def test_backward_chainerx_cuda(self): # TODO(niboshi): Support it if self.dtype == numpy.float16: raise unittest.SkipTest('ChainerX does not support float16') self.check_backward_chainerx( backend.to_chainerx(cuda.to_gpu(self.x)), backend.to_chainerx(cuda.to_gpu(self.t)))
def test_backward_chainerx_cuda_nobias(self): self._skip_if_not_chainerx_supported() self.check_backward( backend.to_chainerx(cuda.to_gpu(self.x)), backend.to_chainerx(cuda.to_gpu(self.W)), None, backend.to_chainerx(cuda.to_gpu(self.gy)))
def test_double_backward_chainerx(self): # TODO(imanishi): Support it if self.dtype == numpy.float16: raise unittest.SkipTest('ChainerX does not support float16') self.check_double_backward( backend.to_chainerx(self.xs), backend.to_chainerx(self.g), backend.to_chainerx(self.ggs))
def test_backward_chainerx_cpu(self): # TODO(imanishi): Support float16 if numpy.float16 in [self.x_dtype, self.W_dtype]: raise unittest.SkipTest('ChainerX does not support float16') self.check_backward( backend.to_chainerx(self.x), backend.to_chainerx(self.W), backend.to_chainerx(self.b), backend.to_chainerx(self.gy))
def test_double_backward_chainerx_cpu(self): # TODO(imanishi): Support float16 if numpy.float16 in [self.x_dtype, self.W_dtype]: raise unittest.SkipTest('ChainerX does not support float16') inputs = [backend.to_chainerx(_) for _ in [self.x, self.W, self.b]] grad_outputs = [backend.to_chainerx(_) for _ in [self.gy]] grad_grad_inputs = [backend.to_chainerx(_) for _ in [self.ggx, self.ggW, self.ggb]] self.check_double_backward( inputs, grad_outputs, grad_grad_inputs, use_cudnn='never')
def _chainerx_apply_fallback_postprocess( self, chainerx_in_data, inputs, outputs): # TODO(hvy): Take configuration.config.enable_backprop into # account? chainerx_out_data = backend.to_chainerx(outputs) # Insert a ChainerX op-node that calls FunctionNode.backward in # backprop. Note that chainerx_out_data may not require gradients. chainerx._core._function_node_forward( self, chainerx_in_data, chainerx_out_data, [] if self._input_indexes_to_retain is None else self._input_indexes_to_retain, [] if self._output_indexes_to_retain is None else self._output_indexes_to_retain) self.inputs = tuple( [variable._ChainerxVariableNodeProps(x) for x in inputs]) ret = tuple([ _to_variable_with_chainerx_fallback_array( chainerx_out_array, out_array) for chainerx_out_array, out_array in six.moves.zip(chainerx_out_data, outputs)]) return ret
def to_chainerx(self): """Converts parameter variables and persistent values to ChainerX \ without any copy. This method does not handle non-registered attributes. If some of such attributes must be copied to ChainerX, the link implementation must override this method to do so. Returns: self """ # NOQA if not chainerx.is_available(): raise RuntimeError('ChainerX is not available.') xp = self._device.xp if xp is chainerx: return self d = self.__dict__ for name in self._params: d[name].to_chainerx() for name in self._persistent: if not numpy.isscalar(d[name]): d[name] = backend.to_chainerx(d[name]) self._device = ( backend.ChainerxDevice.from_fallback_device(self._device)) return self
def _extract_apply_in_data(inputs): # Extracts arrays from FunctionNode.apply() inputs. # # A flag that indicates whether inputs are chainerx arrays is also # returned. # # Each object in `inputs` may be `Variable` or an array. # If it's a `Variable` and its underlying array is a chainerx array, # `Variable._data[0]` (which is backproppable in contrast to # `Variable.array`) is returned. # # If at least one of the arrays is a ChainerX array, all other NumPy/CuPy # arrays are converted to ChainerX arrays without copy. if len(inputs) == 0: return False, () # Unwrap arrays arrays = [ (x._data[0] if x.xp is chainerx else x.array) if isinstance(x, variable.Variable) else x for x in inputs] if (chainerx.is_available() and any([isinstance(arr, chainerx.ndarray) for arr in arrays])): return True, tuple(backend.to_chainerx(arrays)) return False, tuple(arrays)
def test_double_backward_chainerx_native(self): self._skip_if_not_chainerx_supported() self.check_double_backward( backend.to_chainerx(self.x), backend.to_chainerx(self.W), backend.to_chainerx(self.b), backend.to_chainerx(self.gy), backend.to_chainerx(self.ggx), backend.to_chainerx(self.ggW), backend.to_chainerx(self.ggb))
def check_mix_xp(self, xp): xp_x1 = xp.random.randn(2, 3).astype(numpy.float32) xp_x2 = xp.random.randn(2, 3).astype(numpy.float32) x2 = backend.to_chainerx(xp_x2) y, = self.SimpleFunctionNode(xp).apply((xp_x1, x2)) assert isinstance(y.array, chainerx.ndarray) chainerx.testing.assert_array_equal( backend.CpuDevice().send(xp_x1 * xp_x2), y.array)
def test_to_chainerx(self, backend_config): arr = backend_config.get_array(numpy.ones((2, 3), numpy.float32)) arr_converted = backend.to_chainerx(arr) src_device = backend_config.device assert isinstance(arr_converted, chainerx.ndarray) if src_device.xp is chainerx: assert arr is arr_converted elif src_device.xp is cuda.cupy: assert arr.device.id == arr_converted.device.index self.check_equal_memory_shared(arr, arr_converted)
