def _getitem(arr, key):
if not isinstance(arr, chainerx.ndarray):
return arr[key]
try:
return arr[key]
except (IndexError, chainerx.DimensionError):
pass
is_backprop_required = arr.is_backprop_required()
arr = backend.from_chainerx(arr)
if isinstance(key, chainerx.ndarray):
key = backend.from_chainerx(key)
if isinstance(arr, cuda.ndarray):
with arr.device:
ret = arr[key]
else:
ret = arr[key]
# Doing this check after the fallback __getitem__ because the error which
# caused the fallback might not be due to advanced indexing. In such
# case the fallback __getitem__ should also raise the error.
if is_backprop_required:
raise RuntimeError(
'ChainerX getitem fallback for advanced indexing is not supported '
'for arrays that are connected to a graph.')
return backend.to_chainerx(ret)
def update_core_chainerx(self, param):
"""Updates the ChainerX parameter.
This method can be overridden to implement custom update logic.
The default implementation is to convert the parameter to a
memory-shared NumPy/CuPy parameter and call the corresponding update
method.
See :meth:`update_core` for details.
Args:
param (~chainer.Variable): Variable to be updated.
"""
grad_array = param.grad
backend_name = param.array.device.backend.name
if backend_name == 'native':
update_core = self.update_core_cpu
elif backend_name == 'cuda':
update_core = self.update_core_gpu
else:
raise RuntimeError(
'Default implementation of Optimizer.update_core_chainerx is '
'only provided for native or cuda backends (actual: {}). '
'Override Optimizer.update_core_chainerx() to implement '
'custom update logic.'.format(backend_name))
# Convert state arrays to NumPy/CuPy
chainerx_state_arrays = {}
for state_name, st in self.state.items():
st = self.state[state_name]
if isinstance(st, chainerx.ndarray):
self.state[state_name] = backend.from_chainerx(st)
chainerx_state_arrays[state_name] = st
# Create a temporary parameter with memory-shared NumPy/CuPy array
# If the ChainerX parameter has a cached NumPy/CuPy copy, use the
# cache and avoid redundant conversion. Else, create the cache here
# and use it.
if param._chainerx_fallback_array is None:
param._chainerx_fallback_array = backend.from_chainerx(param.array)
temp_param = variable.Variable._init_unchecked(
param._chainerx_fallback_array, is_chainerx_array=False)
if grad_array is not None:
temp_param._set_grad_without_check(
backend.from_chainerx(grad_array))
# Update
update_core(temp_param)
# Restore state arrays
for state_name, arr in chainerx_state_arrays.items():
cur_arr = self.state[state_name]
if cur_arr is not arr:
arr = backend.to_chainerx(cur_arr)
self.state[state_name] = arr
def update_core_chainerx(self, param):
"""Updates the ChainerX parameter.
This method can be overridden to implement custom update logic.
The default implementation is to convert the parameter to a
memory-shared NumPy/CuPy parameter and call the corresponding update
method.
See :meth:`update_core` for details.
Args:
param (~chainer.Variable): Variable to be updated.
"""
grad_array = param.grad
backend_name = param.array.device.backend.name
if backend_name == 'native':
update_core = self.update_core_cpu
elif backend_name == 'cuda':
update_core = self.update_core_gpu
else:
raise RuntimeError(
'Default implementation of Optimizer.update_core_chainerx is '
'only provided for native or cuda backends (actual: {}). '
'Override Optimizer.update_core_chainerx() to implement '
'custom update logic.'.format(backend_name))
# Convert state arrays to NumPy/CuPy
chainerx_state_arrays = {}
for state_name, st in self.state.items():
st = self.state[state_name]
if isinstance(st, chainerx.ndarray):
self.state[state_name] = backend.from_chainerx(st)
chainerx_state_arrays[state_name] = st
# Create a temporary parameter with memory-shared NumPy/CuPy array
# If the ChainerX parameter has a cached NumPy/CuPy copy, use the
# cache and avoid redundant conversion. Else, create the cache here
# and use it.
if param._chainerx_fallback_array is None:
param._chainerx_fallback_array = backend.from_chainerx(
param.array)
temp_param = variable.Variable(param._chainerx_fallback_array)
if grad_array is not None:
temp_param._set_grad_without_check(
backend.from_chainerx(grad_array))
# Update
update_core(temp_param)
# Restore state arrays
for state_name, arr in chainerx_state_arrays.items():
self.state[state_name] = arr
def _setitem(arr, key, value):
"""Sets arr[key] to value by falling back to a non-ChainerX arrays.
Supports both basic and advanced indexing.
Note:
With the ``cuda`` backend, the behavior differs from NumPy when
integer arrays in ``slices`` reference the same location
multiple times. In that case, the value that is actually stored
is undefined.
>>> import chainerx
>>> chainerx.set_default_device('cuda:0')
>>> a = chainerx.zeros((2,), dtype=chainerx.float)
>>> i = chainerx.array([0, 1, 0, 1, 0, 1])
>>> v = chainerx.arange(6).astype(chainerx.float)
>>> a[i] = v
>>> a # doctest: +SKIP
array([2., 3.], shape=(2,), dtype=float64, device='cuda:0')
On the other hand, NumPy and ``native`` backend store the value
corresponding to the last index among the indices referencing
duplicate locations.
