[DEBUG] GE(12837,python):2023-03-30-09:12:32.835.991 [shape_refiner.cc:689]12837 CreateInferenceContext:Input node[relu_bias_add], type[BiasAdd], context_map size[0]. 

[DEBUG] GE(12837,python):2023-03-30-09:12:32.835.997 [shape_refiner.cc:689]12837 CreateInferenceContext:Input node[shape_bias_add], type[BiasAdd], context_map size[0]. 

[DEBUG] GE(12837,python):2023-03-30-09:12:32.836.001 [shape_refiner.cc:689]12837 CreateInferenceContext:Input node[mul_bias_add], type[BiasAdd], context_map size[0]. 

[DEBUG] GE(12837,python):2023-03-30-09:12:32.836.006 [shape_refiner.cc:722]12837 CreateInferenceContext:Node: convolution_backward_filter, marks size: 0. 

[DEBUG] GE(12837,python):2023-03-30-09:12:32.836.010 [infershape_pass.cc:151]12837 InferShapeAndType:create context for node:convolution_backward_filter, marks 0 

[DEBUG] GE(12837,python):2023-03-30-09:12:32.836.015 [infershape_pass.cc:277]12837 CallInferShapeFunc:get shape locked flag:0 from node:convolution_backward_filter. 

[DEBUG] GE(12837,python):2023-03-30-09:12:32.836.023 [transformer_utils.cc:78]12837 CatchFormatAndShape:Node is convolution_backward_filter, input tensor idx is 0. ori format: ND, format: ND, ori shape:4,512,7,7, shape:4,512,7,7 is same! No need to catch format&shape! 

[DEBUG] GE(12837,python):2023-03-30-09:12:32.836.030 [transformer_utils.cc:78]12837 CatchFormatAndShape:Node is convolution_backward_filter, input tensor idx is 1. ori format: ND, format: ND, ori shape:4, shape:4 is same! No need to catch format&shape! 

[DEBUG] GE(12837,python):2023-03-30-09:12:32.836.037 [transformer_utils.cc:78]12837 CatchFormatAndShape:Node is convolution_backward_filter, input tensor idx is 2. ori format: ND, format: ND, ori shape:4,512,7,7, shape:4,512,7,7 is same! No need to catch format&shape! 

[DEBUG] GE(12837,python):2023-03-30-09:12:32.836.042 [transformer_utils.cc:101]12837 CatchFormatAndShape:Node is convolution_backward_filter, output tensor idx is 0. ori format: ND, format: ND, ori shape:, shape: is same!or output original not initialized. No need to catch format&shape! 

[DEBUG] OP(12837,python):2023-03-30-09:12:32.836.053 [nn_calculation_ops.cc:4361][OP_PROTO] Conv2DBackpropFilterInfer:4361 OpName:[convolution_backward_filter] "Enter Conv2DBackpropFilter inferfunction!" 

[ERROR] OP(12837,python):2023-03-30-09:12:32.836.061 [nn_calculation_ops.cc:4378][OP_PROTO] Conv2DBackpropFilterInfer:4378 OpName:[] "not found filter_format in format2str" 

[INFO] GE(12837,python):2023-03-30-09:12:32.836.085 [error_manager.cc:252]12837 ReportInterErrMessage:report error_message, error_code:E69999, work_stream_id:6 

[ERROR] GE(12837,python):2023-03-30-09:12:32.836.117 [op_desc.cc:1317]12837 CallInferFuncV1: ErrorNo: 4294967295(failed) [COMP][PRE_OPT][Call][InferFunc] for convolution_backward_filter failed. ret:4294967295 

以及尝试了把 input 数据都用 transdata 给转成 nchw 格式了，但是还是报错，构图片段代码如下

    auto convolution_backward_filter = op::Conv2DBackpropFilter("convolution_backward_filter")

        .set_input_x(relu_bias_add)

        .set_input_filter_size(shape_bias_add)

        .set_input_out_backprop(mul_bias_add)

        .set_attr_strides({1,1,1,1})

        .set_attr_pads({1,1,1,1})

        .set_attr_dilations({1,1,1,1})

        .set_attr_groups(1)

        .set_attr_data_format("NCHW");

    graph.AddOp(convolution_backward_filter);