rv1106，a simple net ，it work well in the conversation step， but when i run rknnmodel on rv1106， it‘s outputs is different with pc。。。。。。一个简单的分类模型，onnx转rknn及在板子上运行都没有报错但是结果就是不对 #174
this link is the net i uesd. when i convert it to rknn， and do accuracy_analysis， everything seems work well until i run it on rv1106.the results is different. what's wrong with the net?
https://github.com/HuKai97/YOLOv5-LPRNet-Licence-Recognition/blob/master/models/LPRNet.py
this link is the net i uesd. when i convert it to rknn， and do accuracy_analysis， everything seems work well until i run it on rv1106.the results is different. what's wrong with the net?
用的就是这个模型，在pc上进行onnx转rknn并测试、分析是没有问题的，但是转到板子上输出的结果也不报错，就是分类完全随机了
下面是我在pc转模型的log I rknn-toolkit2 version: 2.1.0+708089d1 --> Config model done --> Loading model I Loading : 100%|███████████████████████████████████████████████| 22/22 [00:00<00:00, 103796.05it/s] done --> Building model D base_optimize ... D base_optimize done. D D fold_constant ... D fold_constant done. D D correct_ops ... D correct_ops done. D D fuse_ops ... D fuse_ops results: D convert_squeeze_to_reshape: remove node = ['Squeeze_25'], add node = ['Squeeze_25_2rs'] D unsqueeze_to_4d_transpose: remove node = [], add node = ['onnx::Transpose_108_rs', 'output-rs'] D bypass_two_reshape: remove node = ['onnx::Transpose_108_rs', 'Squeeze_25_2rs'] D fold_constant ... D fold_constant done. D fuse_ops done. D D sparse_weight ... D sparse_weight done. D I GraphPreparing : 100%|███████████████████████████████████████████| 27/27 [00:00<00:00, 562.96it/s] I Quantizating : 100%|██████████████████████████████████████████████| 27/27 [00:00<00:00, 45.57it/s] D D quant_optimizer ... D quant_optimizer results: D adjust_relu: ['Relu_22', 'Relu_20', 'Relu_17', 'Relu_15', 'Relu_12', 'Relu_10', 'Relu_7', 'Relu_5', 'Relu_3', 'Relu_1'] D quant_optimizer done. D W build: The default input dtype of 'input' is changed from 'float32' to 'int8' in rknn model for performance! Please take care of this change when deploy rknn model with Runtime API! W build: The default output dtype of 'output' is changed from 'float32' to 'int8' in rknn model for performance! Please take care of this change when deploy rknn model with Runtime API! I rknn building ... I RKNN: [20:28:03.666] compress = 0, conv_eltwise_activation_fuse = 1, global_fuse = 1, multi-core-model-mode = 7, output_optimize = 1, layout_match = 1, enable_argb_group = 0, pipeline_fuse = 1, enable_flash_attention = 0 I RKNN: librknnc version: 2.1.0 (967d001cc8@2024-08-07T11:32:45) D RKNN: [20:28:03.667] RKNN is invoked D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNExtractCustomOpAttrs D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNExtractCustomOpAttrs D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNSetOpTargetPass D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNSetOpTargetPass D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNBindNorm D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNBindNorm D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNEliminateQATDataConvert D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNEliminateQATDataConvert D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNTileGroupConv D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNTileGroupConv D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNTileFcBatchFuse D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNTileFcBatchFuse D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNAddConvBias D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNAddConvBias D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNTileChannel D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNTileChannel D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNPerChannelPrep D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNPerChannelPrep D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNBnQuant D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNBnQuant D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNFuseOptimizerPass D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNFuseOptimizerPass D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNTurnAutoPad D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNTurnAutoPad D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNInitRNNConst D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNInitRNNConst D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNInitCastConst D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNInitCastConst D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNMultiSurfacePass D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNMultiSurfacePass D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNReplaceConstantTensorPass D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNReplaceConstantTensorPass D RKNN: [20:28:03.673] >>>>>> start: rknn::RKNNSubgraphManager D RKNN: [20:28:03.673] <<<<<<<< end: rknn::RKNNSubgraphManager D RKNN: [20:28:03.673] >>>>>> start: OpEmit D RKNN: [20:28:03.674] <<<<<<<< end: OpEmit D RKNN: [20:28:03.674] >>>>>> start: rknn::RKNNAddFirstConv D RKNN: [20:28:03.674] <<<<<<<< end: rknn::RKNNAddFirstConv D RKNN: [20:28:03.674] >>>>>> start: rknn::RKNNLayoutMatchPass D RKNN: [20:28:03.674] <<<<<<<< end: rknn::RKNNLayoutMatchPass D RKNN: [20:28:03.674] >>>>>> start: rknn::RKNNAddSecondaryNode D RKNN: [20:28:03.674] <<<<<<<< end: rknn::RKNNAddSecondaryNode D RKNN: [20:28:03.674] >>>>>> start: rknn::RKNNAllocateConvCachePass D RKNN: [20:28:03.674] <<<<<<<< end: rknn::RKNNAllocateConvCachePass D RKNN: [20:28:03.674] >>>>>> start: OpEmit D RKNN: [20:28:03.674] finish initComputeZoneMapByStepsVector D RKNN: [20:28:03.674] finish initComputeZoneMapByStepsVector D RKNN: [20:28:03.674] finish initComputeZoneMap D RKNN: [20:28:03.674] <<<<<<<< end: OpEmit D RKNN: [20:28:03.674] >>>>>> start: rknn::RKNNSubGraphMemoryPlanPass D RKNN: [20:28:03.674] <<<<<<<< end: rknn::RKNNSubGraphMemoryPlanPass D RKNN: [20:28:03.674] >>>>>> start: rknn::RKNNProfileAnalysisPass D RKNN: [20:28:03.674] <<<<<<<< end: rknn::RKNNProfileAnalysisPass D RKNN: [20:28:03.674] >>>>>> start: rknn::RKNNOperatorIdGenPass D RKNN: [20:28:03.674] <<<<<<<< end: rknn::RKNNOperatorIdGenPass D RKNN: [20:28:03.674] >>>>>> start: rknn::RKNNWeightTransposePass W RKNN: [20:28:03.676] Warning: Tensor output-rs_i1 need paramter qtype, type is set to float16 by default! W RKNN: [20:28:03.676] Warning: Tensor output-rs_i1 need paramter qtype, type is set to float16 by default! D RKNN: [20:28:03.676] <<<<<<<< end: rknn::RKNNWeightTransposePass D RKNN: [20:28:03.676] >>>>>> start: rknn::RKNNCPUWeightTransposePass D RKNN: [20:28:03.676] <<<<<<<< end: rknn::RKNNCPUWeightTransposePass D RKNN: [20:28:03.676] >>>>>> start: rknn::RKNNModelBuildPass D RKNN: [20:28:03.676] <<<<<<<< end: rknn::RKNNModelBuildPass D RKNN: [20:28:03.676] >>>>>> start: rknn::RKNNModelRegCmdbuildPass D RKNN: [20:28:03.676] -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- D RKNN: [20:28:03.676] Network Layer Information Table D RKNN: [20:28:03.676] -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- D RKNN: [20:28:03.676] ID OpType DataType Target InputShape OutputShape Cycles(DDR/NPU/Total) RW(KB) FullName D RKNN: [20:28:03.676] -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- D RKNN: [20:28:03.676] 0 InputOperator INT8 CPU \ (1,3,48,168) 0/0/0 0 InputOperator:input D RKNN: [20:28:03.676] 1 ConvRelu INT8 NPU (1,3,48,168),(8,3,5,5),(8) (1,8,44,164) 22823/180400/180400 24 Conv:Conv_0 D RKNN: [20:28:03.676] 2 ConvRelu INT8 NPU (1,8,44,164),(8,8,3,3),(8) (1,8,44,164) 37697/64944/64944 114 Conv:Conv_2 D RKNN: [20:28:03.676] 3 ConvRelu INT8 NPU (1,8,44,164),(16,8,3,3),(16) (1,16,44,164) 37884/64944/64944 115 Conv:Conv_4 D RKNN: [20:28:03.676] 4 ConvRelu INT8 NPU (1,16,44,164),(16,16,3,3),(16) (1,16,44,164) 37884/64944/64944 115 Conv:Conv_6 D RKNN: [20:28:03.676] 5 MaxPool INT8 NPU (1,16,44,164) (1,16,22,82) 0/0/0 112 MaxPool:MaxPool_8 D RKNN: [20:28:03.676] 6 ConvRelu INT8 NPU (1,16,22,82),(32,16,3,3),(32) (1,32,22,82) 14848/32544/32544 32 Conv:Conv_9 D RKNN: [20:28:03.676] 7 ConvRelu INT8 NPU (1,32,22,82),(32,32,3,3),(32) (1,32,22,82) 20283/32544/32544 65 Conv:Conv_11 D RKNN: [20:28:03.676] 8 MaxPool INT8 NPU (1,32,22,82) (1,32,11,41) 0/0/0 56 MaxPool:MaxPool_13 D RKNN: [20:28:03.676] 9 ConvRelu INT8 NPU (1,32,11,41),(48,32,3,3),(48) (1,48,11,41) 8165/12528/12528 27 Conv:Conv_14 D RKNN: [20:28:03.676] 10 ConvRelu INT8 NPU (1,48,11,41),(48,48,3,3),(48) (1,48,11,41) 10458/25056/25056 41 Conv:Conv_16 D RKNN: [20:28:03.676] 11 MaxPool INT8 NPU (1,48,11,41) (1,48,5,20) 0/0/0 21 MaxPool:MaxPool_18 D RKNN: [20:28:03.676] 12 ConvRelu INT8 NPU (1,48,5,20),(64,48,3,3),(64) (1,64,5,20) 6391/8064/8064 32 Conv:Conv_19 D RKNN: [20:28:03.676] 13 ConvRelu INT8 NPU (1,64,5,20),(128,64,3,3),(128) (1,128,5,20) 15254/16128/16128 79 Conv:Conv_21 D RKNN: [20:28:03.676] 14 MaxPool INT8 NPU (1,128,5,20) (1,128,1,21) 0/0/0 12 MaxPool:MaxPool_23 D RKNN: [20:28:03.676] 15 Conv INT8 NPU (1,128,1,21),(78,128,1,1),(78) (1,78,1,21) 2434/640/2434 13 Conv:Conv_24 D RKNN: [20:28:03.676] 16 Transpose INT8 NPU (1,78,1,21) (1,21,1,78) 0/0/0 1 Transpose:Transpose_26 D RKNN: [20:28:03.676] 17 Reshape INT8 NPU (1,21,1,78),(3) (1,21,78) 0/0/0 2 Reshape:output-rs D RKNN: [20:28:03.676] 18 OutputOperator INT8 CPU (1,21,78) \ 0/0/0 1 OutputOperator:output D RKNN: [20:28:03.676] -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- D RKNN: [20:28:03.676] <<<<<<<< end: rknn::RKNNModelRegCmdbuildPass D RKNN: [20:28:03.676] >>>>>> start: rknn::RKNNFlatcModelBuildPass D RKNN: [20:28:03.677] Export Mini RKNN model to /tmp/tmpqwtc_u17/check.rknn D RKNN: [20:28:03.677] >>>>>> end: rknn::RKNNFlatcModelBuildPass D RKNN: [20:28:03.677] >>>>>> start: rknn::RKNNMemStatisticsPass D RKNN: [20:28:03.677] ---------------------------------------------------------------------------------------------------------------------------------------- D RKNN: [20:28:03.677] Feature Tensor Information Table D RKNN: [20:28:03.677] ------------------------------------------------------------------------------------------------------+--------------------------------- D