>>> import numpy
>>> a_cpu = numpy.zeros((2,), dtype=numpy.float)
>>> i_cpu = numpy.array([0, 1, 0, 1, 0, 1])
>>> v_cpu = numpy.arange(6).astype(numpy.float)
>>> a_cpu[i_cpu] = v_cpu
>>> a_cpu
array([4., 5.])
"""
if not isinstance(arr, chainerx.ndarray):
arr[key] = value
return
if arr.is_backprop_required():
raise RuntimeError(
'ChainerX setitem fallback for advanced indexing is not supported '
'for arrays that are connected to a graph.')
arr = backend.from_chainerx(arr)
if isinstance(key, chainerx.ndarray):
key = backend.from_chainerx(key)
if isinstance(value, chainerx.ndarray):
value = backend.from_chainerx(value)
if isinstance(arr, cuda.ndarray):
with arr.device:
arr[key] = value
else:
arr[key] = value
def backward(self, target_input_indexes, grad_outputs):
retained_inputs = self.get_retained_inputs()
inputs = [None] * len(self.inputs)
in_data = [None] * len(self.inputs)
for retained, i_in in six.moves.zip(retained_inputs,
self._input_indexes_to_retain):
inputs[i_in] = retained
in_data[i_in] = None if retained is None else retained.array
in_data = tuple(in_data)
grad_out_data = tuple(
[None if grad is None else grad.data for grad in grad_outputs])
is_chainerx_fallback_mode = self._is_chainerx_fallback_mode
if is_chainerx_fallback_mode:
# Convert input and output gradients to numpy/cupy
in_data = backend.from_chainerx(in_data)
grad_out_data = backend.from_chainerx(grad_out_data)
# Call Function.backward
with cuda.get_device_from_array(*(in_data + grad_out_data)):
if is_chainerx_fallback_mode:
# Enable attribute fallback
with function_node._chainerx_attribute_fallback(
self._function, self.chainerx_device):
gxs = self._function.backward(in_data, grad_out_data)
else:
gxs = self._function.backward(in_data, grad_out_data)
# Check gradients
for x, gx in six.moves.zip(self.inputs, gxs):
if gx is not None:
variable._check_grad_type(self, x, True, gx)
# Convert input gradients back to ChainerX
if is_chainerx_fallback_mode:
gxs = backend.to_chainerx(gxs)
ret = []
for i in target_input_indexes:
if gxs[i] is None:
g = None
else:
# Intentionally not passing requires_grad=False so that
# backprop routines can raise an error when a further backprop
# is attempted against this gradient variable.
g = variable.Variable(gxs[i])
if g.xp is not chainerx:
g.node._old_style_grad_generator = self._function.label
ret.append(g)
return tuple(ret)
def backward(self, target_input_indexes, grad_outputs):
retained_inputs = self.get_retained_inputs()
inputs = [None] * len(self.inputs)
in_data = [None] * len(self.inputs)
for retained, i_in in six.moves.zip(
retained_inputs, self._input_indexes_to_retain):
inputs[i_in] = retained
in_data[i_in] = retained.array
in_data = tuple(in_data)
grad_out_data = tuple([None if grad is None else grad.data
for grad in grad_outputs])
is_chainex_fallback_mode = self._is_chainex_fallback_mode
if is_chainex_fallback_mode:
# Convert input and output gradients to numpy/cupy
in_data = backend.from_chainerx(in_data)
grad_out_data = backend.from_chainerx(grad_out_data)
# Call Function.backward
with cuda.get_device_from_array(*(in_data + grad_out_data)):
if is_chainex_fallback_mode:
# Enable attribute fallback
with function_node._chainerx_attribute_fallback(
self._function, self.chainerx_device):
gxs = self._function.backward(in_data, grad_out_data)
else:
gxs = self._function.backward(in_data, grad_out_data)
for x, gx in six.moves.zip(self.inputs, gxs):
variable._check_grad_type(self, x, True, gx, False)
# Convert input gradients back to ChainerX
if is_chainex_fallback_mode:
gxs = backend.to_chainerx(gxs)
ret = []
for i in target_input_indexes:
if gxs[i] is None:
g = None
else:
# Intentionally not passing requires_grad=False so that
# backprop routines can raise an error when a further backprop
# is attempted against this gradient variable.
g = variable.Variable(gxs[i])
if g.xp is not chainerx:
g.node._old_style_grad_generator = self._function.label
ret.append(g)
return tuple(ret)
def _chainerx_apply_fallback_preprocess(self, in_data, inputs):
chainerx_in_data = in_data
in_data = []
device = None
for data, x in six.moves.zip(chainerx_in_data, inputs):
if data is None:
fallback_data = None
else:
# Use the cached fallback arrays as inputs if they exist.
x_is_variable = isinstance(x, variable.Variable)
if x_is_variable and x._chainerx_fallback_array is not None:
fallback_data = x._chainerx_fallback_array
if device is None:
device = x.device
else:
fallback_data = backend.from_chainerx(data)
if device is None:
device = backend.ChainerxDevice(data.device)
# Update the fallback cache if possible.
if x_is_variable:
x._chainerx_fallback_array = fallback_data
in_data.append(fallback_data)
in_data = tuple(in_data)
return chainerx_in_data, in_data, device
def forward_cpu(self, inputs):
class_weight = backend.from_chainerx(self.class_weight)
self.retain_inputs((0, 1))
x, t = inputs
if chainer.is_debug():
_check_input_values(x, t, self.ignore_label)
log_y = log_softmax._log_softmax(x)
if self.cache_score:
self.y = numpy.exp(log_y)
if class_weight is not None:
shape = [1 if d != 1 else -1 for d in six.moves.range(x.ndim)]
log_y *= _broadcast_to(class_weight.reshape(shape), x.shape)
log_yd = numpy.rollaxis(log_y, 1)
log_yd = log_yd.reshape(len(log_yd), -1)
t_valid = t != self.ignore_label
t = t * t_valid
log_p = log_yd[t.ravel(), numpy.arange(t.size)]
log_p *= t_valid.ravel()
if self.reduce == 'mean':
# deal with the case where the SoftmaxCrossEntropy is
# unpickled from the old version
if self.normalize:
count = t_valid.sum()
else:
count = len(x)
self._coeff = 1.0 / max(count, 1)
y = log_p.sum(keepdims=True) * (-self._coeff)
return y.reshape(()),
else:
return -log_p.reshape(t.shape),
def forward_cpu(self, inputs):
class_weight = backend.from_chainerx(self.class_weight)
self.retain_inputs((0, 1))
x, t = inputs
if chainer.is_debug():
_check_input_values(x, t, self.ignore_label)
log_y = log_softmax._log_softmax(x)
if self.cache_score:
self.y = numpy.exp(log_y)
if class_weight is not None:
shape = [1 if d != 1 else -1 for d in six.moves.range(x.ndim)]
log_y *= _broadcast_to(class_weight.reshape(shape), x.shape)
log_yd = numpy.rollaxis(log_y, 1)
log_yd = log_yd.reshape(len(log_yd), -1)
t_valid = t != self.ignore_label
t = t * t_valid
log_p = log_yd[t.ravel(), numpy.arange(t.size)]
log_p *= t_valid.ravel()
if self.reduce == 'mean':
# deal with the case where the SoftmaxCrossEntropy is
# unpickled from the old version
if self.normalize:
count = t_valid.sum()
else:
count = len(x)
self._coeff = 1.0 / max(count, 1)
y = log_p.sum(keepdims=True) * (-self._coeff)
return y.reshape(()),
else:
return -log_p.reshape(t.shape),
def _make_samples(self, t):
size = int(t.shape[0])
# first one is the positive, and others are sampled negatives
samples = self.sampler((size, self.sample_size + 1))
samples = backend.from_chainerx(samples)
samples[:, 0] = t
return samples
def forward_gpu(self, inputs):
t = backend.from_chainerx(self.t) # Workaround for ChainerX.
gx = cuda.cupy.zeros(self.shape, self.dtype)
gx = cuda.elementwise('S t, T gloss', 'raw T gx',
'int ind[] = {i, t}; gx[ind] = gloss;',
'getitem_bwd')(t, inputs[0], gx)
return gx,
def getattribute(self, name):
value = sup.__getattribute__(name)
if isinstance(value, chainerx.ndarray):
fallback_arr = fallback_array_cache.get(name)
if fallback_arr is None:
fallback_arr = backend.from_chainerx(value)
fallback_array_cache[name] = fallback_arr
return fallback_arr
return value
def getattribute(self, name):
value = sup.__getattribute__(name)
if isinstance(value, chainerx.ndarray):
fallback_arr = fallback_array_cache.get(name)
if fallback_arr is None:
fallback_arr = backend.from_chainerx(value)
fallback_array_cache[name] = fallback_arr
return fallback_arr
return value
def output_data(self):
"""A tuple of the retained output arrays.
It has the same length as the :attr:`outputs`. Elements that are not
retained are set to ``None``.
"""
if self.node._is_chainerx_fallback_mode:
return backend.from_chainerx(self.node.output_data)
return self.node.output_data
def output_data(self):
"""A tuple of the retained output arrays.
It has the same length as the :attr:`outputs`. Elements that are not
retained are set to ``None``.
