from_array(gx_data).use()
if cuda.get_array_module(gx_data).isnan(gx_data).any():
raise RuntimeError(
'NaN is detected on backward computation of '
'{}'.format(func.label))
if not retain_grad:
for y in outputs:
if y is not None and y is not self.node:
grads[y] = None
y_var = y.get_variable_or_none()
if y_var is not None:
y_var._grad_var = None
for i, gx in enumerate(gxs):
if gx is None:
continue
x = target_inputs[i]
if not x.requires_grad:
continue
_check_grad_type(func, x, gx.data)
if x in target_inputs[:i]:
# Accumulate the duplicated gradients here. See the comment
# above the code that builds ``in_grad``.
cur_gx = grads[x]
grads[x] = gx if cur_gx is None else gx + cur_gx
else:
grads[x] = gx
x_var = x.get_variable_or_none()
if x_var is not None:
x_var._grad_var = grads[x]
if x.creator_node is not None:
add_cand(x.creator_node)
del gxs # to reduce memory usage
if initial_device is not None:
initial_device.use()
def apply(self, inputs):
"""Computes output variables and grows the computational graph.
Basic behavior is expressed in the documentation of
:class:`FunctionNode`.
.. note::
If the :data:`~Variable.data` attribute of input variables exist on
a GPU device, that device is made current before calling
:meth:`forward`, so implementors do not need to take care of device
selection in most cases.
Args:
inputs: Tuple of input variables. Each element can be either
:class:`Variable`, :class:`numpy.ndarray`,
or :class:`cupy.ndarray`. If the element is an ndarray, it is
automatically wrapped with :class:`Variable`.
Returns:
A tuple of output :class:`Variable` objects.
"""
input_vars = [chainer.as_variable(x) for x in inputs]
in_data = tuple([x.data for x in input_vars])
requires_grad = any([x.requires_grad for x in input_vars])
if chainer.is_debug():
self.stack = traceback.extract_stack()
if configuration.config.type_check:
self._check_data_type_forward(in_data)
hooks = chainer.get_function_hooks()
if self._n_local_function_hooks > 0:
hooks = collections.OrderedDict(hooks)
hooks.update(self.local_function_hooks)
hooks = hooks.values() # avoid six for performance
for hook in hooks:
hook.forward_preprocess(self, in_data)
# Forward propagation
with cuda.get_device_from_array(*in_data):
self._input_indexes_to_retain = None
self._output_indexes_to_retain = None
outputs = self.forward(in_data)
assert type(outputs) is tuple
for hook in hooks:
hook.forward_postprocess(self, in_data)
# NaN check of output values
if chainer.is_debug():
if any(out.dtype.kind == 'f'
and cuda.get_array_module(out).isnan(out).any()
for out in outputs):
msg = ('NaN is detected on forward computation of '
'{}'.format(self.label))
raise RuntimeError(msg)
ret = tuple([
variable.Variable(y, requires_grad=requires_grad) for y in outputs
])
if configuration.config.enable_backprop:
# Topological ordering
self.rank = max([x.rank for x in input_vars]) if input_vars else 0
# Add backward edges
for i, y in enumerate(ret):
y.creator_node = self
self.inputs = tuple([x.node for x in input_vars])
# Add forward edges (must be weak references)
self.outputs = tuple([weakref.ref(y.node) for y in ret])
if self._input_indexes_to_retain is not None:
for index in self._input_indexes_to_retain:
input_vars[index].retain_data()
if self._output_indexes_to_retain is not None:
retained_data = []
for index in self._output_indexes_to_retain:
ret[index].retain_data()
retained_data.append(outputs[index])
self._retained_output_data = tuple(retained_data)
return ret
def __exit__(self, *_):
del chainer.get_function_hooks()[self.name]
def _backprop(outputs, inputs, grad_required, retain_grad, grads, loss_scale):
