我们从Python开源项目中，提取了以下6个代码示例，用于说明如何使用torch.backends.cudnn.version()。
def _update_output(self, input, weight, bias): self.use_cudnn = cudnn.is_acceptable(input) if self.use_cudnn and cudnn.version() < 6000: self.use_cudnn = not self.is_dilated() if self.use_cudnn: output = input.new(*self._output_size(input, weight)) if self.transposed: self._cudnn_info = ( torch._C._cudnn_convolution_transpose_full_forward( input, weight, bias, output, self.padding, self.stride, self.dilation, self.groups, cudnn.benchmark)) else: self._cudnn_info = torch._C._cudnn_convolution_full_forward( input, weight, bias, output, self.padding, self.stride, self.dilation, self.groups, cudnn.benchmark) if not self.requires_grad: del self._cudnn_info return output self._bufs = [[] for g in range(self.groups)] output = self._thnn('update_output', input, weight, bias) if not self.requires_grad: del self._bufs return output
def backward_weight(fn, input, hx, output, weight, grad_weight): with torch.cuda.device_of(input): is_input_packed = fn.batch_sizes is not None handle = cudnn.get_handle() if fn.mode == cudnn.CUDNN_LSTM: hx, cx = hx else: cx = None if fn.batch_first and not is_input_packed: input = input.transpose(0, 1) output = output.transpose(0, 1) input_size = _input_size(fn, input) hidden_size = _hidden_size(fn) if not fn.requires_grad: raise RuntimeError('backward_weight can only be called when the function requires grad!') if fn.dropout != 0 and cudnn.version() < 5103: raise RuntimeError('dropout supported only in cudnn v 5.1 and above') if tuple(input.size()) != input_size: raise RuntimeError('Expected input size {}, got {}'.format( input_size, tuple(input.size()))) if tuple(hx.size()) != hidden_size: raise RuntimeError('Expected input size {}, got {}'.format( hidden_size, hx.size())) assert hx.is_contiguous() assert cx is None or cx.is_contiguous() x = input.contiguous() y = output dw = fn.weight_buf.new().resize_as_(fn.weight_buf).zero_() check_error(cudnn.lib.cudnnRNNBackwardWeights( handle, fn.rnn_desc, fn.seq_length, fn.x_descs, ctypes.c_void_p(x.data_ptr()), fn.hx_desc, ctypes.c_void_p(hx.data_ptr()), fn.y_descs, ctypes.c_void_p(y.data_ptr()), ctypes.c_void_p(fn.workspace.data_ptr()), fn.workspace.size(0), fn.w_desc, ctypes.c_void_p(dw.data_ptr()), ctypes.c_void_p(fn.reserve.data_ptr()), fn.reserve.size(0) )) # copy the weights from the weight_buf into grad_weight grad_params = get_parameters(fn, handle, dw) _copyParams(grad_params, grad_weight) return grad_weight
def backward_weight(fn, input, hx, output, weight, grad_weight): with torch.cuda.device_of(input): is_input_packed = fn.batch_sizes is not None handle = cudnn.get_handle() if fn.mode == cudnn.CUDNN_LSTM: hx, cx = hx else: cx = None if fn.batch_first and not is_input_packed: input = input.transpose(0, 1) output = output.transpose(0, 1) input_size = _input_size(fn, input) hidden_size = _hidden_size(fn) if not fn.requires_grad: raise RuntimeError('backward_weight can only be called when the function requires grad!') if fn.dropout != 0 and cudnn.version() < 5103: raise RuntimeError('dropout supported only in cudnn v 5.1 and above') if tuple(input.size()) != input_size: raise RuntimeError('Expected input size {}, got {}'.format( input_size, tuple(input.size()))) if tuple(hx.size()) != hidden_size: raise RuntimeError('Expected input size {}, got {}'.format( hidden_size, hx.size())) assert hx.is_contiguous() assert cx is None or cx.is_contiguous() x = input.contiguous() y = output dw = fn.weight_buf.new().resize_as_(fn.weight_buf).zero_() with torch.cuda.device_of(input): workspace = torch.cuda.ByteTensor(fn.workspace_size) check_error(cudnn.lib.cudnnRNNBackwardWeights( handle, fn.rnn_desc, fn.seq_length, fn.x_descs, ctypes.c_void_p(x.data_ptr()), fn.hx_desc, ctypes.c_void_p(hx.data_ptr()), fn.y_descs, ctypes.c_void_p(y.data_ptr()), ctypes.c_void_p(workspace.data_ptr()), workspace.size(0), fn.w_desc, ctypes.c_void_p(dw.data_ptr()), ctypes.c_void_p(fn.reserve.data_ptr()), fn.reserve.size(0) )) # copy the weights from the weight_buf into grad_weight grad_params = get_parameters(fn, handle, dw) _copyParams(grad_params, grad_weight) return grad_weight