"""
if self.node._is_chainerx:
return backend.from_chainerx(self.node.output_data)
return self.node.output_data
def forward_gpu(self, inputs):
t = backend.from_chainerx(self.t) # Workaround for ChainerX.
gx = cuda.cupy.zeros(self.shape, self.dtype)
gx = cuda.elementwise(
'S t, T gloss',
'raw T gx',
'int ind[] = {i, t}; gx[ind] = gloss;',
'getitem_bwd'
)(t, inputs[0], gx)
return gx,
def _getitem(arr, key):
try:
return arr[key]
except (IndexError, chainerx.DimensionError):
pass
if isinstance(arr, chainerx.ndarray):
arr = backend.from_chainerx(arr)
is_arr_chainerx = True
else:
is_arr_chainerx = False
if isinstance(key, chainerx.ndarray):
key = backend.from_chainerx(key)
if isinstance(arr, cuda.ndarray):
with arr.device:
ret = arr[key]
else:
ret = arr[key]
if is_arr_chainerx:
ret = backend.to_chainerx(ret)
return ret
def forward_gpu(self, inputs):
class_weight = backend.from_chainerx(self.class_weight)
self.retain_inputs((0, 1))
cupy = cuda.cupy
x, t = inputs
if chainer.is_debug():
_check_input_values(x, t, self.ignore_label)
if x.size == 0:
y = cupy.zeros(t.shape, dtype=x.dtype)
if self.cache_score:
self.y = y
if self.reduce == 'mean':
return y.sum(),
else:
return y,
log_y = log_softmax._log_softmax(x)
if self.cache_score:
self.y = cupy.exp(log_y)
if class_weight is not None:
shape = [1 if d != 1 else -1 for d in six.moves.range(x.ndim)]
log_y *= cupy.broadcast_to(class_weight.reshape(shape), x.shape)
if self.normalize:
coeff = cupy.maximum(1, (t != self.ignore_label).sum())
else:
coeff = max(1, len(t))
self._coeff = cupy.divide(1.0, coeff, dtype=x.dtype)
log_y = cupy.rollaxis(log_y, 1, log_y.ndim)
if self.reduce == 'mean':
ret = cuda.reduce(
'S t, raw T log_y, int32 n_channel, raw T coeff, '
'S ignore_label',
'T out',
't == ignore_label ? T(0) : log_y[_j * n_channel + t]',
'a + b', 'out = a * -coeff[0]', '0', 'crossent_fwd'
)(t, log_y.reduced_view(), log_y.shape[-1],
self._coeff, self.ignore_label)
else:
ret = cuda.elementwise(
'S t, raw T log_y, int32 n_channel, T ignore', 'T out',
'''
if (t == ignore) {
out = 0;
} else {
out = -log_y[i * n_channel + t];
}
''',
'softmax_crossent_no_reduce_fwd'
)(t, log_y.reduced_view(), log_y.shape[-1], self.ignore_label)
ret = ret.reshape(t.shape)
return ret,
def forward_gpu(self, inputs):
class_weight = backend.from_chainerx(self.class_weight)
self.retain_inputs((0, 1))
cupy = cuda.cupy
x, t = inputs
if chainer.is_debug():
_check_input_values(x, t, self.ignore_label)
if x.size == 0:
y = cupy.zeros(t.shape, dtype=x.dtype)
if self.cache_score:
self.y = y
if self.reduce == 'mean':
return y.sum(),
else:
return y,
log_y = log_softmax._log_softmax(x)
if self.cache_score:
self.y = cupy.exp(log_y)
if class_weight is not None:
shape = [1 if d != 1 else -1 for d in six.moves.range(x.ndim)]
log_y *= cupy.broadcast_to(class_weight.reshape(shape), x.shape)
if self.normalize:
coeff = cupy.maximum(1, (t != self.ignore_label).sum())
else:
coeff = max(1, len(t))
self._coeff = cupy.divide(1.0, coeff, dtype=x.dtype)
log_y = cupy.rollaxis(log_y, 1, log_y.ndim)
if self.reduce == 'mean':
ret = cuda.reduce(
'S t, raw T log_y, int32 n_channel, raw T coeff, '
'S ignore_label', 'T out',
't == ignore_label ? T(0) : log_y[_j * n_channel + t]',
'a + b', 'out = a * -coeff[0]', '0',
'crossent_fwd')(t, log_y.reduced_view(), log_y.shape[-1],
self._coeff, self.ignore_label)
else:
ret = cuda.elementwise(
'S t, raw T log_y, int32 n_channel, T ignore', 'T out', '''
if (t == ignore) {
out = 0;
} else {
out = -log_y[i * n_channel + t];
}
''', 'softmax_crossent_no_reduce_fwd')(t, log_y.reduced_view(),
log_y.shape[-1],
self.ignore_label)
ret = ret.reshape(t.shape)
return ret,
def from_chainerx(self):
"""Converts parameter variables and persistent values from ChainerX \
to NumPy/CuPy devices without any copy."""
