我们从Python开源项目中，提取了以下15个代码示例，用于说明如何使用chainer.cuda.cudnn()。
def forward_gpu(self, x): # Implementation using cudnn x = x[0] n, c, h, w = x.shape y_h = conv.get_conv_outsize( h, self.kh, self.sy, self.ph, self.cover_all) y_w = conv.get_conv_outsize( w, self.kw, self.sx, self.pw, self.cover_all) y = cuda.cupy.empty((n, c, y_h, y_w), dtype=x.dtype) handle = cudnn.get_handle() pool_desc = self.create_pool_desc() x_desc = cudnn.create_tensor_descriptor(x) y_desc = cudnn.create_tensor_descriptor(y) oz_dtype = 'd' if x.dtype == 'd' else 'f' one = numpy.array(1, dtype=oz_dtype).ctypes zero = numpy.array(0, dtype=oz_dtype).ctypes libcudnn.poolingForward( handle, pool_desc.value, one.data, x_desc.value, x.data.ptr, zero.data, y_desc.value, y.data.ptr) self.y = y return y,
def backward_gpu(self, x, gy): # Implementation using cudnn x = x[0] handle = cudnn.get_handle() pool_desc = self.create_pool_desc() # Pooling of cuDNNv2 does not seem to support non-contiguous gradients gy = cuda.cupy.ascontiguousarray(gy[0]) x_desc = cudnn.create_tensor_descriptor(x) y_desc = cudnn.create_tensor_descriptor(gy) oz_dtype = 'd' if x.dtype == 'd' else 'f' one = numpy.array(1, dtype=oz_dtype).ctypes zero = numpy.array(0, dtype=oz_dtype).ctypes gx = cuda.cupy.empty_like(x) libcudnn.poolingBackward( handle, pool_desc.value, one.data, y_desc.value, self.y.data.ptr, y_desc.value, gy.data.ptr, x_desc.value, x.data.ptr, zero.data, x_desc.value, gx.data.ptr) return gx,
def forward(self, xs): x = xs[0] xp = cuda.get_array_module(x) if (xp != numpy and cuda.cudnn_enabled and self.use_cudnn and _cudnn_version >= 3000): oz_dtype = 'd' if x.dtype == 'd' else 'f' one = numpy.array(1, dtype=oz_dtype).ctypes zero = numpy.array(0, dtype=oz_dtype).ctypes handle = cudnn.get_handle() x_cube = x.reshape(x.shape[:2] + (-1, 1)) desc = cudnn.create_tensor_descriptor(x_cube) self.y = xp.empty_like(x) libcudnn.softmaxForward( handle, _algorithm, _mode, one.data, desc.value, x_cube.data.ptr, zero.data, desc.value, self.y.data.ptr) return self.y, else: log_z = logsumexp(x) self.y = x - log_z return self.y,
def forward(self, x): xp = cuda.get_array_module(*x) if (xp != numpy and cuda.cudnn_enabled and self.use_cudnn and (_cudnn_version >= 3000 or x[0].dtype != numpy.float16)): oz_dtype = 'd' if x[0].dtype == 'd' else 'f' one = numpy.array(1, dtype=oz_dtype).ctypes zero = numpy.array(0, dtype=oz_dtype).ctypes handle = cudnn.get_handle() x_cube = x[0].reshape(x[0].shape[:2] + (-1, 1)) desc = cudnn.create_tensor_descriptor(x_cube) self.y = xp.empty_like(x[0]) libcudnn.softmaxForward( handle, _algorithm, _mode, one.data, desc.value, x_cube.data.ptr, zero.data, desc.value, self.y.data.ptr) else: self.y = x[0] - x[0].max(axis=1, keepdims=True) xp.exp(self.y, out=self.y) self.y /= self.y.sum(axis=1, keepdims=True) return self.y,
def backward(self, x, gy): xp = cuda.get_array_module(*x) if (xp != numpy and cuda.cudnn_enabled and self.use_cudnn and (_cudnn_version >= 3000 or x[0].dtype != numpy.float16)): oz_dtype = 'd' if x[0].dtype == 'd' else 'f' one = numpy.array(1, dtype=oz_dtype).ctypes zero = numpy.array(0, dtype=oz_dtype).ctypes handle = cudnn.get_handle() gx = xp.empty_like(x[0]) gx_cube = gx.reshape(gx.shape[:2] + (-1, 1)) desc = cudnn.create_tensor_descriptor(gx_cube) libcudnn.softmaxBackward( handle, _algorithm, _mode, one.data, desc.value, self.y.data.ptr, desc.value, gy[0].data.ptr, zero.data, desc.value, gx.data.ptr) else: gx = self.y * gy[0] sumdx = gx.sum(axis=1, keepdims=True) gx -= self.y * sumdx return gx,
def softmax_log(x, use_cudnn): xp = cuda.get_array_module(x) if (xp != numpy and cuda.cudnn_enabled and use_cudnn and _cudnn_version >= 3000): oz_dtype = 'd' if x.dtype == 'd' else 'f' one = numpy.array(1, dtype=oz_dtype).ctypes zero = numpy.array(0, dtype=oz_dtype).ctypes handle = cudnn.get_handle() x_cube = x.reshape(x.shape[:2] + (-1, 1)) desc = cudnn.create_tensor_descriptor(x_cube) y = xp.empty_like(x) libcudnn.softmaxForward( handle, _algorithm, _mode, one.data, desc.value, x_cube.data.ptr, zero.data, desc.value, y.data.ptr) return y else: log_z = logsumexp(xp, x) return x - log_z
def _make_tensor_descriptor_array(xs, length): """Make an array of pointers denoting pointers of tensor descriptors. """ descs = [] batch_size = xs.shape[0] // length for i in range(length): x = xs[i*batch_size:(i+1)*batch_size] if x.ndim < 3: shape = x.shape + (1,) * (3 - x.ndim) x = x.reshape(shape) desc = cudnn.create_tensor_nd_descriptor(x) descs.append(desc) return PointerArray([d.value for d in descs], descs)
def create_pool_desc(self): return cudnn.create_pooling_descriptor( (self.kh, self.kw), (self.sy, self.sx), (self.ph, self.pw), libcudnn.CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING)
def create_pool_desc(self): return cudnn.create_pooling_descriptor( (self.kh, self.kw), (self.sy, self.sx), (self.ph, self.pw), libcudnn.CUDNN_POOLING_MAX)
def forward_gpu(self, x): if (cuda.cudnn_enabled and self.use_cudnn and (_cudnn_version >= 3000 or x[0].dtype != numpy.float16)): y = cudnn.activation_forward(x[0], _mode) self.y = y else: y = cuda.cupy.maximum(x[0], 0) return y,
def forward_gpu(self, inputs): x = inputs[0] if (cuda.cudnn_enabled and self.use_cudnn and (_cudnn_version >= 3000 or x.dtype != numpy.float16)): self.y = cuda.cupy.cudnn.activation_forward(x, _mode) else: self.y = cuda.elementwise( 'T x', 'T y', 'y = 1 / (1 + exp(-x))', 'sigmoid_fwd')(x) return self.y,
def backward_gpu(self, inputs, grads): x = inputs[0] gy = grads[0] if (cuda.cudnn_enabled and self.use_cudnn and (_cudnn_version >= 3000 or x.dtype != numpy.float16)): gx = cuda.cupy.cudnn.activation_backward(x, self.y, gy, _mode) else: gx = cuda.elementwise( 'T y, T gy', 'T gx', 'gx = gy * y * (1 - y)', 'sigmoid_bwd')(self.y, gy) return gx,
def forward_gpu(self, x): if (cuda.cudnn_enabled and self.use_cudnn and (_cudnn_version >= 3000 or x[0].dtype != numpy.float16)): self.y = cudnn.activation_forward(x[0], _mode) else: self.y = cuda.cupy.empty_like(x[0]) cuda.cupy.tanh(x[0], out=self.y) return self.y,
def backward_gpu(self, x, gy): if (cuda.cudnn_enabled and self.use_cudnn and (_cudnn_version >= 3000 or x[0].dtype != numpy.float16)): gx = cudnn.activation_backward(x[0], self.y, gy[0], _mode) else: gx = cuda.elementwise( 'T y, T gy', 'T gx', 'gx = gy * (1 - y * y)', 'tanh_bwd')(self.y, gy[0]) return gx,