def test_double_backward_chainerx_cuda(self): self._skip_if_not_chainerx_supported() self.check_double_backward( backend.to_chainerx(cuda.to_gpu(self.x)), backend.to_chainerx(cuda.to_gpu(self.W)), backend.to_chainerx(cuda.to_gpu(self.b)), backend.to_chainerx(cuda.to_gpu(self.gy)), backend.to_chainerx(cuda.to_gpu(self.ggx)), backend.to_chainerx(cuda.to_gpu(self.ggW)), backend.to_chainerx(cuda.to_gpu(self.ggb)))
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 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 test_backward_chainerx_native(self): self._skip_if_not_chainerx_supported() self.check_backward( backend.to_chainerx(self.x), backend.to_chainerx(self.W), backend.to_chainerx(self.b), backend.to_chainerx(self.gy))
def test_backward_chainerx_native_nobias(self): self._skip_if_not_chainerx_supported() self.check_backward( backend.to_chainerx(self.x), backend.to_chainerx(self.W), None, backend.to_chainerx(self.gy))
def test_forward_chainerx_native(self): # TODO(niboshi): Support it if self.dtype == numpy.float16: raise unittest.SkipTest('ChainerX does not support float16') self.check_forward( backend.to_chainerx(self.x), backend.to_chainerx(self.t))
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 test_backward_chainerx_cuda_nobias(self): self.check_backward(backend.to_chainerx(cuda.to_gpu(self.x)), backend.to_chainerx(cuda.to_gpu(self.W)), None, backend.to_chainerx(cuda.to_gpu(self.gy)))
def test_backward_chainerx_native(self): self.check_backward(backend.to_chainerx(self.x), backend.to_chainerx(self.W), backend.to_chainerx(self.b), backend.to_chainerx(self.gy))
def test_forward_chainerx_native(self): self.check_forward(backend.to_chainerx(self.x), 'never')
def test_forward_chainerx_cuda(self): self.check_forward(backend.to_chainerx(cuda.to_gpu(self.x)), 'never')
def test_backward_chainerx_gpu(self): self.check_backward( backend.to_chainerx(self.x).to_device('cuda'), backend.to_chainerx(self.W).to_device('cuda'), backend.to_chainerx(self.b).to_device('cuda'), backend.to_chainerx(self.gy).to_device('cuda'))
def test_forward_chainerx_cuda(self): # TODO(niboshi): Support it if self.dtype == numpy.float16: raise unittest.SkipTest('ChainerX does not support float16') self.check_forward(backend.to_chainerx(cuda.to_gpu(self.x)), 'never')
def test_backward_chainerx_cuda(self): self.check_backward_chainerx( backend.to_chainerx(cuda.to_gpu(self.x)), backend.to_chainerx(cuda.to_gpu(self.t)))
def test_backward_chainerx_native(self): self.check_backward_chainerx( backend.to_chainerx(self.x), backend.to_chainerx(self.t))
def test_backward_chainerx_native(self): # TODO(niboshi): Support it if self.dtype == numpy.float16: raise unittest.SkipTest('ChainerX does not support float16') self.check_backward(backend.to_chainerx(self.x), backend.to_chainerx(self.gy), 'never')
def test_double_backward_chainerx_native_nobias(self): self.check_double_backward(backend.to_chainerx(self.x), backend.to_chainerx(self.W), None, backend.to_chainerx(self.gy), backend.to_chainerx(self.ggx), backend.to_chainerx(self.ggW), None)
def test_double_backward_chainerx_native(self): self.check_double_backward(backend.to_chainerx(self.x), backend.to_chainerx(self.gy), backend.to_chainerx(self.ggx), 'never')
def test_double_backward_chainerx_native_nobias(self): self._skip_if_not_chainerx_supported() self.check_double_backward( backend.to_chainerx(self.x), backend.to_chainerx(self.W), None, backend.to_chainerx(self.gy), backend.to_chainerx(self.ggx), backend.to_chainerx(self.ggW), None)
def conv(a): return backend.to_chainerx(cuda.to_gpu(a))
def setattr(self, name, value): if isinstance(value, fallback_device.xp.ndarray): fallback_array_cache[name] = value sup.__setattr__(name, backend.to_chainerx(value)) return sup.__setattr__(name, value)
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 test_forward_chainerx_cuda(self): self.check_forward_consistency( lambda xs: backend.to_chainerx(cuda.to_gpu(xs)), nobias=False)
def test_forward_chainerx_cuda_nobias(self): self._skip_if_not_chainerx_supported() self.check_forward_consistency( lambda xs: backend.to_chainerx(cuda.to_gpu(xs)), nobias=True)
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): fallback_arr = backend.from_chainerx(st) self.state[state_name] = fallback_arr chainerx_state_arrays[state_name] = (st, fallback_arr) # 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, fallback_arr) in chainerx_state_arrays.items(): cur_arr = self.state[state_name] if cur_arr is not fallback_arr: # The optimizer altered the reference of the state, instead of # updating it in-place. We need to convert the new state back # to ChainerX. arr = backend.to_chainerx(cur_arr) self.state[state_name] = arr