d = self.__dict__
for name in self._params:
d[name].from_chainerx()
for name in self._persistent:
if not numpy.isscalar(d[name]):
d[name] = backend.from_chainerx(d[name])
if isinstance(self._device, backend.ChainerxDevice):
self._device = self._device.fallback_device
return self
def test_from_chainerx(self, backend_config):
arr = backend_config.get_array(numpy.ones((2, 3), numpy.float32))
arr_converted = backend.from_chainerx(arr)
src_device = backend_config.device
if src_device.xp is chainerx:
dst_xp = src_device.fallback_device.xp
assert isinstance(arr_converted, dst_xp.ndarray)
if dst_xp is cuda.cupy:
assert arr_converted.device.id == src_device.device.index
else:
assert arr is arr_converted
with backend_config:
self.check_equal_memory_shared(arr, arr_converted)
def test_from_chainerx(self, backend_config):
arr = backend_config.get_array(numpy.ones((2, 3), numpy.float32))
arr_converted = backend.from_chainerx(arr)
src_device = backend_config.device
if src_device.xp is chainerx:
dst_xp = src_device.fallback_device.xp
assert isinstance(arr_converted, dst_xp.ndarray)
if dst_xp is cuda.cupy:
assert arr_converted.device.id == src_device.device.index
else:
assert arr is arr_converted
with backend_config:
self.check_equal_memory_shared(arr, arr_converted)
def forward(self, xs):
a = xs[0]
b = xs[1]
y = a.copy()
xp = backend.get_array_module(a)
slices = tuple([
backend.from_chainerx(s) if isinstance(s, chainerx.ndarray) else s
for s in self.slices
])
if y[slices].shape != b.shape:
raise ValueError('Chainer does not support automatic broadcasting '
'of variables.')
if xp is numpy:
numpy.add.at(y, slices, b),
else:
cuda.cupyx.scatter_add(y, slices, b),
return y,
def _chainerx_apply_fallback_preprocess(self, in_data, inputs):
chainerx_in_data = in_data
in_data = []
for i in six.moves.range(len(inputs)):
# Use the cached fallback arrays as inputs if they exist.
x = inputs[i]
x_is_variable = isinstance(x, variable.Variable)
if x_is_variable and x._chainerx_fallback_array is not None:
x_data = x._chainerx_fallback_array
else:
x_data = backend.from_chainerx(chainerx_in_data[i])
# Update the fallback cache if possible.
if x_is_variable:
x._chainerx_fallback_array = x_data
in_data.append(x_data)
in_data = tuple(in_data)
return chainerx_in_data, in_data
def _chainerx_apply_fallback_preprocess(self, in_data, inputs):
chainerx_in_data = in_data
in_data = []
device = None
for data, x in six.moves.zip(chainerx_in_data, inputs):
# Use the cached fallback arrays as inputs if they exist.
x_is_variable = isinstance(x, variable.Variable)
if x_is_variable and x._chainerx_fallback_array is not None:
fallback_data = x._chainerx_fallback_array
if device is None:
device = x.device
else:
fallback_data = backend.from_chainerx(data)
if device is None:
device = backend.ChainerxDevice(data.device)
# Update the fallback cache if possible.
if x_is_variable:
x._chainerx_fallback_array = fallback_data
in_data.append(fallback_data)
in_data = tuple(in_data)
return chainerx_in_data, in_data, device
def forward(self, inputs):
self.retain_inputs((0, 1))
x, gamma, beta = inputs
xp = backend.get_array_module(x)
if self.running_mean is None:
self.running_mean = xp.zeros_like(gamma)
self.running_var = xp.zeros_like(gamma)
self.axis = _compute_axis(x.ndim, gamma.ndim, self.axis)
self.key_axis = _compute_key_axis(x.ndim, gamma.ndim, self.axis)
if all(x.shape[i] == 1 for i in self.axis):
if 0 in self.axis:
warnings.warn(
'A batch with no more than one sample has been given'
' to F.batch_normalization. F.batch_normalization'
' will always output a zero tensor for such batches.'
' This could be caused by incorrect configuration in'
' your code (such as running evaluation while'
' chainer.config.train=True),'
' but could also happen in the last batch of training'
' if non-repeating iterator is used.',
UserWarning)
else:
warnings.warn(
'F.batch_normalization received a batch with single'
' dimensions along all axes that are used for aggregating'
' statistics. F.batch_normalization'
' will always output a zero tensor for such batches.',
UserWarning)
# TODO(niboshi): Refactor calculation of expander and axis into a
# function and call it just before they are used.
# expander inserts singleton dimensions to gamma and beta so that they
# can be broadcasted with x.
expander = [None for _ in range(x.ndim)]
for i in self.key_axis:
expander[i] = slice(None)
expander = tuple(expander)
self.expander = expander
self.mode = _BNMode(x, gamma, self.key_axis)
self.use_cudnn = self.mode.can_use_cudnn(xp)
self.use_ideep = self.mode.can_use_ideep()
if self.use_ideep:
# TODO(niboshi): Refactor iDeep part into a separate method
expand_dim = False
if x.ndim == 2:
expand_dim = True
x = x[:, :, None, None]
y, self.mean, self.var, self.inv_std = (
intel64.ideep.batchNormalization.Forward(
intel64.ideep.array(x),
intel64.ideep.array(gamma),
intel64.ideep.array(beta),
None,
None,
self.eps
))
m = x.size // gamma.size
adjust = m / max(m - 1., 1.)