RKNN: [20:28:03.677] ID User Tensor DataType DataFormat OrigShape NativeShape | [Start End) Size D RKNN: [20:28:03.677] ------------------------------------------------------------------------------------------------------+--------------------------------- D RKNN: [20:28:03.677] 1 ConvRelu input INT8 NC1HWC2 (1,3,48,168) (1,1,48,168,3) | 0x0002e040 0x00034340 0x00006300 D RKNN: [20:28:03.677] 2 ConvRelu onnx::Conv_75 INT8 NC1HWC2 (1,8,44,164) (1,1,44,164,16) | 0x00034340 0x00050640 0x0001c300 D RKNN: [20:28:03.677] 3 ConvRelu onnx::Conv_78 INT8 NC1HWC2 (1,8,44,164) (1,1,44,164,16) | 0x00050640 0x0006c940 0x0001c300 D RKNN: [20:28:03.677] 4 ConvRelu onnx::Conv_81 INT8 NC1HWC2 (1,16,44,164) (1,1,44,164,16) | 0x0002e040 0x0004a340 0x0001c300 D RKNN: [20:28:03.677] 5 MaxPool onnx::MaxPool_84 INT8 NC1HWC2 (1,16,44,164) (1,1,44,164,16) | 0x0004a340 0x00066640 0x0001c300 D RKNN: [20:28:03.677] 6 ConvRelu input.32 INT8 NC1HWC2 (1,16,22,82) (1,1,22,82,16) | 0x0002e040 0x00035100 0x000070c0 D RKNN: [20:28:03.677] 7 ConvRelu onnx::Conv_88 INT8 NC1HWC2 (1,32,22,82) (1,2,22,82,16) | 0x00035100 0x00043280 0x0000e180 D RKNN: [20:28:03.677] 8 MaxPool onnx::MaxPool_91 INT8 NC1HWC2 (1,32,22,82) (1,2,22,82,16) | 0x00043280 0x00051400 0x0000e180 D RKNN: [20:28:03.677] 9 ConvRelu input.52 INT8 NC1HWC2 (1,32,11,41) (1,2,11,41,16) | 0x0002e040 0x000318c0 0x00003880 D RKNN: [20:28:03.677] 10 ConvRelu onnx::Conv_95 INT8 NC1HWC2 (1,48,11,41) (1,3,11,41,16) | 0x000318c0 0x00036d80 0x000054c0 D RKNN: [20:28:03.677] 11 MaxPool onnx::MaxPool_98 INT8 NC1HWC2 (1,48,11,41) (1,3,11,41,16) | 0x00036d80 0x0003c240 0x000054c0 D RKNN: [20:28:03.677] 12 ConvRelu input.72 INT8 NC1HWC2 (1,48,5,20) (1,3,5,20,16) | 0x0002e040 0x0002f300 0x000012c0 D RKNN: [20:28:03.677] 13 ConvRelu onnx::Conv_102 INT8 NC1HWC2 (1,64,5,20) (1,4,5,20,16) | 0x0002f300 0x00030c00 0x00001900 D RKNN: [20:28:03.677] 14 MaxPool onnx::MaxPool_105 INT8 NC1HWC2 (1,128,5,20) (1,8,5,20,16) | 0x00030c00 0x00033e00 0x00003200 D RKNN: [20:28:03.677] 15 Conv input.92 INT8 NC1HWC2 (1,128,1,21) (1,9,1,21,16) | 0x0002e040 0x0002ec40 0x00000c00 D RKNN: [20:28:03.677] 16 Transpose onnx::Squeeze_107 INT8 NC1HWC2 (1,78,1,21) (1,5,1,21,16) | 0x0002ec40 0x0002f3c0 0x00000780 D RKNN: [20:28:03.677] 16 Transpose onnx::Squeeze_107_exSecondary INT8 NC1HWC2 (1,78,1,21) (1,15,1,21,16) | 0x0002f3c0 0x000307c0 0x00001400 D RKNN: [20:28:03.677] 17 Reshape output-rs INT8 NC1HWC2 (1,21,1,78) (1,2,1,78,16) | 0x0002e040 0x0002ea40 0x00000a00 D RKNN: [20:28:03.677] 17 Reshape output-rs_exSecondary INT8 NC1HWC2 (1,21,1,78) (1,5,1,78,16) | 0x0002ea40 0x000304b0 0x00001a70 D RKNN: [20:28:03.677] 18 OutputOperator output INT8 UNDEFINED (1,21,78) (1,21,78) | 0x000304c0 0x00030ec0 0x00000a00 D RKNN: [20:28:03.677] ------------------------------------------------------------------------------------------------------+--------------------------------- D RKNN: [20:28:03.677] ------------------------------------------------------------------------------------- D RKNN: [20:28:03.677] Const Tensor Information Table D RKNN: [20:28:03.677] ---------------------------------------------------+--------------------------------- D RKNN: [20:28:03.677] ID User Tensor DataType OrigShape | [Start End) Size D RKNN: [20:28:03.677] ---------------------------------------------------+--------------------------------- D