candidate_funcs, push_candidate, pop_candidate = _get_ordered_func_heap()
for y in outputs:
creator = y.creator_node
if creator is not None:
push_candidate(creator)
input_nodes = set(x.node for x in inputs)
ret_dict = {}
while candidate_funcs:
func = pop_candidate()
# Collect the gradients w.r.t. the outputs
ys = [y() for y in func.outputs] # access via weak ref
gys = tuple([grads.pop(y) for y in ys])
for node, gy in six.moves.zip(ys, gys):
if node is not None:
if node in input_nodes:
ret_dict[node] = gy
if retain_grad:
y = node.get_variable_or_none()
if y is not None:
y.grad_var = gy
y._loss_scale = loss_scale
# Collect the gradients w.r.t. the inputs
input_indexes = []
x_grads = collections.OrderedDict()
for i, x in enumerate(func.inputs):
if x not in grad_required:
continue
input_indexes.append(i)
if x not in x_grads:
x_grads[x] = grads.get_as_list(x)
if not input_indexes:
continue
input_indexes = tuple(input_indexes)
# Do backward
# Call pre-backward hooks
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
hooks = hooks.values() # avoid six for performance
in_data = tuple([x.data for x in func.inputs])
out_grad_data = tuple([None if g is None else g.data for g in gys])
with cuda.get_device_from_array(*in_data):
for hook in hooks:
hook.backward_preprocess(func, in_data, out_grad_data)
_backprop_utils.backprop_step(func, input_indexes, gys, x_grads)
# Call post-backward hooks
for hook in hooks:
hook.backward_postprocess(func, in_data, out_grad_data)
# Update grads
for node, g in x_grads.items():
if not g: # gradient == None
continue
creator = node.creator_node
if creator is not None:
push_candidate(creator)
for x in input_nodes:
if x not in ret_dict:
ret_dict[x] = grads.pop(x)
return ret_dict
def _backward_main(self, retain_grad, loss_scale):
self._node._check_old_style_gradient()
if self.creator_node is None:
return
cand_funcs = []
seen_set = set()
grads = _backprop_utils.GradTable(load_if_new=True)
# Initialize error by 1, if this is a loss variable
if self.array.size == 1 and self._grad_var is None:
if self.array.ndim != 0:
warnings.warn(
'Treating a scalar as a variable with only one element'
' in Variable.backward is deprecated. A scalar variable'
' must be a 0-dimensional array. Apply'
' chainer.functions.squeeze to obtain a scalar variable.'
' If the size of this variable accidentally becomes one,'
' set zero to grad.',
DeprecationWarning)
with cuda.get_device_from_array(self.array) as device:
if device is cuda.DummyDevice:
self.grad = numpy.ones_like(self.array)
else:
self.grad = cuda.cupy.ones_like(self.array)
if loss_scale is not None:
self.grad *= loss_scale
grads[self._node] = self._grad_var
def add_cand(cand):
if cand not in seen_set:
# Negate since heapq is min-heap
heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand))
seen_set.add(cand)
add_cand(self.creator_node)
leaf_nodes = set()
while cand_funcs:
_, _, func = heapq.heappop(cand_funcs)
inputs = func.inputs
target_input_indexes = tuple([
i for i, x in enumerate(inputs) if x.requires_grad
])
outputs = [y() for y in func.outputs] # access via weak ref
out_grad = tuple([grads.pop(y) for y in outputs])
if not target_input_indexes:
continue
in_data = tuple([x.data for x in inputs])
out_grad_array = tuple(
[None if g is None else g.array for g in out_grad])
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
hooks = hooks.values() # avoid six for performance
with cuda.get_device_from_array(*(in_data + out_grad_array)):
for hook in hooks:
hook.backward_preprocess(func, in_data, out_grad_array)
# Collect the current input gradients.
target_inputs = [inputs[i] for i in target_input_indexes]
# Keep the order for the portability, rather than
# in_grad = {x: grads.get_as_list(x)
# for x in set(target_inputs)}
in_grad = collections.OrderedDict()
for x in target_inputs:
if x not in in_grad:
in_grad[x] = grads.get_as_list(x)
# to reduce memory usage
x._set_grad_var_if_available(None)
_backprop_utils.backprop_step(
func, target_input_indexes, out_grad, in_grad)
for hook in hooks:
hook.backward_postprocess(func, in_data, out_grad_array)
for y, gy in six.moves.zip(outputs, out_grad):
if y is not None and y is not self.node:
y._set_grad_var_if_available(
gy if retain_grad else None)
del gy, out_grad # to reduce memory usage
for x, gx in in_grad.items():
if not gx: # gradient == None
continue
for gx_elem in gx:
_check_grad_type(func, x, gx_elem.array)
del gx_elem # to reduce memory usage
if x.creator_node is None: # leaf
leaf_nodes.add(x)
else:
add_cand(x.creator_node)
del gx, in_grad # to reduce memory usage
for x in leaf_nodes:
x_var = x.get_variable_or_none()
gx = grads.pop(x)
if x_var is not None:
x_var._grad_var = gx
x_var._loss_scale = loss_scale
grads.assert_no_grads()
def apply(self, inputs):
"""Computes output variables and grows the computational graph.
Basic behavior is expressed in the documentation of
:class:`FunctionNode`.
.. note::
If the :data:`~Variable.data` attribute of input variables exist on
a GPU device, that device is made current before calling
:meth:`forward`, so implementors do not need to take care of device
selection in most cases.
Args:
inputs: Tuple of input variables. Each element can be either
:class:`~chainer.Variable`, :class:`numpy.ndarray`,
or :class:`cupy.ndarray`. If the element is an ndarray, it is
automatically wrapped with :class:`~chainer.Variable`.
Returns:
A tuple of output :class:`~chainer.Variable` objects.
"""
chainerx_in_data = None
chainerx_device = None
is_chainerx, in_data = _extract_apply_in_data(inputs)
if is_chainerx:
# Try ChainerX C++ implementation.
# If it's supported, the output arrays are wrapped with Variables
# and returned.
# If not supported, FunctionNode.forward_chainerx should return
# Fallback.
# In that case the input arrays are converted to numpy.ndarray
# or cupy.ndarray (depending on the ChainerX backend) and
# forward computation falls back to the conventional
# FunctionNode.forward() implementaion.
outputs = self.forward_chainerx(in_data)
if outputs is not chainer.Fallback:
# Supported. Wrap with variables and return
assert isinstance(outputs, tuple)
return tuple([
variable.Variable._init_unchecked(
y,
requires_grad=y.is_backprop_required(),
is_chainerx_array=True) for y in outputs
])
# Fall back to FunctionNode.forward()
chainerx_in_data, in_data, chainerx_device = (
self._chainerx_apply_fallback_preprocess(in_data, inputs))
self._is_chainerx_fallback_mode = True
self.chainerx_device = chainerx_device
utils._check_arrays_forward_compatible(in_data, self.label)
is_debug = chainer.is_debug()
if is_debug:
# Keep stack trace for debug
self.stack = traceback.extract_stack()
if configuration.config.type_check:
self._check_data_type_forward(in_data)
hooks = chainer.get_function_hooks()
if self._n_local_function_hooks > 0:
hooks = collections.OrderedDict(hooks)
hooks.update(self.local_function_hooks)
hooks = hooks.values() # avoid six for performance
for hook in hooks:
hook.forward_preprocess(self, in_data)
# Forward propagation
with cuda.get_device_from_array(*in_data):
self._input_indexes_to_retain = None
self._output_indexes_to_retain = None
if chainer.config.schedule_func is not None:
outputs = static_forward_optimizations(self, in_data)
elif self._is_chainerx_fallback_mode:
# In ChainerX fallback, __class__ is temporarily replaced with
# the fabricated one with automatic attirbute fallback.
with _chainerx_attribute_fallback(self, chainerx_device):
outputs = self.forward(in_data)
else:
# In normal case, simply run the forward method.
outputs = self.forward(in_data)
# Check for output array types
if not isinstance(outputs, tuple):
raise TypeError('forward output must be a tuple ({})\n'
'Actual: {}'.format(self.label, type(outputs)))
if not chainer.is_arrays_compatible(outputs):
raise TypeError(
'incompatible array types are mixed in the forward output '
'({}).\n'
'Actual: {}'.format(self.label,
', '.join(str(type(x)) for x in outputs)))
for hook in hooks:
hook.forward_postprocess(self, in_data)
# NaN check of output values
if is_debug:
if any(chainer.backend._contains_nan(out) for out in outputs):
msg = ('NaN is detected on forward computation of '
'{}'.format(self.label))
raise RuntimeError(msg)
self._output_count = len(outputs)
if self._is_chainerx_fallback_mode:
ret = self._chainerx_apply_fallback_postprocess(
chainerx_in_data, inputs, outputs)
else:
input_vars = [chainer.as_variable(x) for x in inputs]
requires_grad = any([x.requires_grad for x in input_vars])
ret = tuple([
variable.Variable(y, requires_grad=requires_grad)
for y in outputs
])
if configuration.config.enable_backprop:
# Topological ordering
self.rank = max([x.rank
for x in input_vars]) if input_vars else 0
# Add backward edges
for y in ret:
y.creator_node = self
self.inputs = tuple([x.node for x in input_vars])
# Add forward edges (must be weak references)
self.outputs = tuple([weakref.ref(y.node) for y in ret])
if self._input_indexes_to_retain is not None:
for index in self._input_indexes_to_retain:
input_vars[index].retain_data()
if self._output_indexes_to_retain is not None:
retained_data = []
for index in self._output_indexes_to_retain:
ret[index].retain_data()
retained_data.append(outputs[index])
self._retained_output_data = tuple(retained_data)
self.lazy_grad_sum = configuration.config.lazy_grad_sum
return ret
def backward(self, retain_grad=False):
"""Runs error backpropagation (a.k.a. backprop) from this variable.
On backprop, :meth:`Function.backward` is called on each
:class:`Function` object appearing in the backward graph starting from
this variable. The backward graph is represented by backward references
from variables to their creators, and from functions to their inputs.
The backprop stops at all root variables. Some functions set ``None``
as gradients of some inputs, where further backprop does not take place
at such input variables.
This method uses :data:`grad` as the initial error array. User can
manually set a gradient array before calling this method. If
:data:`data` contains only one element (i.e., it is scalar) and
:data:`grad` is ``None``, then this method automatically complements
1.0 as the initial error. This is useful on starting backprop from
some scalar loss value.
Args:
retain_grad (bool): If ``True``, the gradient arrays of all
intermediate variables are kept. Otherwise, :data:`grad` of the
intermediate variables are set to ``None`` on appropriate
timing, which may reduce the maximum memory consumption.
In most cases of training some models, the purpose of backprop
is to compute gradients of parameters, not of variables, so it
is recommended to set this flag ``False``.
"""
if self.creator is None:
return
initial_device = None
if cuda.available and isinstance(self.data, cuda.cupy.ndarray):
try:
initial_device = cuda.Device()
except cuda.cupy.cuda.runtime.CUDARuntimeError as e:
if e.status != 38: # cudaErrorNoDevice
raise
is_debug = chainer.is_debug()
cand_funcs = []
seen_set = set()
seen_vars = set()
need_copy = set()
# Initialize error by 1, if this is a loss variable
if self.data.size == 1 and self.grad is None:
with cuda.get_device_from_array(self.data) as device:
if device is cuda.DummyDevice:
self.grad = numpy.ones_like(self.data)
else:
self.grad = cuda.cupy.ones_like(self.data)
def add_cand(cand):
if cand not in seen_set:
# Negate since heapq is min-heap
heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand))
seen_set.add(cand)
add_cand(self.creator)
while cand_funcs:
_, _, func = heapq.heappop(cand_funcs)
outputs = [y() for y in func.outputs] # access via weak ref
in_data = tuple([x.data for x in func.inputs])
out_grad = tuple([None if y is None else y.grad for y in outputs])
hooks = chainer.get_function_hooks()
if func._n_local_function_hooks != 0:
hooks = collections.OrderedDict(hooks)
hooks.update(func.local_function_hooks)
cuda.get_device_from_array(*(in_data + out_grad)).use()
for hook in six.itervalues(hooks):
hook.backward_preprocess(func, in_data, out_grad)
gxs = func.backward(in_data, out_grad)
assert len(gxs) == len(in_data)
for hook in six.itervalues(hooks):
hook.backward_postprocess(func, in_data, out_grad)
if is_debug:
for gx in gxs:
if gx is None:
continue
cuda.get_device_from_array(gx).use()
if cuda.get_array_module(gx).isnan(gx).any():
msg = 'NaN is detected on backward computation'
raise RuntimeError(msg)
if not retain_grad:
for y in outputs:
if y is not None and y is not self:
y.grad = None
for x, gx in zip(func.inputs, gxs):
if gx is None:
continue
_check_grad_type(func, x, gx)
# Accumulate the gradient to x. It is a bit tricky to handle
# branches and parameter gradient accumulation correctly.
id_x = id(x)
if x.creator is None: # leaf
if x._grad is None:
x.grad = gx
need_copy.add(id_x)
else:
cuda.get_device_from_array(gx).use()
if id_x in need_copy:
x.grad = utils.force_array(x.grad + gx) # copy
need_copy.remove(id_x)
else:
x._grad += gx
else: # not a leaf
add_cand(x.creator)
if id_x not in seen_vars: # 1st visit
x.grad = gx
seen_vars.add(id_x)
need_copy.add(id_x)
else:
cuda.get_device_from_array(gx).use()
if id_x in need_copy: # 2nd visit
x._grad = utils.force_array(gx + x._grad) # copied
need_copy.remove(id_x)
else: # 3rd or later visit
x._grad += gx
del gxs # to reduce memory usage
if initial_device is not None:
initial_device.use()
def apply(self, inputs):
"""Computes output variables and grows the computational graph.
Basic behavior is expressed in the documentation of
:class:`FunctionNode`.
.. note::
If the :data:`~Variable.data` attribute of input variables exist on
a GPU device, that device is made current before calling
:meth:`forward`, so implementors do not need to take care of device
selection in most cases.
Args:
inputs: Tuple of input variables. Each element can be either
:class:`Variable`, :class:`numpy.ndarray`,
or :class:`cupy.ndarray`. If the element is an ndarray, it is
automatically wrapped with :class:`Variable`.
Returns:
A tuple of output :class:`Variable` objects.
"""
input_vars = [x if isinstance(x, variable.Variable)
else variable.Variable(x, requires_grad=False)
for x in inputs]
in_data = tuple([x.data for x in input_vars])
requires_grad = any([x.requires_grad for x in input_vars])
if chainer.is_debug():
self.stack = traceback.extract_stack()
if configuration.config.type_check:
self._check_data_type_forward(in_data)
hooks = chainer.get_function_hooks()
if self._n_local_function_hooks > 0:
hooks = collections.OrderedDict(hooks)
hooks.update(self.local_function_hooks)
hooks = hooks.values() # avoid six for performance
for hook in hooks:
hook.forward_preprocess(self, in_data)
# Forward propagation
with cuda.get_device_from_array(*in_data):
self._input_indexes_to_retain = None
self._output_indexes_to_retain = None
outputs = self.forward(in_data)
assert type(outputs) is tuple
for hook in hooks:
hook.forward_postprocess(self, in_data)
# NaN check of output values
if chainer.is_debug():
if any(out.dtype.kind == 'f' and
cuda.get_array_module(out).isnan(out).any()
for out in outputs):
msg = 'NaN is detected on forward computation'
raise RuntimeError(msg)
ret = tuple([variable.Variable(y, requires_grad=requires_grad)
for y in outputs])
if configuration.config.enable_backprop:
# Topological ordering
self.rank = max([x.rank for x in input_vars]) if input_vars else 0
# Add backward edges
for i, y in enumerate(ret):
y.creator_node = self
self.inputs = tuple([x.node for x in input_vars])
# Add forward edges (must be weak references)
self.outputs = tuple([weakref.ref(y.node) for y in ret])
if self._input_indexes_to_retain is not None:
for index in self._input_indexes_to_retain:
input_vars[index].retain_data()
if self._output_indexes_to_retain is not None:
retained_data = []
for index in self._output_indexes_to_retain:
ret[index].retain_data()
retained_data.append(outputs[index])
self._retained_output_data = tuple(retained_data)
return ret