# Update running_mean
if isinstance(self.running_mean, intel64.ideep.mdarray):
self.running_mean.inplace_axpby(
self.decay, (1 - self.decay), self.mean)
else:
self.running_mean *= self.decay
self.running_mean += self.mean * (1 - self.decay)
# Update running_var
if isinstance(self.running_var, intel64.ideep.mdarray):
self.running_var.inplace_axpby(
self.decay, (1 - self.decay), self.var * adjust)
else:
self.running_var *= self.decay
self.running_var += self.var * adjust * (1 - self.decay)
if expand_dim:
y = numpy.squeeze(y, axis=(2, 3))
elif self.use_cudnn:
if self.mean is None:
# Output cache to speed up backward pass.
self.mean = xp.empty_like(gamma)
# Output cache to speed up backward pass.
self.inv_std = xp.empty_like(gamma)
y = cudnn.batch_normalization_forward_training(
x, gamma, beta, self.running_mean, self.running_var,
self.mean, self.inv_std, self.eps, self.decay,
self.mode.is_for_conv2d, self.mode.get_cudnn_mode(),
configuration.config.debug)
else:
# Generic CPU and GPU implementation
gamma = gamma[expander]
beta = beta[expander]
self.mean = x.mean(axis=self.axis)
var = x.var(axis=self.axis)
if xp is numpy:
self.inv_std = numpy.reciprocal(numpy.sqrt(
var + self.eps, dtype=x.dtype))
else:
self.inv_std = cuda.cupyx.rsqrt(var + self.eps)
y = _apply_bn_fwd(xp, x, self.mean[expander],
self.inv_std[expander], gamma, beta)
# Update running statistics
m = x.size // gamma.size
adjust = m / max(m - 1., 1.) # unbiased estimation
xp = backend.get_array_module(self.running_mean, self.running_var)
if xp is chainerx:
self.running_mean, self.running_var = backend.from_chainerx(
(self.running_mean, self.running_var))
self.running_mean *= self.decay
self.running_mean += (1 - self.decay) * self.mean
self.running_var *= self.decay
self.running_var += (1 - self.decay) * adjust * var
if xp is chainerx:
self.running_mean = backend.to_chainerx(self.running_mean)
self.running_var = backend.to_chainerx(self.running_var)
return y,
def forward(self, inputs):
self.retain_inputs((0, 1))
x, gamma, beta = inputs
xp = backend.get_array_module(x)
if self.running_mean is None:
self.running_mean = xp.zeros_like(gamma)
self.running_var = xp.zeros_like(gamma)
self.axis = _compute_axis(x.ndim, gamma.ndim, self.axis)
self.key_axis = _compute_key_axis(x.ndim, gamma.ndim, self.axis)
if all(x.shape[i] == 1 for i in self.axis):
if 0 in self.axis:
warnings.warn(
'A batch with no more than one sample has been given'
' to F.batch_normalization. F.batch_normalization'
' will always output a zero tensor for such batches.'
' This could be caused by incorrect configuration in'
' your code (such as running evaluation while'
' chainer.config.train=True),'
' but could also happen in the last batch of training'
' if non-repeating iterator is used.', UserWarning)
else:
warnings.warn(
'F.batch_normalization received a batch with single'
' dimensions along all axes that are used for aggregating'
' statistics. F.batch_normalization'
' will always output a zero tensor for such batches.',
UserWarning)
# TODO(niboshi): Refactor calculation of expander and axis into a
# function and call it just before they are used.
# expander inserts singleton dimensions to gamma and beta so that they
# can be broadcasted with x.
expander = [None for _ in range(x.ndim)]
for i in self.key_axis:
expander[i] = slice(None)
expander = tuple(expander)
self.expander = expander
self.mode = _BNMode(x, gamma, self.key_axis)
self.use_cudnn = self.mode.can_use_cudnn(xp)
self.use_ideep = self.mode.can_use_ideep()
if self.use_ideep:
# TODO(niboshi): Refactor iDeep part into a separate method
expand_dim = False
if x.ndim == 2:
expand_dim = True
x = x[:, :, None, None]
y, self.mean, self.var, self.inv_std = (
intel64.ideep.batchNormalization.Forward(
intel64.ideep.array(x), intel64.ideep.array(gamma),
intel64.ideep.array(beta), None, None, self.eps))
m = x.size // gamma.size
adjust = m / max(m - 1., 1.)
# Update running_mean
if isinstance(self.running_mean, intel64.ideep.mdarray):
self.running_mean.inplace_axpby(self.decay, (1 - self.decay),
self.mean)
else:
self.running_mean *= self.decay
self.running_mean += self.mean * (1 - self.decay)
# Update running_var
if isinstance(self.running_var, intel64.ideep.mdarray):
self.running_var.inplace_axpby(self.decay, (1 - self.decay),
self.var * adjust)
else:
self.running_var *= self.decay
self.running_var += self.var * adjust * (1 - self.decay)
if expand_dim:
y = numpy.squeeze(y, axis=(2, 3))
elif self.use_cudnn:
if self.mean is None:
# Output cache to speed up backward pass.
self.mean = xp.empty_like(gamma)