RKNN: [20:28:03.677] 1 ConvRelu onnx::Conv_111 INT8 (8,3,5,5) | 0x00002980 0x00002cc0 0x00000340 D RKNN: [20:28:03.677] 1 ConvRelu onnx::Conv_112 INT32 (8) | 0x00002cc0 0x00002d40 0x00000080 D RKNN: [20:28:03.677] 2 ConvRelu onnx::Conv_114 INT8 (8,8,3,3) | 0x00002d40 0x000031c0 0x00000480 D RKNN: [20:28:03.677] 2 ConvRelu onnx::Conv_115 INT32 (8) | 0x000031c0 0x00003240 0x00000080 D RKNN: [20:28:03.677] 3 ConvRelu onnx::Conv_117 INT8 (16,8,3,3) | 0x00003240 0x00003b40 0x00000900 D RKNN: [20:28:03.677] 3 ConvRelu onnx::Conv_118 INT32 (16) | 0x00003b40 0x00003bc0 0x00000080 D RKNN: [20:28:03.677] 4 ConvRelu onnx::Conv_120 INT8 (16,16,3,3) | 0x00003bc0 0x000044c0 0x00000900 D RKNN: [20:28:03.677] 4 ConvRelu onnx::Conv_121 INT32 (16) | 0x000044c0 0x00004540 0x00000080 D RKNN: [20:28:03.677] 6 ConvRelu onnx::Conv_123 INT8 (32,16,3,3) | 0x00004540 0x00005740 0x00001200 D RKNN: [20:28:03.677] 6 ConvRelu onnx::Conv_124 INT32 (32) | 0x00005740 0x00005840 0x00000100 D RKNN: [20:28:03.677] 7 ConvRelu onnx::Conv_126 INT8 (32,32,3,3) | 0x00005840 0x00007c40 0x00002400 D RKNN: [20:28:03.677] 7 ConvRelu onnx::Conv_127 INT32 (32) | 0x00007c40 0x00007d40 0x00000100 D RKNN: [20:28:03.677] 9 ConvRelu onnx::Conv_129 INT8 (48,32,3,3) | 0x00007d40 0x0000b340 0x00003600 D RKNN: [20:28:03.677] 9 ConvRelu onnx::Conv_130 INT32 (48) | 0x0000b340 0x0000b4c0 0x00000180 D RKNN: [20:28:03.677] 10 ConvRelu onnx::Conv_132 INT8 (48,48,3,3) | 0x0000b4c0 0x000105c0 0x00005100 D RKNN: [20:28:03.677] 10 ConvRelu onnx::Conv_133 INT32 (48) | 0x000105c0 0x00010740 0x00000180 D RKNN: [20:28:03.677] 12 ConvRelu onnx::Conv_135 INT8 (64,48,3,3) | 0x00010740 0x00017340 0x00006c00 D RKNN: [20:28:03.677] 12 ConvRelu onnx::Conv_136 INT32 (64) | 0x00017340 0x00017540 0x00000200 D RKNN: [20:28:03.677] 13 ConvRelu onnx::Conv_138 INT8 (128,64,3,3) | 0x00017540 0x00029540 0x00012000 D RKNN: [20:28:03.677] 13 ConvRelu onnx::Conv_139 INT32 (128) | 0x00029540 0x00029940 0x00000400 D RKNN: [20:28:03.677] 15 Conv newCnn.weight INT8 (78,128,1,1) | 0x00000000 0x00002700 0x00002700 D RKNN: [20:28:03.677] 15 Conv newCnn.bias INT32 (78) | 0x00002700 0x00002980 0x00000280 D RKNN: [20:28:03.677] 17 Reshape output-rs_i1 INT64 (3) | 0x00029940*0x00029980 0x00000040 D RKNN: [20:28:03.677] ---------------------------------------------------+--------------------------------- D RKNN: [20:28:03.677] ---------------------------------------- D RKNN: [20:28:03.677] Total Internal Memory Size: 250.25KB D RKNN: [20:28:03.677] Total Weight Memory Size: 166.375KB D RKNN: [20:28:03.677] ---------------------------------------- D RKNN: [20:28:03.677] <<<<<<<< end: rknn::RKNNMemStatisticsPass I rknn buiding done. done --> Export rknn model, ../model_pldr/lpr.rknn done --> Init runtime environment I Target is None, use simulator! done --> Running model I GraphPreparing : 100%|██████████████████████████████████████████| 29/29 [00:00<00:00, 2977.60it/s] I SessionPreparing : 100%|████████████████████████████████████████| 29/29 [00:00<00:00, 2697.66it/s] I GraphPreparing : 100%|██████████████████████████████████████████| 29/29 [00:00<00:00, 3108.64it/s] I AccuracyAnalysing : 100%|█████████████████████████████████████████| 29/29 [00:00<00:00, 97.96it/s]
simulator_error: calculate the output error of each layer of the simulator (compared to the 'golden' value).
entire: output error of each layer between 'golden' and 'simulator', these errors will accumulate layer by layer.
single: single-layer output error between 'golden' and 'simulator', can better reflect the single-layer accuracy of the simulator.
layer_name simulator_error entire single cos euc cos euc
[Input] input 1.00000 | 0.0 1.00000 | 0.0 [exDataConvert] input_int8 1.00000 | 4.3257 1.00000 | 4.3257 [Conv] input.4 [Relu] onnx::Conv_75 0.99996 | 1.2561 0.99996 | 1.2561 [Conv] input.12 [Relu] onnx::Conv_78 0.99975 | 2.6626 0.99989 | 1.7904 [Conv] input.20 [Relu] onnx::Conv_81 0.99922 | 5.5223 0.99989 | 2.0373 [Conv] input.28 [Relu] onnx::MaxPool_84 0.99944 | 7.5979 0.99997 | 1.6664 [MaxPool] input.32 0.99963 | 3.9089 0.99999 | 0.5475 [Conv] input.40 [Relu] onnx::Conv_88 0.99883 | 4.9569 0.99992 | 1.2693 [Conv] input.48 [Relu] onnx::MaxPool_91 0.99928 | 4.6993 0.99996 | 1.1559 [MaxPool] input.52 0.99948 | 2.8090 0.99999 | 0.4502 [Conv] input.60 [Relu] onnx::Conv_95 0.99930 | 2.4040 0.99995 | 0.6363 [Conv] input.68 [Relu] onnx::MaxPool_98 0.99858 | 1.8201 0.99990 | 0.4723 [MaxPool] input.72 0.99909 | 1.3663 0.99997 | 0.2254 [Conv] input.80 [Relu] onnx::Conv_102 0.99876 | 1.3209 0.99992 | 0.3279 [Conv] input.88 [Relu] onnx::MaxPool_105 0.99662 | 5.5412 0.99984 | 1.1720 [MaxPool] input.92 0.99825 | 4.3240 0.99996 | 0.5895 [Conv] onnx::Squeeze_107 0.99947 | 21.785 0.99996 | 5.1620 [Transpose] output-rs 0.99947 | 21.785 0.99997 | 4.4925 [Reshape] output_int8 0.99947 | 21.785 0.99997 | 4.4925 [exDataConvert] output 0.99947 | 21.785 0.99997 | 4.4925 I The error analysis results save to: ./snapshot/error_analysis.txt W accuracy_analysis: The mapping of layer_name & file_name save to: ./snapshot/map_name_to_file.txt [ 5.7592 12.67024 5.7592 3.0715735 6.527094 7.6789336 16.893654 9.214721 1.5357867 1.9197334 6.527094 -3.0715735 1.5357867 10.366561 -1.5357867 1.9197334 11.902348 7.6789336 0.3839467 9.598667 13.438134 ]
(1, 21, 78) [ 6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 20 0 0 0 0 0 15 15 13 14 18 1 20 0 18 18 18 16 18 15 17 8 7 19 18 20 3 18 3 17 18 18 8 0 20 16 6 18 3 18 20 3 3 18 18 16 18 18 18 17 20 9]