diff --git a/docs/inference_model_convertor/op_list.md b/docs/inference_model_convertor/op_list.md index 497c9a2f036eaea6ce24238113dc6e3d4c83ee92..ccc301d8c6d36b56be49788632fac323b9c9ac93 100755 --- a/docs/inference_model_convertor/op_list.md +++ b/docs/inference_model_convertor/op_list.md @@ -117,7 +117,8 @@ Aten: | 125 | aten::complex | 126 | aten::real | 127 | aten::imag | 128 | aten::fft\_rfftn | | 129 | aten::fft\_irfftn | 130 | aten::hardsigmoid | 131 | aten::hardswish | 132 | aten::linear | | 133 | aten::rsqrt | 134 | aten::replication\_pad1d | 135 | aten::full | 136 | aten::group\_norm | -| 137 | aten::argmax | 138 | aten::copy | 139 | aten::upsample\_trilinear3d | | | +| 137 | aten::argmax | 138 | aten::copy | 139 | aten::upsample\_trilinear3d | 140 | aten::clone | +| 141 | aten::rand | 142 | aten::randn | | | | | Prim: | 序号 | OP | 序号 | OP | 序号 | OP | 序号 | OP | diff --git a/x2paddle/op_mapper/pytorch2paddle/aten.py b/x2paddle/op_mapper/pytorch2paddle/aten.py index 0212f3deff5958f0452529f6a7dd81076a2a20d6..b393edf68bf6d1c64eab1cccb341b88c832470d1 100755 --- a/x2paddle/op_mapper/pytorch2paddle/aten.py +++ b/x2paddle/op_mapper/pytorch2paddle/aten.py @@ -1087,6 +1087,36 @@ def aten_clamp_min(mapper, graph, node): return current_inputs, current_outputs +def aten_clone(mapper, graph, node): + """ + TorchScript Code: + %55 : Tensor = aten::clone(%54) + Parameter meaning: + %55 (Tensor): output tensor + %54 (Tensor): input tensor + """ + scope_name = mapper.normalize_scope_name(node) + output_name = mapper._get_outputs_name(node)[0] + layer_outputs = [output_name] + layer_inputs = {} + inputs_name, inputs_node = mapper._get_inputs_name(node) + # outputs list + current_outputs = [output_name] + # inputs list + mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs, + scope_name) + layer_inputs["x"] = inputs_name[0] + + current_inputs = list(layer_inputs.values()) + + graph.add_layer( + "paddle.clone", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name) + return current_inputs, current_outputs + + def aten_complex(mapper, graph, node): """ TorchScript示例: @@ -1347,123 +1377,191 @@ def aten_conv2d(mapper, graph, node): def aten__convolution(mapper, graph, node): - """ 构造conv2d的PaddleLayer。 - TorchScript示例: + """ + TorchScript Code: %input.10 : Tensor = aten::_convolution(%input.1, %18, %10, %19, %20, %21, %13, %22, %12, %13, %13, %15) - 参数含义: - %input.10 (Tensor): 输出，卷积后的结果。 - %input.8 (Tensor): 需要进行卷积的特征层。 - %18 (Tensor): weights。 - %10 (Tensor): bias。 - %19 (list): 步长大小。 - %20 (list): 填充大小。 - %21 (list): 空洞大小。 - %13 (bool): 是否进行转置卷积。 - %22 (list): 输出形状上一侧额外添加的大小。 - %12 (int): 卷积的组数。 + Parameter meaning: + %input.10 (Tensor): Output Tensor + %input.8 (Tensor): Input Tensor + %18 (Tensor): weights + %10 (Tensor): bias + %19 (list): stride + %20 (list): padding + %21 (list): dilation + %13 (bool): whether transpose + %22 (list): output_padding + %12 (int): groups """ scope_name = mapper.normalize_scope_name(node) inputs_name, inputs_node = mapper._get_inputs_name(node) - weights = mapper.pytorch_params[inputs_name[1]] - if len(weights.shape) == 3: - op_name = name_generator("conv1d", mapper.nn_name2id) - elif len(weights.shape) == 4: - op_name = name_generator("conv2d", mapper.nn_name2id) - else: - op_name = name_generator("conv3d", mapper.nn_name2id) - output_name = mapper._get_outputs_name(node)[0] - layer_outputs = [op_name, output_name] - layer_inputs = {} layer_attrs = {} - # 获取当前节点输出的list - current_outputs = [output_name] - # 处理输入0，即%input.8 - mapper._check_input(graph, inputs_node[0], inputs_name[0], current_outputs, - scope_name) - layer_inputs["input"] = inputs_name[0] - # 获取当前节点输入的list - current_inputs = list(layer_inputs.values()) - # 处理输入1，即%18 - mapper.paddle_params[op_name + - ".weight"] = weights #np.swapaxes(weights, 0, 1) - if mapper.attrs[inputs_name[6]]: - layer_attrs["out_channels"] = weights.shape[1] - else: - layer_attrs["out_channels"] = weights.shape[0] - layer_attrs["kernel_size"] = weights.shape[2:] - # 处理输入2，即%10 - if inputs_name[2] in mapper.pytorch_params: - bias = mapper.pytorch_params[inputs_name[2]] - if bias is not None: - mapper.paddle_params[op_name + ".bias"] = bias - else: - layer_attrs["bias_attr"] = False - else: - layer_attrs["bias_attr"] = False - # 处理输入3，即%19 + # deal with