# Output cache to speed up backward pass.
self.inv_std = xp.empty_like(gamma)
y = cudnn.batch_normalization_forward_training(
x, gamma, beta, self.running_mean, self.running_var, self.mean,
self.inv_std, self.eps, self.decay, self.mode.is_for_conv2d,
self.mode.get_cudnn_mode(), chainer.is_debug())
else:
# Generic CPU and GPU implementation
gamma = gamma[expander]
beta = beta[expander]
self.mean = x.mean(axis=self.axis)
var = x.var(axis=self.axis)
if xp is numpy:
self.inv_std = numpy.reciprocal(
numpy.sqrt(var + self.eps, dtype=x.dtype))
else:
self.inv_std = cuda.cupyx.rsqrt(var + self.eps)
y = _apply_bn_fwd(xp, x, self.mean[expander],
self.inv_std[expander], gamma, beta)
# Update running statistics
m = x.size // gamma.size
adjust = m / max(m - 1., 1.) # unbiased estimation
xp = backend.get_array_module(self.running_mean, self.running_var)
if xp is chainerx:
self.running_mean, self.running_var = backend.from_chainerx(
(self.running_mean, self.running_var))
self.running_mean *= self.decay
self.running_mean += (1 - self.decay) * self.mean
self.running_var *= self.decay
self.running_var += (1 - self.decay) * adjust * var
if xp is chainerx:
self.running_mean = backend.to_chainerx(self.running_mean)
self.running_var = backend.to_chainerx(self.running_var)
return y,
def forward(self, inputs):
self.retain_inputs((0, 1))
x, gamma, beta = inputs
xp = backend.get_array_module(x)
if self.running_mean is None:
self.running_mean = xp.zeros_like(gamma)
self.running_var = xp.zeros_like(gamma)
self.axis = _compute_axis(x.ndim, gamma.ndim, self.axis)
self.key_axis = _compute_key_axis(x.ndim, gamma.ndim, self.axis)
if all(x.shape[i] == 1 for i in self.axis):
if 0 in self.axis:
warnings.warn(
'A batch with no more than one sample has been given'
' to F.batch_normalization. F.batch_normalization'
' will always output a zero tensor for such batches.'
' This could be caused by incorrect configuration in'
' your code (such as running evaluation while'
' chainer.config.train=True),'
' but could also happen in the last batch of training'
' if non-repeating iterator is used.',
UserWarning)
else:
warnings.warn(
'F.batch_normalization received a batch with single'
' dimensions along all axes that are used for aggregating'
' statistics. F.batch_normalization'
' will always output a zero tensor for such batches.',
UserWarning)
# TODO(niboshi): Refactor calculation of expander and axis into a
# function and call it just before they are used.
# expander inserts singleton dimensions to gamma and beta so that they
# can be broadcasted with x.
expander = [None for _ in range(x.ndim)]
for i in self.key_axis:
expander[i] = slice(None)
expander = tuple(expander)
self.expander = expander
self.mode = _BNMode(x, gamma, self.key_axis)
self.use_cudnn = self.mode.can_use_cudnn(xp)
self.use_ideep = self.mode.can_use_ideep()
if self.use_ideep:
# TODO(niboshi): Refactor iDeep part into a separate method
expand_dim = False
if x.ndim == 2:
expand_dim = True
x = x[:, :, None, None]
y, self.mean, self.var, self.inv_std = (
intel64.ideep.batchNormalization.Forward(
intel64.ideep.array(x),
intel64.ideep.array(gamma),
intel64.ideep.array(beta),
None,
None,
self.eps
))
m = x.size // gamma.size
adjust = m / max(m - 1., 1.)
# Update running_mean
if isinstance(self.running_mean, intel64.ideep.mdarray):
self.running_mean.inplace_axpby(
self.decay, (1 - self.decay), self.mean)
else:
self.running_mean *= self.decay
self.running_mean += self.mean * (1 - self.decay)
# Update running_var
if isinstance(self.running_var, intel64.ideep.mdarray):
self.running_var.inplace_axpby(
self.decay, (1 - self.decay), self.var * adjust)
else:
self.running_var *= self.decay
self.running_var += self.var * adjust * (1 - self.decay)
if expand_dim:
y = numpy.squeeze(y, axis=(2, 3))
elif self.use_cudnn:
# TODO(niboshi): Refactor cuDNN part into a separate method
x = cuda.cupy.ascontiguousarray(x)
gamma = cuda.cupy.ascontiguousarray(gamma)
beta = cuda.cupy.ascontiguousarray(beta)
dtype = x.dtype
handle = cudnn.get_handle()
x_desc = cudnn.create_tensor_descriptor(
_as4darray(x, self.mode))
cudnn_mode = self.mode.get_cudnn_mode()
derivedBnDesc = cudnn.create_uninitialized_tensor_descriptor()
libcudnn.deriveBNTensorDescriptor(derivedBnDesc.value,
x_desc.value, cudnn_mode)
dtype_param = _get_dtype_of_tensor_descriptor(derivedBnDesc)
if dtype_param is not dtype:
gamma = gamma.astype(dtype_param)
beta = beta.astype(dtype_param)
running_mean = self.running_mean.astype(dtype_param)
running_var = self.running_var.astype(dtype_param)
else:
running_mean = self.running_mean
running_var = self.running_var
oz_dtype = (
numpy.float64 if x.dtype == numpy.float64 else numpy.float32)
one = numpy.array(1, dtype=oz_dtype).ctypes
zero = numpy.array(0, dtype=oz_dtype).ctypes
y = cuda.cupy.empty_like(x)
# Factor used in the moving average
factor = 1 - self.decay
if self.mean is None:
# Output cache to speed up backward pass.
self.mean = xp.empty_like(gamma)
# Output cache to speed up backward pass.
self.inv_std = xp.empty_like(gamma)
# Note: cuDNN computes the mini-batch mean and variance
# internally. We can simply (optionally) pass
# it the running-average mean and variance arrays.
# Note: This API seems to set the inverse of the standard deviation
# (instead of variance) to resultSaveInvVariance argument. The
# current implementation of our BN depends on this behavior so that
# we can reduce the number of reduction kernels.
libcudnn.batchNormalizationForwardTraining(
handle, cudnn_mode, one.data, zero.data,
x_desc.value, x.data.ptr, x_desc.value,
y.data.ptr, derivedBnDesc.value, gamma.data.ptr,
beta.data.ptr, factor, running_mean.data.ptr,
running_var.data.ptr, self.eps,
self.mean.data.ptr, self.inv_std.data.ptr)
# Note: When the CUDNN_BATCHNORM_SPATIAL_PERSISTENT mode is used,
# there is a possibility of numerical overflow. You can use
# queryRuntimeError() to make sure whether the overflow actually
# occured or not during the batch normalization.
if (cudnn_mode is libcudnn.CUDNN_BATCHNORM_SPATIAL_PERSISTENT and
configuration.config.debug):
query_mode = libcudnn.CUDNN_ERRQUERY_BLOCKING
rstatus = libcudnn.queryRuntimeError(handle, query_mode)
if rstatus is not libcudnn.CUDNN_STATUS_SUCCESS:
warnings.warn(
'A numerical overflow might have happend in cuDNN'
'batch normalization (status:{})'.format(rstatus))
if dtype_param is not dtype:
# When data type of prameters is converted, say, from fp16
# to fp32, the values of fp32 arrays of running_mean and
# running_var updated by batchNormalizationForwardTraining
# must be explicitly written back to their original fp16
# arrays.
running_mean = running_mean.astype(dtype)
running_var = running_var.astype(dtype)
self.running_mean.data.copy_from(running_mean.data,
running_mean.nbytes)
self.running_var.data.copy_from(running_var.data,
running_var.nbytes)
else:
# Generic CPU and GPU implementation
gamma = gamma[expander]
beta = beta[expander]
self.mean = x.mean(axis=self.axis)
var = x.var(axis=self.axis)
if xp is numpy:
self.inv_std = numpy.reciprocal(numpy.sqrt(
var + self.eps, dtype=x.dtype))
else:
self.inv_std = cuda.cupyx.rsqrt(var + self.eps)
y = _apply_bn_fwd(xp, x, self.mean[expander],
self.inv_std[expander], gamma, beta)
# Update running statistics
m = x.size // gamma.size
adjust = m / max(m - 1., 1.) # unbiased estimation
xp = backend.get_array_module(self.running_mean, self.running_var)
if xp is chainerx:
self.running_mean, self.running_var = backend.from_chainerx(
(self.running_mean, self.running_var))
self.running_mean *= self.decay
self.running_mean += (1 - self.decay) * self.mean
self.running_var *= self.decay
self.running_var += (1 - self.decay) * adjust * var
if xp is chainerx:
self.running_mean = backend.to_chainerx(self.running_mean)
self.running_var = backend.to_chainerx(self.running_var)
return y,
def forward_cpu(self, inputs):
t = backend.from_chainerx(self.t) # Workaround for ChainerX.
gx = numpy.zeros(self.shape, self.dtype)
gx[six.moves.range(self.t.size), t] = inputs[0]
return gx,
def forward_cpu(self, inputs):
t = backend.from_chainerx(self.t) # Workaround for ChainerX.
gx = numpy.zeros(self.shape, self.dtype)
gx[six.moves.range(self.t.size), t] = inputs[0]
return gx,
def forward_cpu(self, inputs):
b = backend.from_chainerx(self.b) # Workaround for ChainerX
y = (b > 0) * inputs[0]
return utils.force_array(y, dtype=y.dtype),
def forward(self, xs):
slices = tuple([
backend.from_chainerx(s) if isinstance(s, chainerx.ndarray) else s
for s in self.slices
])
return utils.force_array(xs[0][slices]),
def forward_cpu(self, inputs):
b = backend.from_chainerx(self.b) # Workaround for ChainerX
y = (b > 0) * inputs[0]
return utils.force_array(y, dtype=y.dtype),
def forward_gpu(self, inputs):
b = backend.from_chainerx(self.b) # Workaround for ChainerX
gx = _relu_grad2_kernel(b, inputs[0])
return gx,
def forward_gpu(self, inputs):
b = backend.from_chainerx(self.b) # Workaround for ChainerX
gx = _relu_grad2_kernel(b, inputs[0])
return gx,