stride layer_attrs["stride"] = mapper.attrs[inputs_name[3]] - # 处理输入4，即%20 + # deal with padding layer_attrs["padding"] = mapper.attrs[inputs_name[4]] - # 处理输入5，即%21 + # deal with dilation layer_attrs["dilation"] = mapper.attrs[inputs_name[5]] - # 处理输入6，即%13 - if mapper.attrs[inputs_name[6]]: - # 处理输入7，即%22 - layer_attrs["output_padding"] = mapper.attrs[inputs_name[7]] - # 处理输入8，即%12 + # deal with groups layer_attrs["groups"] = mapper.attrs[inputs_name[8]] - if mapper.attrs[inputs_name[6]]: - layer_attrs['in_channels'] = weights.shape[0] * mapper.attrs[ - inputs_name[8]] + # for weight is nn.functional.conv2d input + if inputs_name[1] not in mapper.pytorch_params: + output_name = mapper._get_outputs_name(node)[0] + layer_outputs = [output_name] + layer_inputs = {} + current_outputs = [output_name] + # input + mapper._check_input(graph, inputs_node[0], inputs_name[0], + current_outputs, scope_name) + layer_inputs["x"] = inputs_name[0] + # weight + mapper._check_input(graph, inputs_node[1], inputs_name[1], + current_outputs, scope_name) + layer_inputs["weight"] = inputs_name[1] + layer_attrs["bias"] = mapper.attrs[inputs_name[2]] + current_inputs = list(layer_inputs.values()) + # Determine whether it is conv or convtranspose according to the attribute + if len(layer_attrs["stride"]) == 1: + if mapper.attrs[inputs_name[6]]: + # only convtranspose have output_padding attr + layer_attrs["output_padding"] = mapper.attrs[inputs_name[7]] + graph.add_layer( + "paddle.nn.functional.conv1d_transpose", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + else: + graph.add_layer( + "paddle.nn.functional.conv1d", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + elif len(layer_attrs["stride"]) == 2: + if mapper.attrs[inputs_name[6]]: + # only convtranspose have output_padding attr + layer_attrs["output_padding"] = mapper.attrs[inputs_name[7]] + graph.add_layer( + "paddle.nn.functional.conv2d_transpose", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + else: + graph.add_layer( + "paddle.nn.functional.conv2d", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + elif len(layer_attrs["stride"]) == 3: + if mapper.attrs[inputs_name[6]]: + # only convtranspose have output_padding attr + layer_attrs["output_padding"] = mapper.attrs[inputs_name[7]] + graph.add_layer( + "paddle.nn.functional.conv3d_transpose", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + else: + graph.add_layer( + "paddle.nn.functional.conv3d", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + return current_inputs, current_outputs else: - layer_attrs['in_channels'] = weights.shape[1] * mapper.attrs[ - inputs_name[8]] - if len(weights.shape) == 3: - if mapper.attrs[inputs_name[6]]: - graph.add_layer( - "paddle.nn.Conv1DTranspose", - inputs=layer_inputs, - outputs=layer_outputs, - scope_name=scope_name, - **layer_attrs) + weights = mapper.pytorch_params[inputs_name[1]] + if len(weights.shape) == 3: + op_name = name_generator("conv1d", mapper.nn_name2id) + elif len(weights.shape) == 4: + op_name = name_generator("conv2d", mapper.nn_name2id) else: - graph.add_layer( - "paddle.nn.Conv1D", - inputs=layer_inputs, - outputs=layer_outputs, - scope_name=scope_name, - **layer_attrs) - elif len(weights.shape) == 4: + op_name = name_generator("conv3d", mapper.nn_name2id) + output_name = mapper._get_outputs_name(node)[0] + layer_outputs = [op_name, output_name] + layer_inputs = {} + current_outputs = [output_name] + mapper._check_input(graph, inputs_node[0], inputs_name[0], + current_outputs, scope_name) + layer_inputs["input"] = inputs_name[0] + current_inputs = list(layer_inputs.values()) + mapper.paddle_params[op_name + + ".weight"] = weights #np.swapaxes(weights, 0, 1) if mapper.attrs[inputs_name[6]]: - graph.add_layer( - "paddle.nn.Conv2DTranspose", - inputs=layer_inputs, - outputs=layer_outputs, - scope_name=scope_name, - **layer_attrs) + layer_attrs["out_channels"] = weights.shape[1] else: - graph.add_layer( - "paddle.nn.Conv2D", - inputs=layer_inputs, - outputs=layer_outputs, - scope_name=scope_name, - **layer_attrs) - else: + layer_attrs["out_channels"] = weights.shape[0] + layer_attrs["kernel_size"] = weights.shape[2:] + # deal with bias + if inputs_name[2] in mapper.pytorch_params: + bias = mapper.pytorch_params[inputs_name[2]] + if bias is not None: + mapper.paddle_params[op_name + ".bias"] = bias + else: + layer_attrs["bias_attr"] = False + else: + layer_attrs["bias_attr"] = False if mapper.attrs[inputs_name[6]]: - graph.add_layer( - "paddle.nn.Conv3DTranspose", - inputs=layer_inputs, - outputs=layer_outputs, - scope_name=scope_name, - **layer_attrs) + layer_attrs['in_channels'] = weights.shape[0] * mapper.attrs[ + inputs_name[8]] else: - graph.add_layer( - "paddle.nn.Conv3D", - inputs=layer_inputs, - outputs=layer_outputs, - scope_name=scope_name, - **layer_attrs) - return current_inputs, current_outputs + layer_attrs['in_channels'] = weights.shape[1] * mapper.attrs[ + inputs_name[8]] + if len(weights.shape) == 3: + if mapper.attrs[inputs_name[6]]: + # only convtranspose have output_padding attr + layer_attrs["output_padding"] = mapper.attrs[inputs_name[7]] + graph.add_layer( + "paddle.nn.Conv1DTranspose", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + else: + graph.add_layer( + "paddle.nn.Conv1D", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + elif len(weights.shape) == 4: + if mapper.attrs[inputs_name[6]]: + # only convtranspose have output_padding attr + layer_attrs["output_padding"] = mapper.attrs[inputs_name[7]] + graph.add_layer( + "paddle.nn.Conv2DTranspose", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + else: + graph.add_layer( + "paddle.nn.Conv2D", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + else: + if mapper.attrs[inputs_name[6]]: + # only convtranspose have output_padding attr + layer_attrs["output_padding"] = mapper.attrs[inputs_name[7]] + graph.add_layer( + "paddle.nn.Conv3DTranspose", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + else: + graph.add_layer( + "paddle.nn.Conv3D", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + return current_inputs, current_outputs def aten_conv_transpose2d(mapper, graph, node): @@ -4385,6 +4483,88 @@ def aten_prelu(mapper, graph, node): return current_inputs, current_outputs +def aten_rand(mapper, graph, node): + """ + TorchScript Code: + %input.49 : Tensor = aten::rand(%23, %8, %6, %24, %5) + Parameter meaning: + %input.49 (Tensor): output tensor + %23 (list): input shape list + %8 (int): dtype。 + %6 (int): layout。 + %4995 (int): device + %4995 (bool): requires_grad + """ + scope_name = mapper.normalize_scope_name(node) + output_name = mapper._get_outputs_name(node)[0] + layer_outputs = [output_name] + layer_inputs = {} + layer_attrs = {} + inputs_name, inputs_node = mapper._get_inputs_name(node) + # 获取当前节点输出的list + current_outputs = [output_name] + current_inputs = [] + # 处理输入0，即%23，代表end + if inputs_name[0] in mapper.attrs: + layer_attrs["shape"] = mapper.attrs[inputs_name[0]] + else: + mapper._check_input(graph, inputs_node[0], inputs_name[0], + current_outputs, scope_name) + layer_inputs["shape"] = inputs_name[0] + current_inputs.append(inputs_name[0]) + # 处理输入1，即%8，代表dtype + layer_attrs["dtype"] = dtype_dict[mapper.attrs[inputs_name[1]]] + + graph.add_layer( + "paddle.rand", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + return current_inputs, current_outputs + + +def aten_randn(mapper, graph, node): + """ + TorchScript Code: + %input.49 : Tensor = aten::randn(%23, %8, %6, %24, %5) + Parameter meaning: + %input.49 (Tensor): output tensor + %23 (list): input shape list + %8 (int): dtype。 + %6 (int): layout。 + %4995 (int): device + %4995 (bool): requires_grad + """ + scope_name = mapper.normalize_scope_name(node) + output_name = mapper._get_outputs_name(node)[0] + layer_outputs = [output_name] + layer_inputs = {} + layer_attrs = {} + inputs_name, inputs_node = mapper._get_inputs_name(node) + # 获取当前节点输出的list + current_outputs = [output_name] + current_inputs = [] + # 处理输入0，即%23，代表end + if inputs_name[0] in mapper.attrs: + layer_attrs["shape"] = mapper.attrs[inputs_name[0]] + else: + mapper._check_input(graph, inputs_node[0], inputs_name[0], + current_outputs, scope_name) + layer_inputs["shape"] = inputs_name[0] + current_inputs.append(inputs_name[0]) + # 处理输入1，即%8，代表dtype + layer_attrs["dtype"] = dtype_dict[mapper.attrs[inputs_name[1]]] + + graph.add_layer( + "paddle.randn", + inputs=layer_inputs, + outputs=layer_outputs, + scope_name=scope_name, + **layer_attrs) + return current_inputs, current_outputs + + def aten_real(mapper, graph, node): """ TorchScript示例: