rdk x5使用model_zoo 量化部署yolo26，几乎无法使用

VelumZhang · 2026 年4 月 30 日 05:51

我训练的yolo26 模型 map>0.9
按照model_zoo操作导出到rdk x5后几乎无法识别

以下是我的操作步骤

参考rdk_model_zoo\samples\vision\ultralytics_yolo26\conversion\README_cn.md
使用export_yolo26_detect_bpu.py 导出 onnx模型

Loading model: ./best26n.pt...
Applying BPU Monkey Patch (Output Layout: NHWC)...
Starting export (imgsz=640)...
Ultralytics 8.4.2  Python-3.10.19 torch-2.9.0+cu128 CPU (Intel Core Ultra 9 275HX)
YOLO26n summary (fused): 122 layers, 2,375,226 parameters, 0 gradients, 5.2 GFLOPs

PyTorch: starting from 'best26n.pt' with input shape (1, 3, 640, 640) BCHW and output shape(s) ((1, 80, 80, 2), (1, 80, 80, 4), (1, 40, 40, 2), (1, 40, 40, 4), (1, 20, 20, 2), (1, 20, 20, 4)) (5.1 MB)

ONNX: starting export with onnx 1.19.1 opset 11...
ONNX: slimming with onnxslim 0.1.80...
ONNX: export success  0.9s, saved as 'best26n.onnx' (9.2 MB)

Export complete (1.2s)
Results saved to E:\RDK_bak\rdk_model_zoo-main
Predict:         yolo predict task=detect model=best26n.onnx imgsz=640  
Validate:        yolo val task=detect model=best26n.onnx imgsz=640 data=E:\jetson_bak\Desktop\Detector_linux_dev\datasets\55555_5555\data.yaml  
Visualize:       https://netron.app
Renamed output to: ./best26n.onnx

准备校准数据集

文件结构
/cal/
     frame001.jpg
     frame002.jpg
     frame003.jpg
     frame004.jpg
   ....

使用Linux 虚拟机量化模型

命令：

python samples/vision/ultralytics_yolo26/conversion/mapper.py --onnx ./best26n.onnx --cal-images ./cal/ --cal-sample-num 50

输出log 节选

         [  0.,   0.,   0., ...,   0.,   0.,   0.]]]], dtype=float32)]}}, 'hbdk_dict': {'hbdk_pass_through_params': '--O3 --debug --core-num 1 --fast --jobs 16 ', 'input-source': {'images': 'pyramid', '_default_value': 'ddr'}}, 'node_dict': {}, 'check_mode': False, 'optimization': ['set_Softmax_input_int8,set_Softmax_output_int8']}
2026-04-30 13:07:54,165 file: model_builder.py func: model_builder line No: 35 Start to Horizon NN Model Convert.
2026-04-30 13:07:54,193 file: model_debugger.py func: model_debugger line No: 67 Loading horizon_nn debug methods:set()
2026-04-30 13:07:54,194 file: quantization_config.py func: quantization_config line No: 305 The activation calibration parameters:
    calibration_type:     ['max', 'kl']
    max_percentile:       [0.99995, 1.0]
    per_channel:          [True, False]
    asymmetric:           [True, False]
The modelwise search parameters:
    similarity:           0.995
    metric:               cosine-similarity
The input of all Softmax nodes are set to : int8
2026-04-30 13:07:54,194 file: input_dict_parser.py func: input_dict_parser line No: 240 input images is from pyramid. Its layout is set to NHWC
2026-04-30 13:07:54,194 file: model_builder.py func: model_builder line No: 197 The specified model compilation architecture: bayes-e.
2026-04-30 13:07:54,194 file: model_builder.py func: model_builder line No: 207 The specified model compilation optimization parameters: [].
2026-04-30 13:07:54,194 file: model_builder.py func: model_builder line No: 35 Start to prepare the onnx model.
2026-04-30 13:07:54,215 file: prepare.py func: prepare line No: 106 Input ONNX Model Information:
ONNX IR version:          6
Opset version:            ['ai.onnx v11', 'horizon v1']
Producer:                 pytorch v2.9.0
Domain:                   None
Version:                  None
Graph input:
    images:               shape=[1, 3, 640, 640], dtype=FLOAT32
Graph output:
    output0:              shape=[1, 80, 80, 2], dtype=FLOAT32
    580:                  shape=[1, 80, 80, 4], dtype=FLOAT32
    594:                  shape=[1, 40, 40, 2], dtype=FLOAT32
    602:                  shape=[1, 40, 40, 4], dtype=FLOAT32
    616:                  shape=[1, 20, 20, 2], dtype=FLOAT32
    624:                  shape=[1, 20, 20, 4], dtype=FLOAT32
2026-04-30 13:07:54,573 file: model_builder.py func: model_builder line No: 38 End to prepare the onnx model.
2026-04-30 13:07:54,617 file: model_builder.py func: model_builder line No: 265 Saving model to: best26n_bayese_640x640_nv12_original_float_model.onnx.
2026-04-30 13:07:54,617 file: model_builder.py func: model_builder line No: 35 Start to optimize the onnx model.
2026-04-30 13:07:54,858 file: constant_folding.py func: constant_folding line No: 66 Summary info for constant_folding:
2026-04-30 13:07:54,858 file: constant_folding.py func: constant_folding line No: 67   After constant_folding, the number of nodes has changed from 356 to 356.
2026-04-30 13:07:54,858 file: constant_folding.py func: constant_folding line No: 71   After constant_folding, the number of parameters has changed from 2375251 to 2375251.
2026-04-30 13:07:54,858 file: constant_folding.py func: constant_folding line No: 76 Detailed info for constant_folding:
2026-04-30 13:07:54,858 file: constant_folding.py func: constant_folding line No: 88 
2026-04-30 13:07:55,134 file: model_builder.py func: model_builder line No: 38 End to optimize the onnx model.
2026-04-30 13:07:55,175 file: model_builder.py func: model_builder line No: 265 Saving model to: best26n_bayese_640x640_nv12_optimized_float_model.onnx.
2026-04-30 13:07:55,175 file: model_builder.py func: model_builder line No: 35 Start to calibrate the model.
2026-04-30 13:07:55,347 file: calibration_data_set.py func: calibration_data_set line No: 111 input name: images,  number_of_samples: 100
2026-04-30 13:07:55,347 file: calibration_data_set.py func: calibration_data_set line No: 123 There are 100 samples in the data set.
2026-04-30 13:07:55,347 file: infer_thresholds.py func: infer_thresholds line No: 84 Run calibration model with modelwise search method.
2026-04-30 13:07:55,702 file: base.py func: base line No: 138 Calibration using batch 8
2026-04-30 13:07:58,583 file: ort.py func: ort line No: 207 Reset batch_size=1 and execute forward again...
2026-04-30 13:10:50,122 file: modelwise_search.py func: modelwise_search line No: 75 Select max-percentile:percentile=0.99995 method.
2026-04-30 13:10:50,725 file: model_builder.py func: model_builder line No: 38 End to calibrate the model.
2026-04-30 13:10:50,791 file: model_builder.py func: model_builder line No: 265 Saving model to: best26n_bayese_640x640_nv12_calibrated_model.onnx.
2026-04-30 13:10:50,791 file: model_builder.py func: model_builder line No: 35 Start to quantize the model.
2026-04-30 13:10:51,794 file: constant_folding.py func: constant_folding line No: 66 Summary info for constant_folding:
2026-04-30 13:10:51,794 file: constant_folding.py func: constant_folding line No: 67   After constant_folding, the number of nodes has changed from 308 to 308.
2026-04-30 13:10:51,794 file: constant_folding.py func: constant_folding line No: 71   After constant_folding, the number of parameters has changed from 2406969 to 2406969.
2026-04-30 13:10:51,794 file: constant_folding.py func: constant_folding line No: 76 Detailed info for constant_folding:
2026-04-30 13:10:51,794 file: constant_folding.py func: constant_folding line No: 88 
2026-04-30 13:10:51,881 file: model_builder.py func: model_builder line No: 38 End to quantize the model.
2026-04-30 13:10:51,916 file: model_builder.py func: model_builder line No: 265 Saving model to: best26n_bayese_640x640_nv12_quantized_model.onnx.
2026-04-30 13:10:51,916 file: model_builder.py func: model_builder line No: 35 Start to compile the model with march bayes-e.
2026-04-30 13:10:52,282 file: hybrid_build.py func: hybrid_build line No: 111 Compile submodel: main_graph_subgraph_0
2026-04-30 13:10:52,307 file: hbdk_cc.py func: hbdk_cc line No: 126 hbdk-cc parameters:['--O3', '--debug', '--core-num', '1', '--fast', '--jobs', '16', '--input-layout', 'NHWC', '--output-layout', 'NHWC', '--input-source', 'pyramid']
2026-04-30 13:10:52,308 file: hbdk_cc.py func: hbdk_cc line No: 127 hbdk-cc command used:hbdk-cc -f hbir -m /tmp/tmpev6mkwb1/main_graph_subgraph_0.hbir -o /tmp/tmpev6mkwb1/main_graph_subgraph_0.hbm --march bayes-e --progressbar --O3 --debug --core-num 1 --fast --jobs 16 --input-layout NHWC --output-layout NHWC --input-source pyramid
2026-04-30 13:13:41,690 file: tool_utils.py func: tool_utils line No: 326 consumed time 169.33
2026-04-30 13:13:41,793 file: tool_utils.py func: tool_utils line No: 326 FPS=112.43, latency = 8894.5 us, DDR = 23388720 bytes   (see main_graph_subgraph_0.html)
2026-04-30 13:13:41,870 file: model_builder.py func: model_builder line No: 38 End to compile the model with march bayes-e.
2026-04-30 13:13:44,163 file: print_info_dict.py func: print_info_dict line No: 72 The main quantized node information:
======================================================================================================================================
Node                                                ON   Subgraph  Type                       Cosine Similarity  Threshold  DataType  
--------------------------------------------------------------------------------------------------------------------------------------
HZ_PREPROCESS_FOR_images                            BPU  id(0)     HzSQuantizedPreprocess     0.999977           127.0      int8      
/model.0/conv/Conv                                  BPU  id(0)     HzSQuantizedConv           0.999582           1.0        int8      
/model.0/act/Mul                                    BPU  id(0)     HzLut                      0.998928           49.2831    int8      
/model.1/conv/Conv                                  BPU  id(0)     HzSQuantizedConv           0.993233           48.177     int8      
/model.1/act/Mul                                    BPU  id(0)     HzLut                      0.994739           47.8601    int8      
/model.2/cv1/conv/Conv                              BPU  id(0)     HzSQuantizedConv           0.991292           41.2605    int8      
/model.2/cv1/act/Mul                                BPU  id(0)     HzLut                      0.992176           50.9909    int8      
/model.2/Split                                      BPU  id(0)     Split                      0.992260           33.6133    int8      
/model.2/m.0/cv1/conv/Conv                          BPU  id(0)     HzSQuantizedConv           0.994036           33.6133    int8      
/model.2/m.0/cv1/act/Mul                            BPU  id(0)     HzLut                      0.994845           4.23858    int8      
/model.2/m.0/cv2/conv/Conv                          BPU  id(0)     HzSQuantizedConv           0.978182           0.918392   int8      
/model.2/m.0/cv2/act/Mul                            BPU  id(0)     HzLut                      0.983361           27.1908    int8      
/model.2/m.0/Add                                    BPU  id(0)     HzSElementwiseAdd          0.989416           33.6133    int8      
/model.2/Split_output_0_calibrated_Requantize       BPU  id(0)     HzRequantize               --                 --         int8      
/model.2/Split_output_1_calibrated_Requantize       BPU  id(0)     HzRequantize               --                 --         int8      
/model.2/Concat                                     BPU  id(0)     Concat                     0.990323           33.6133    int8      
/model.2/cv2/conv/Conv                              BPU  id(0)     HzSQuantizedConv           0.986384           34.6924    int8      
/model.2/cv2/act/Mul                                BPU  id(0)     HzLut                      0.982361           42.8399    int8      
/model.3/conv/Conv                                  BPU  id(0)     HzSQuantizedConv           0.987570           11.3628    int8      
/model.3/act/Mul                                    BPU  id(0)     HzLut                      0.992035           8.02566    int8      
/model.4/cv1/conv/Conv                              BPU  id(0)     HzSQuantizedConv           0.986502           5.97602    int8      
/model.4/cv1/act/Mul                                BPU  id(0)     HzLut                      0.988705           6.16003    int8      
/model.4/Split                                      BPU  id(0)     Split                      0.987318           3.84233    int8      
/model.4/m.0/cv1/conv/Conv                          BPU  id(0)     HzSQuantizedConv           0.991171           3.84233    int8      
/model.4/m.0/cv1/act/Mul                            BPU  id(0)     HzLut                      0.992874           3.75658    int8      
/model.4/m.0/cv2/conv/Conv                          BPU  id(0)     HzSQuantizedConv           0.990146           3.59272    int8      
/model.4/m.0/cv2/act/Mul                            BPU  id(0)     HzLut                      0.991048           6.0697     int8      
/model.4/m.0/Add                                    BPU  id(0)     HzSElementwiseAdd          0.993869           3.84233    int8      
/model.4/Split_output_0_calibrated_Requantize       BPU  id(0)     HzRequantize               --                 --         int8      
/model.4/Split_output_1_calibrated_Requantize       BPU  id(0)     HzRequantize               --                 --         int8      
/model.4/Concat                                     BPU  id(0)     Concat                     0.992462           3.84233    int8      
/model.4/cv2/conv/Conv                              BPU  id(0)     HzSQuantizedConv           0.977627           5.84837    int8      
/model.4/cv2/act/Mul                                BPU  id(0)     HzLut                      0.966566           5.285      int8      
/model.5/conv/Conv                                  BPU  id(0)     HzSQuantizedConv           0.983363           2.80314    int8      
/model.5/act/Mul                                    BPU  id(0)     HzLut                      0.986903           5.26058    int8      
/model.6/cv1/conv/Conv                              BPU  id(0)     HzSQuantizedConv           0.965669           4.12865    int8      
/model.6/cv1/act/Mul                                BPU  id(0)     HzLut                      0.965217           5.81133    int8      
/model.6/Split                                      BPU  id(0)     Split                      0.956085           3.70905    int8      
/model.6/m.0/cv1/conv/Conv                          BPU  id(0)     HzSQuantizedConv           0.979331           3.70905    int8      
/model.6/m.0/cv2/conv/Conv                          BPU  id(0)     HzSQuantizedConv           0.958494           3.70905    int8      
/model.6/m.0/cv1/act/Mul                            BPU  id(0)     HzLut                      0.981890           2.85697    int8      
/model.6/m.0/cv2/act/Mul                            BPU  id(0)     HzLut                      0.955246           9.15907    int8      
/model.6/m.0/m/m.0/cv1/conv/Conv                    BPU  id(0)     HzSQuantizedConv           0.976006           2.25202    int8      
/model.6/m.0/m/m.0/cv1/act/Mul                      BPU  id(0)     HzLut                      0.964667           5.34329    int8      
/model.6/m.0/m/m.0/cv2/conv/Conv                    BPU  id(0)     HzSQuantizedConv           0.976574           2.69569    int8      
/model.6/m.0/m/m.0/cv2/act/Mul                      BPU  id(0)     HzLut                      0.979242           3.30823    int8      
/model.6/m.0/m/m.0/Add                              BPU  id(0)     HzSElementwiseAdd          0.984495           2.25202    int8      
/model.6/m.0/m/m.1/cv1/conv/Conv                    BPU  id(0)     HzSQuantizedConv           0.982179           3.46597    int8      
/model.6/m.0/m/m.1/cv1/act/Mul                      BPU  id(0)     HzLut                      0.975347           4.34563    int8      
/model.6/m.0/m/m.1/cv2/conv/Conv                    BPU  id(0)     HzSQuantizedConv           0.988352           2.99381    int8      
/model.6/m.0/m/m.1/cv2/act/Mul                      BPU  id(0)     HzLut                      0.990266           5.93548    int8      
/model.6/m.0/m/m.1/Add                              BPU  id(0)     HzSElementwiseAdd          0.992224           3.46597    int8      
/model.6/m.0/Concat                                 BPU  id(0)     Concat                     0.988938           7.01479    int8      
/model.6/m.0/cv3/conv/Conv                          BPU  id(0)     HzSQuantizedConv           0.973695           7.01479    int8      
/model.6/m.0/cv3/act/Mul                            BPU  id(0)     HzLut                      0.971947           7.1746     int8      
/model.6/Split_output_0_calibrated_Requantize       BPU  id(0)     HzRequantize               --                 --         int8      
/model.6/Split_output_1_calibrated_Requantize       BPU  id(0)     HzRequantize               --                 --         int8      
/model.6/Concat                                     BPU  id(0)     Concat                     0.967469           3.70905    int8      
/model.6/cv2/conv/Conv                              BPU  id(0)     HzSQuantizedConv           0.969959           4.16062    int8      
/model.6/cv2/act/Mul                                BPU  id(0)     HzLut                      0.959340           4.88968    int8      
/model.7/conv/Conv                                  BPU  id(0)     HzSQuantizedConv           0.957919           2.82507    int8      
/model.7/act/Mul                                    BPU  id(0)     HzLut                      0.952700           6.80349    int8      
/model.8/cv1/conv/Conv                              BPU  id(0)     HzSQuantizedConv           0.940611           3.83789    int8      
/model.8/cv1/act/Mul                                BPU  id(0)     HzLut                      0.925319           7.11303    int8      
/model.8/Split                                      BPU  id(0)     Split                      0.930325           4.95504    int8      
/model.8/m.0/cv1/conv/Conv                          BPU  id(0)     HzSQuantizedConv           0.960385           4.95504    int8      
/model.8/m.0/cv2/conv/Conv                          BPU  id(0)     HzSQuantizedConv           0.913674           4.95504    int8      
/model.8/m.0/cv1/act/Mul                            BPU  id(0)     HzLut                      0.952659           5.03103    int8      
/model.8/m.0/cv2/act/Mul                            BPU  id(0)     HzLut                      0.912352           9.13057    int8      
/model.8/m.0/m/m.0/cv1/conv/Conv                    BPU  id(0)     HzSQuantizedConv           0.957526           2.04927    int8      
/model.8/m.0/m/m.0/cv1/act/Mul                      BPU  id(0)     HzLut                      0.946501           5.49295    int8      
/model.8/m.0/m/m.0/cv2/conv/Conv                    BPU  id(0)     HzSQuantizedConv           0.966301           3.56367    int8      
/model.8/m.0/m/m.0/cv2/act/Mul                      BPU  id(0)     HzLut                      0.961327           7.63598    int8      
/model.8/m.0/m/m.0/Add                              BPU  id(0)     HzSElementwiseAdd          0.962684           2.04927    int8      
/model.8/m.0/m/m.1/cv1/conv/Conv                    BPU  id(0)     HzSQuantizedConv           0.957281           5.08985    int8      
/model.8/m.0/m/m.1/cv1/act/Mul                      BPU  id(0)     HzLut                      0.958437           6.36546    int8      
/model.8/m.0/m/m.1/cv2/conv/Conv                    BPU  id(0)     HzSQuantizedConv           0.959945           4.74152    int8      
/model.8/m.0/m/m.1/cv2/act/Mul                      BPU  id(0)     HzLut                      0.960627           15.4104    int8      
/model.8/m.0/m/m.1/Add                              BPU  id(0)     HzSElementwiseAdd          0.969886           5.08985    int8      
/model.8/m.0/Concat                                 BPU  id(0)     Concat                     0.963206           11.7198    int8      
/model.8/m.0/cv3/conv/Conv                          BPU  id(0)     HzSQuantizedConv           0.947358           11.7198    int8      
/model.8/m.0/cv3/act/Mul                            BPU  id(0)     HzLut                      0.939252           7.90709    int8      
/model.8/Split_output_0_calibrated_Requantize       BPU  id(0)     HzRequantize               --                 --         int8      
/model.8/Split_output_1_calibrated_Requantize       BPU  id(0)     HzRequantize               --                 --         int8      
/model.8/Concat                                     BPU  id(0)     Concat                     0.931873           4.95504    int8      
/model.8/cv2/conv/Conv                              BPU  id(0)     HzSQuantizedConv           0.954973           5.24073    int8      
/model.8/cv2/act/Mul                                BPU  id(0)     HzLut                      0.949352           7.94439    int8      
/model.9/cv1/conv/Conv                              BPU  id(0)     HzSQuantizedConv           0.919595           5.88878    int8      
/model.9/m/MaxPool                                  BPU  id(0)     HzQuantizedMaxPool         0.975137           12.2629    int8      
/model.9/m_1/MaxPool                                BPU  id(0)     HzQuantizedMaxPool         0.984824           12.2629    int8      
/model.9/m_2/MaxPool                                BPU  id(0)     HzQuantizedMaxPool         0.987853           12.2629    int8      
/model.9/Concat                                     BPU  id(0)     Concat                     0.977351           12.2629    int8      
/model.9/cv2/conv/Conv                              BPU  id(0)     HzSQuantizedConv           0.957602           12.2629    int8      
/model.9/cv2/act/Mul                                BPU  id(0)     HzLut                      0.921119           12.8324    int8      
/model.9/Add                                        BPU  id(0)     HzSElementwiseAdd          0.943476           8.70204    int8      
/model.10/cv1/conv/Conv                             BPU  id(0)     HzSQuantizedConv           0.925221           9.3442     int8      
/model.10/cv1/act/Mul                               BPU  id(0)     HzLut                      0.948974           20.5783    int8      
/model.10/Split                                     BPU  id(0)     Split                      0.947308           20.5783    int8      
/model.10/m/m.0/attn/qkv/conv/Conv                  BPU  id(0)     HzSQuantizedConv           0.909879           20.5783    int8      
/model.10/m/m.0/attn/Reshape                        BPU  id(0)     Reshape                    0.909879           9.48591    int8      
/model.10/m/m.0/attn/Split                          BPU  id(0)     Split                      0.943776           9.48591    int8      
/model.10/m/m.0/attn/Transpose                      BPU  id(0)     Transpose                  0.943777           9.48591    int8      
/model.10/m/m.0/attn/Reshape_2                      BPU  id(0)     Reshape                    0.895889           9.48591    int8      
/model.10/m/m.0/attn/MatMul                         BPU  id(0)     HzSQuantizedMatmul         0.920568           9.48591    int8      
/model.10/m/m.0/attn/Mul                            BPU  id(0)     HzSQuantizedConv           0.920568           67.5368    int8      
...0/attn/Softmax_reducemax_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzQuantizedReduceMax       0.994295           11.9389    int8      
...0/m/m.0/attn/Softmax_sub_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzSElementwiseSub          0.939837           11.9389    int8      
...0/m/m.0/attn/Softmax_exp_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzLut                      0.905658           5.64418    int8      
...0/attn/Softmax_reducesum_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzSQuantizedReduceSum      0.973537           1.0        int8      
.../attn/Softmax_reciprocal_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzLut                      0.919691           113.568    int8      
...0/m/m.0/attn/Softmax_mul_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzSElementwiseMul          0.870533           1.0        int8      
/model.10/m/m.0/attn/Transpose_1                    BPU  id(0)     Transpose                  0.870532           0.300665   int8      
/model.10/m/m.0/attn/MatMul_1                       BPU  id(0)     HzSQuantizedMatmul         0.947740           9.48591    int8      
/model.10/m/m.0/attn/Reshape_1                      BPU  id(0)     Reshape                    0.947740           6.3334     int8      
/model.10/m/m.0/attn/pe/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.906148           9.48591    int8      
/model.10/m/m.0/attn/proj/conv/Conv                 BPU  id(0)     HzSQuantizedConv           0.864568           5.34381    int8      
/model.10/m/m.0/ffn/ffn.0/conv/Conv                 BPU  id(0)     HzSQuantizedConv           0.971892           5.73318    int8      
/model.10/m/m.0/ffn/ffn.0/act/Mul                   BPU  id(0)     HzLut                      0.951269           5.64755    int8      
/model.10/m/m.0/ffn/ffn.1/conv/Conv                 BPU  id(0)     HzSQuantizedConv           0.781569           1.97536    int8      
/model.10/Concat                                    BPU  id(0)     Concat                     0.928312           20.5783    int8      
/model.10/cv2/conv/Conv                             BPU  id(0)     HzSQuantizedConv           0.945669           20.5783    int8      
/model.10/cv2/act/Mul                               BPU  id(0)     HzLut                      0.919415           8.27366    int8      
/model.11/Resize                                    BPU  id(0)     HzQuantizedResizeUpsample  0.919400           4.21709    int8      
/model.11/Resize_output_0_calibrated_Requantize     BPU  id(0)     HzRequantize               --                 --         int8      
...el.6/cv2/act/Mul_output_0_calibrated_Requantize  BPU  id(0)     HzRequantize               --                 --         int8      
/model.12/Concat                                    BPU  id(0)     Concat                     0.931564           4.21709    int8      
/model.13/cv1/conv/Conv                             BPU  id(0)     HzSQuantizedConv           0.947009           4.21529    int8      
/model.13/cv1/act/Mul                               BPU  id(0)     HzLut                      0.955289           7.09418    int8      
/model.13/Split                                     BPU  id(0)     Split                      0.951366           3.99445    int8      
/model.13/m.0/cv1/conv/Conv                         BPU  id(0)     HzSQuantizedConv           0.957274           3.99445    int8      
/model.13/m.0/cv2/conv/Conv                         BPU  id(0)     HzSQuantizedConv           0.919471           3.99445    int8      
/model.13/m.0/cv1/act/Mul                           BPU  id(0)     HzLut                      0.967630           3.25442    int8      
/model.13/m.0/cv2/act/Mul                           BPU  id(0)     HzLut                      0.896833           8.98231    int8      
/model.13/m.0/m/m.0/cv1/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.969446           1.51745    int8      
/model.13/m.0/m/m.0/cv1/act/Mul                     BPU  id(0)     HzLut                      0.952117           4.53908    int8      
/model.13/m.0/m/m.0/cv2/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.976392           2.41541    int8      
/model.13/m.0/m/m.0/cv2/act/Mul                     BPU  id(0)     HzLut                      0.982481           4.86362    int8      
/model.13/m.0/m/m.0/Add                             BPU  id(0)     HzSElementwiseAdd          0.980348           1.51745    int8      
/model.13/m.0/m/m.1/cv1/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.970932           3.36905    int8      
/model.13/m.0/m/m.1/cv1/act/Mul                     BPU  id(0)     HzLut                      0.965228           3.84044    int8      
/model.13/m.0/m/m.1/cv2/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.990089           2.06876    int8      
/model.13/m.0/m/m.1/cv2/act/Mul                     BPU  id(0)     HzLut                      0.989747           6.52589    int8      
/model.13/m.0/m/m.1/Add                             BPU  id(0)     HzSElementwiseAdd          0.991309           3.36905    int8      
/model.13/m.0/Concat                                BPU  id(0)     Concat                     0.987641           8.23091    int8      
/model.13/m.0/cv3/conv/Conv                         BPU  id(0)     HzSQuantizedConv           0.957342           8.23091    int8      
/model.13/m.0/cv3/act/Mul                           BPU  id(0)     HzLut                      0.949348           4.23047    int8      
/model.13/Split_output_0_calibrated_Requantize      BPU  id(0)     HzRequantize               --                 --         int8      
/model.13/Split_output_1_calibrated_Requantize      BPU  id(0)     HzRequantize               --                 --         int8      
/model.13/Concat                                    BPU  id(0)     Concat                     0.954095           3.99445    int8      
/model.13/cv2/conv/Conv                             BPU  id(0)     HzSQuantizedConv           0.955105           3.80039    int8      
/model.13/cv2/act/Mul                               BPU  id(0)     HzLut                      0.958404           5.82707    int8      
/model.14/Resize                                    BPU  id(0)     HzQuantizedResizeUpsample  0.958458           3.00218    int8      
/model.14/Resize_output_0_calibrated_Requantize     BPU  id(0)     HzRequantize               --                 --         int8      
...el.4/cv2/act/Mul_output_0_calibrated_Requantize  BPU  id(0)     HzRequantize               --                 --         int8      
/model.15/Concat                                    BPU  id(0)     Concat                     0.961011           3.00218    int8      
/model.16/cv1/conv/Conv                             BPU  id(0)     HzSQuantizedConv           0.984929           2.64325    int8      
/model.16/cv1/act/Mul                               BPU  id(0)     HzLut                      0.987454           3.60601    int8      
/model.16/Split                                     BPU  id(0)     Split                      0.985717           3.34793    int8      
/model.16/m.0/cv1/conv/Conv                         BPU  id(0)     HzSQuantizedConv           0.991058           3.34793    int8      
/model.16/m.0/cv2/conv/Conv                         BPU  id(0)     HzSQuantizedConv           0.988863           3.34793    int8      
/model.16/m.0/cv1/act/Mul                           BPU  id(0)     HzLut                      0.993833           1.29608    int8      
/model.16/m.0/cv2/act/Mul                           BPU  id(0)     HzLut                      0.992083           5.38116    int8      
/model.16/m.0/m/m.0/cv1/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.972915           1.00723    int8      
/model.16/m.0/m/m.0/cv1/act/Mul                     BPU  id(0)     HzLut                      0.967981           3.39708    int8      
/model.16/m.0/m/m.0/cv2/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.978850           2.99559    int8      
/model.16/m.0/m/m.0/cv2/act/Mul                     BPU  id(0)     HzLut                      0.987576           3.59757    int8      
/model.16/m.0/m/m.0/Add                             BPU  id(0)     HzSElementwiseAdd          0.987586           1.00723    int8      
/model.16/m.0/m/m.1/cv1/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.989137           2.9602     int8      
/model.16/m.0/m/m.1/cv1/act/Mul                     BPU  id(0)     HzLut                      0.995549           5.01332    int8      
/model.16/m.0/m/m.1/cv2/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.995177           2.99879    int8      
/model.16/m.0/m/m.1/cv2/act/Mul                     BPU  id(0)     HzLut                      0.995613           10.4563    int8      
/model.16/m.0/m/m.1/Add                             BPU  id(0)     HzSElementwiseAdd          0.995611           2.9602     int8      
/model.16/m.0/Concat                                BPU  id(0)     Concat                     0.995279           11.2084    int8      
/model.16/m.0/cv3/conv/Conv                         BPU  id(0)     HzSQuantizedConv           0.985486           11.2084    int8      
/model.16/m.0/cv3/act/Mul                           BPU  id(0)     HzLut                      0.986524           7.34149    int8      
/model.16/Split_output_0_calibrated_Requantize      BPU  id(0)     HzRequantize               --                 --         int8      
/model.16/Split_output_1_calibrated_Requantize      BPU  id(0)     HzRequantize               --                 --         int8      
/model.16/Concat                                    BPU  id(0)     Concat                     0.986972           3.34793    int8      
/model.16/cv2/conv/Conv                             BPU  id(0)     HzSQuantizedConv           0.981286           5.32721    int8      
/model.16/cv2/act/Mul                               BPU  id(0)     HzLut                      0.991466           9.86162    int8      
/model.17/conv/Conv                                 BPU  id(0)     HzSQuantizedConv           0.977645           4.44043    int8      
...3.0/one2one_cv3.0.0/one2one_cv3.0.0.0/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.997363           4.44043    int8      
/model.23/one2one_cv2.0/one2one_cv2.0.0/conv/Conv   BPU  id(0)     HzSQuantizedConv           0.993177           4.44043    int8      
/model.17/act/Mul                                   BPU  id(0)     HzLut                      0.974019           5.94218    int8      
...cv3.0/one2one_cv3.0.0/one2one_cv3.0.0.0/act/Mul  BPU  id(0)     HzLut                      0.996707           6.98937    int8      
/model.23/one2one_cv2.0/one2one_cv2.0.0/act/Mul     BPU  id(0)     HzLut                      0.993933           12.1241    int8      
/model.18/Concat                                    BPU  id(0)     Concat                     0.962423           3.00218    int8      
...3.0/one2one_cv3.0.0/one2one_cv3.0.0.1/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.940910           6.98293    int8      
/model.23/one2one_cv2.0/one2one_cv2.0.1/conv/Conv   BPU  id(0)     HzSQuantizedConv           0.982400           11.5376    int8      
/model.19/cv1/conv/Conv                             BPU  id(0)     HzSQuantizedConv           0.952182           3.00218    int8      
...cv3.0/one2one_cv3.0.0/one2one_cv3.0.0.1/act/Mul  BPU  id(0)     HzLut                      0.940317           3.70321    int8      
/model.23/one2one_cv2.0/one2one_cv2.0.1/act/Mul     BPU  id(0)     HzLut                      0.982003           10.2031    int8      
/model.19/cv1/act/Mul                               BPU  id(0)     HzLut                      0.954909           4.70628    int8      
...3.0/one2one_cv3.0.1/one2one_cv3.0.1.0/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.918869           3.56602    int8      
/model.23/one2one_cv2.0/one2one_cv2.0.2/Conv        BPU  id(0)     HzSQuantizedConv           0.991031           3.64807    int8      
/model.19/Split                                     BPU  id(0)     Split                      0.939838           4.46392    int8      
/model.19/m.0/cv1/conv/Conv                         BPU  id(0)     HzSQuantizedConv           0.989980           4.46392    int8      
/model.19/m.0/cv2/conv/Conv                         BPU  id(0)     HzSQuantizedConv           0.976776           4.46392    int8      
...cv3.0/one2one_cv3.0.1/one2one_cv3.0.1.0/act/Mul  BPU  id(0)     HzLut                      0.911491           5.67313    int8      
...3.0/one2one_cv3.0.1/one2one_cv3.0.1.1/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.905940           4.64853    int8      
/model.19/m.0/cv1/act/Mul                           BPU  id(0)     HzLut                      0.992725           2.64108    int8      
/model.19/m.0/cv2/act/Mul                           BPU  id(0)     HzLut                      0.968877           6.29187    int8      
/model.19/m.0/m/m.0/cv1/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.976733           0.835912   int8      
...cv3.0/one2one_cv3.0.1/one2one_cv3.0.1.1/act/Mul  BPU  id(0)     HzLut                      0.903295           4.24024    int8      
/model.23/one2one_cv3.0/one2one_cv3.0.2/Conv        BPU  id(0)     HzSQuantizedConv           0.999801           4.15891    int8      
/model.19/m.0/m/m.0/cv1/act/Mul                     BPU  id(0)     HzLut                      0.965798           5.90733    int8      
/model.19/m.0/m/m.0/cv2/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.981310           2.33814    int8      
/model.19/m.0/m/m.0/cv2/act/Mul                     BPU  id(0)     HzLut                      0.984562           4.96914    int8      
/model.19/m.0/m/m.0/Add                             BPU  id(0)     HzSElementwiseAdd          0.988171           0.835912   int8      
/model.19/m.0/m/m.1/cv1/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.977087           2.59399    int8      
/model.19/m.0/m/m.1/cv1/act/Mul                     BPU  id(0)     HzLut                      0.972357           6.62685    int8      
/model.19/m.0/m/m.1/cv2/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.986858           3.03324    int8      
/model.19/m.0/m/m.1/cv2/act/Mul                     BPU  id(0)     HzLut                      0.990025           11.7915    int8      
/model.19/m.0/m/m.1/Add                             BPU  id(0)     HzSElementwiseAdd          0.990814           2.59399    int8      
/model.19/m.0/Concat                                BPU  id(0)     Concat                     0.989853           11.2476    int8      
/model.19/m.0/cv3/conv/Conv                         BPU  id(0)     HzSQuantizedConv           0.979148           11.2476    int8      
/model.19/m.0/cv3/act/Mul                           BPU  id(0)     HzLut                      0.979345           12.7928    int8      
/model.19/Split_output_0_calibrated_Requantize      BPU  id(0)     HzRequantize               --                 --         int8      
/model.19/Split_output_1_calibrated_Requantize      BPU  id(0)     HzRequantize               --                 --         int8      
/model.19/Concat                                    BPU  id(0)     Concat                     0.971184           4.46392    int8      
/model.19/cv2/conv/Conv                             BPU  id(0)     HzSQuantizedConv           0.977676           5.4348     int8      
/model.19/cv2/act/Mul                               BPU  id(0)     HzLut                      0.984003           11.9404    int8      
/model.20/conv/Conv                                 BPU  id(0)     HzSQuantizedConv           0.969865           4.35456    int8      
...3.1/one2one_cv3.1.0/one2one_cv3.1.0.0/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.993468           4.35456    int8      
/model.23/one2one_cv2.1/one2one_cv2.1.0/conv/Conv   BPU  id(0)     HzSQuantizedConv           0.991981           4.35456    int8      
/model.20/act/Mul                                   BPU  id(0)     HzLut                      0.953748           7.5289     int8      
...cv3.1/one2one_cv3.1.0/one2one_cv3.1.0.0/act/Mul  BPU  id(0)     HzLut                      0.993290           11.9308    int8      
/model.23/one2one_cv2.1/one2one_cv2.1.0/act/Mul     BPU  id(0)     HzLut                      0.993349           9.85252    int8      
/model.21/Concat                                    BPU  id(0)     Concat                     0.931940           4.21709    int8      
...3.1/one2one_cv3.1.0/one2one_cv3.1.0.1/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.938509           11.9307    int8      
/model.23/one2one_cv2.1/one2one_cv2.1.1/conv/Conv   BPU  id(0)     HzSQuantizedConv           0.986387           9.85201    int8      
/model.22/cv1/conv/Conv                             BPU  id(0)     HzSQuantizedConv           0.914396           4.21709    int8      
...cv3.1/one2one_cv3.1.0/one2one_cv3.1.0.1/act/Mul  BPU  id(0)     HzLut                      0.934071           5.21448    int8      
/model.23/one2one_cv2.1/one2one_cv2.1.1/act/Mul     BPU  id(0)     HzLut                      0.985686           5.85388    int8      
/model.22/cv1/act/Mul                               BPU  id(0)     HzLut                      0.884339           8.98337    int8      
...3.1/one2one_cv3.1.1/one2one_cv3.1.1.0/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.926885           4.59897    int8      
/model.23/one2one_cv2.1/one2one_cv2.1.2/Conv        BPU  id(0)     HzSQuantizedConv           0.993526           4.3885     int8      
/model.22/Split                                     BPU  id(0)     Split                      0.860255           6.05575    int8      
/model.22/m.0/m.0.0/cv1/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.977903           6.05575    int8      
...cv3.1/one2one_cv3.1.1/one2one_cv3.1.1.0/act/Mul  BPU  id(0)     HzLut                      0.919584           7.71591    int8      
...3.1/one2one_cv3.1.1/one2one_cv3.1.1.1/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.898816           7.43917    int8      
/model.22/m.0/m.0.0/cv1/act/Mul                     BPU  id(0)     HzLut                      0.984141           5.91858    int8      
/model.22/m.0/m.0.0/cv2/conv/Conv                   BPU  id(0)     HzSQuantizedConv           0.983478           4.17787    int8      
...cv3.1/one2one_cv3.1.1/one2one_cv3.1.1.1/act/Mul  BPU  id(0)     HzLut                      0.893130           3.98663    int8      
/model.23/one2one_cv3.1/one2one_cv3.1.2/Conv        BPU  id(0)     HzSQuantizedConv           0.999497           3.65477    int8      
/model.22/m.0/m.0.0/cv2/act/Mul                     BPU  id(0)     HzLut                      0.983533           8.44104    int8      
/model.22/m.0/m.0.0/Add                             BPU  id(0)     HzSElementwiseAdd          0.982657           6.05575    int8      
/model.22/m.0/m.0.1/attn/qkv/conv/Conv              BPU  id(0)     HzSQuantizedConv           0.979321           8.02413    int8      
/model.22/m.0/m.0.1/attn/Reshape                    BPU  id(0)     Reshape                    0.979321           10.9123    int8      
/model.22/m.0/m.0.1/attn/Split                      BPU  id(0)     Split                      0.974744           10.9123    int8      
/model.22/m.0/m.0.1/attn/Transpose                  BPU  id(0)     Transpose                  0.974744           10.9123    int8      
/model.22/m.0/m.0.1/attn/Reshape_2                  BPU  id(0)     Reshape                    0.979533           10.9123    int8      
/model.22/m.0/m.0.1/attn/MatMul                     BPU  id(0)     HzSQuantizedMatmul         0.978635           10.9123    int8      
/model.22/m.0/m.0.1/attn/Mul                        BPU  id(0)     HzSQuantizedConv           0.978635           110.91     int8      
...1/attn/Softmax_reducemax_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzQuantizedReduceMax       0.995463           19.6062    int8      
...0/m.0.1/attn/Softmax_sub_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzSElementwiseSub          0.932269           19.6062    int8      
...0/m.0.1/attn/Softmax_exp_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzLut                      0.878896           5.64418    int8      
...1/attn/Softmax_reducesum_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzSQuantizedReduceSum      0.959095           1.0        int8      
.../attn/Softmax_reciprocal_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzLut                      0.963902           84.2815    int8      
...0/m.0.1/attn/Softmax_mul_FROM_QUANTIZED_SOFTMAX  BPU  id(0)     HzSElementwiseMul          0.917670           1.0        int8      
/model.22/m.0/m.0.1/attn/Transpose_1                BPU  id(0)     Transpose                  0.917671           0.336908   int8      
/model.22/m.0/m.0.1/attn/MatMul_1                   BPU  id(0)     HzSQuantizedMatmul         0.977507           10.9123    int8      
/model.22/m.0/m.0.1/attn/Reshape_1                  BPU  id(0)     Reshape                    0.977507           9.33229    int8      
/model.22/m.0/m.0.1/attn/pe/conv/Conv               BPU  id(0)     HzSQuantizedConv           0.968197           10.9123    int8      
/model.22/m.0/m.0.1/attn/proj/conv/Conv             BPU  id(0)     HzSQuantizedConv           0.942881           6.099      int8      
/model.22/m.0/m.0.1/ffn/ffn.0/conv/Conv             BPU  id(0)     HzSQuantizedConv           0.971395           8.02227    int8      
/model.22/m.0/m.0.1/ffn/ffn.0/act/Mul               BPU  id(0)     HzLut                      0.926983           11.2912    int8      
/model.22/m.0/m.0.1/ffn/ffn.1/conv/Conv             BPU  id(0)     HzSQuantizedConv           0.925082           6.26859    int8      
/model.22/Split_output_0_calibrated_Requantize      BPU  id(0)     HzRequantize               --                 --         int8      
/model.22/Split_output_1_calibrated_Requantize      BPU  id(0)     HzRequantize               --                 --         int8      
/model.22/Concat                                    BPU  id(0)     Concat                     0.940496           6.05575    int8      
/model.22/cv2/conv/Conv                             BPU  id(0)     HzSQuantizedConv           0.963748           6.66716    int8      
/model.22/cv2/act/Mul                               BPU  id(0)     HzLut                      0.966216           12.6563    int8      
...3.2/one2one_cv3.2.0/one2one_cv3.2.0.0/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.970342           11.4354    int8      
/model.23/one2one_cv2.2/one2one_cv2.2.0/conv/Conv   BPU  id(0)     HzSQuantizedConv           0.982549           11.4354    int8      
...cv3.2/one2one_cv3.2.0/one2one_cv3.2.0.0/act/Mul  BPU  id(0)     HzLut                      0.968683           10.9199    int8      
/model.23/one2one_cv2.2/one2one_cv2.2.0/act/Mul     BPU  id(0)     HzLut                      0.987581           10.9047    int8      
...3.2/one2one_cv3.2.0/one2one_cv3.2.0.1/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.940824           6.71417    int8      
/model.23/one2one_cv2.2/one2one_cv2.2.1/conv/Conv   BPU  id(0)     HzSQuantizedConv           0.981390           10.9045    int8      
...cv3.2/one2one_cv3.2.0/one2one_cv3.2.0.1/act/Mul  BPU  id(0)     HzLut                      0.942234           3.8541     int8      
/model.23/one2one_cv2.2/one2one_cv2.2.1/act/Mul     BPU  id(0)     HzLut                      0.983515           21.3939    int8      
...3.2/one2one_cv3.2.1/one2one_cv3.2.1.0/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.933661           3.3179     int8      
/model.23/one2one_cv2.2/one2one_cv2.2.2/Conv        BPU  id(0)     HzSQuantizedConv           0.992204           21.3939    int8      
...cv3.2/one2one_cv3.2.1/one2one_cv3.2.1.0/act/Mul  BPU  id(0)     HzLut                      0.926668           6.20664    int8      
...3.2/one2one_cv3.2.1/one2one_cv3.2.1.1/conv/Conv  BPU  id(0)     HzSQuantizedConv           0.925398           4.90883    int8      
...cv3.2/one2one_cv3.2.1/one2one_cv3.2.1.1/act/Mul  BPU  id(0)     HzLut                      0.919776           3.81159    int8      
/model.23/one2one_cv3.2/one2one_cv3.2.2/Conv        BPU  id(0)     HzSQuantizedConv           0.999860           3.72568    int8
2026-04-30 13:13:44,165 file: print_info_dict.py func: print_info_dict line No: 72 The quantized model output:
=============================================================================
Output      Cosine Similarity  L1 Distance  L2 Distance  Chebyshev Distance  
-----------------------------------------------------------------------------
output0     0.999802           0.139915     0.001662     1.488247            
580         0.991031           0.313810     0.002931     3.019216            
594         0.999497           0.185436     0.004304     1.495047            
602         0.993526           0.369644     0.007280     3.608502            
616         0.999860           0.083254     0.003808     0.388244            
624         0.992204           0.634017     0.026592     7.302471
2026-04-30 13:13:44,170 file: model_builder.py func: model_builder line No: 38 End to Horizon NN Model Convert.
2026-04-30 13:13:44,177 file: hb_mapper_makertbin.py func: hb_mapper_makertbin line No: 601 start convert to *.bin file....
2026-04-30 13:13:44,193 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4326 ONNX model output num : 6
2026-04-30 13:13:44,195 file: layout_util.py func: layout_util line No: 15 set_featuremap_layout start
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4060 model_deps_info: {'hb_mapper_version': '1.24.3', 'hbdk_version': '3.49.15', 'hbdk_runtime_version': ' 3.15.55.0', 'horizon_nn_version': '1.1.0', 'onnx_model': '/home/ddd/rdk_model_zoo-main/best26n.onnx', 'march': 'bayes-e', 'layer_out_dump': False, 'log_level': 'DEBUG', 'working_dir': '/home/ddd/rdk_model_zoo-main/.temporary_workspace/bpu_model_output', 'model_prefix': 'best26n_bayese_640x640_nv12', 'input_names': ['images'], 'input_type_rt': ['nv12'], 'input_space_and_range': ['regular'], 'input_type_train': ['rgb'], 'input_layout_rt': [''], 'input_layout_train': ['NCHW'], 'norm_type': ['data_scale'], 'scale_value': ['0.003921568627451,'], 'mean_value': [''], 'input_shape': ['1x3x640x640'], 'input_batch': [], 'cal_dir': ['/home/ddd/rdk_model_zoo-main/.temporary_workspace/calibration_data'], 'cal_data_type': ['float32'], 'preprocess_on': False, 'calibration_type': 'default', 'per_channel': 'False', 'optimization': ['set_Softmax_input_int8,set_Softmax_output_int8'], 'hbdk_params': {'hbdk_pass_through_params': '--O3 --debug --core-num 1 --fast --jobs 16 ', 'input-source': {'images': 'pyramid', '_default_value': 'ddr'}}, 'debug': True, 'compile_mode': 'latency'}
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4183 ############# model deps info #############
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4184 hb_mapper version   : 1.24.3
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4187 hbdk version        : 3.49.15
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4189 hbdk runtime version: 3.15.55.0
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4192 horizon_nn version  : 1.1.0
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4196 ############# model_parameters info #############
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4202 onnx_model          : /home/ddd/rdk_model_zoo-main/best26n.onnx
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4203 BPU march           : bayes-e
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4204 layer_out_dump      : False
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4205 log_level           : DEBUG
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4206 working dir         : /home/ddd/rdk_model_zoo-main/.temporary_workspace/bpu_model_output
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4207 output_model_file_prefix: best26n_bayese_640x640_nv12
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4228 ############# input_parameters info #############
2026-04-30 13:13:44,195 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4246 ------------------------------------------
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4248 ---------input info : images ---------
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4249 input_name          : images
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4250 input_type_rt       : nv12
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4252 input_space&range   : regular
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4254 input_layout_rt     : None
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4255 input_type_train    : rgb
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4256 input_layout_train  : NCHW
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4257 norm_type           : data_scale
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4258 input_shape         : 1x3x640x640
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4266 scale_value         : 0.003921568627451,
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4268 cal_data_dir        : /home/ddd/rdk_model_zoo-main/.temporary_workspace/calibration_data
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4270 cal_data_type       : float32
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4271 ---------input info : images end -------
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4272 ------------------------------------------
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4274 ############# calibration_parameters info #############
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4275 preprocess_on       : False
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4276 calibration_type:   : default
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4278 optimization        : set_Softmax_input_int8,set_Softmax_output_int8;
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4284 per_channel         : False
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4303 ############# compiler_parameters info #############
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4305 debug               : True
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4307 compile_mode        : latency
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4310 hbdk_pass_through_params: --O3 --debug --core-num 1 --fast --jobs 16
2026-04-30 13:13:44,196 file: onnx2horizonrt.py func: onnx2horizonrt line No: 4310 input-source        : {'images': 'pyramid', '_default_value': 'ddr'}
2026-04-30 13:13:44,199 file: hb_mapper_makertbin.py func: hb_mapper_makertbin line No: 783 Convert to runtime bin file successfully!
2026-04-30 13:13:44,199 file: hb_mapper_makertbin.py func: hb_mapper_makertbin line No: 784 End Model Convert

完整log:
hb_mapper_makertbin.txt (175.0 KB)

测试图片
rdk_model_zoo\samples\vision\ultralytics_yolo26\runtime\python\yolo26_det.py
使用数据集测试图片，最后结果几乎不可检出，仅有少数检出

地瓜小助手_bot · 2026 年4 月 30 日 05:52

我来帮你分析这个问题。从日志看，量化过程中有几个节点的 Cosine Similarity 较低，这可能是导致检测性能下降的关键原因。

先搜索一下社区是否有类似的 YOLO 量化问题：

问题分析

根据你的日志，核心问题是量化后 cls 输出分布异常：

阶段	cls 输出范围	问题
ONNX	`[-229, 82]`	正常
量化后	`[-14, -0.1]`	全为负值且范围极窄

量化后 cls 输出经过 sigmoid 后置信度约为 0.0001-0.47，大量目标被阈值过滤。

排查步骤

步骤 1：检查输入格式一致性

最关键：确认训练、导出、量化三阶段的输入格式完全一致。

# 检查你的训练预处理代码
# 关键问题：训练时用 RGB 还是 BGR？是否除以 255？

# 查看 mapper.py 默认配置
cd rdk_model_zoo/samples/vision/ultralytics_yolo26/conversion
grep -A 10 "input_type" mapper.py

常见错误：

训练用 RGB，量化配置用 BGR
训练时已除以 255，量化时又重复归一化

步骤 2：使用官方校准数据集测试

先用仓库自带的 cal/ 目录测试，排除校准数据问题：

cd rdk_model_zoo/samples/vision/ultralytics_yolo26/conversion

# 使用官方示例校准数据重新量化
python mapper.py \
  --onnx ./best26n.onnx \
  --cal-images ./cal/ \
  --cal-sample-num 100 \
  --output-dir ./test_quant

# 验证输出分布
./inference.sh ./test_quant/best26n.bin ./cal/frame_000118.jpg

预期输出（正常）：

output[0]: shape=(1, 80, 80, 9), min=-200, max=80  # cls
output[1]: shape=(1, 80, 80, 4), min=-5, max=30   # box

步骤 3：检查量化配置参数

创建显式配置文件 yolo26_config.yaml：

model:
  model_name: "yolo26n"
  input_type: "rgb"  # ⚠️ 必须与训练时一致
  input_type_rt: "rgb"
  normalization: [0.0, 1.0, 0.0, 1.0, 0.0, 1.0]
  mean: [0, 0, 0]
  std: [255, 255, 255]  # 训练时如果除以255，这里也要对应
  input_shape: [1, 3, 640, 640]

quantization:
  method: "default"
  calibration_dataset: "./cal/"
  calibration_size: 200  # 增加到200张
  batch_size: 1

compiler:
  target_chip: "x5"
  optimization_level: 2

重新量化：

python mapper.py \
  --config yolo26_config.yaml \
  --onnx ./best26n.onnx

步骤 4：临时验证（降低阈值）

如果急需测试，大幅降低置信度阈值验证是否是置信度问题：

# 修改 detect.py 或你的推理代码
score_thres = 0.05  # 从 0.25 降到 0.05
nms_thres = 0.5

# 如果降低阈值后检测数量明显增加
# 说明是量化导致置信度整体偏低

步骤 5：尝试跳过 YOLO Head 量化

如果上述步骤无效，尝试跳过输出层量化：

quantization:
  # 添加跳过层（根据实际模型结构调整）
  skip_layers: [
    "model.23/one2one_cv2.2/one2one_cv2.2.2/Conv",
    "model.23/one2one_cv3.2/one2one_cv3.2.2/Conv"
  ]

快速验证清单

检查项	命令/位置	预期值
训练输入格式	检查训练代码	RGB 或 BGR
量化 input_type	`mapper.py` 或 yaml	与训练一致
归一化参数	训练 vs 量化配置	只归一化一次
校准数据数量	`ls cal/*.jpg	wc -l`
校准数据质量	覆盖多种场景	光照/角度/大小多样
量化后输出范围	inference.sh 验证	cls: [-200, 100]

参考资源

建议

先用官方预训练模型跑通全流程，确认环境没问题
校准数据用训练集的一部分（200 张），不要用测试集
对比输出范围：官方模型量化后 vs 你的模型量化后
如果问题仍未解决，请补充：
- 训练时的预处理代码片段
- 完整的量化命令和配置文件
- 量化前后的输出范围对比

社区已有多个类似案例，基本都是输入格式不一致或校准数据问题导致的。

Marcelo6151 · 2026 年4 月 30 日 06:13

你好，收到问题，可以先用官方的预训练模型测试一下？

地瓜小助手_bot · 2026 年4 月 30 日 06:14

问题分析

@VelumZhang 你好，从你的日志可以明确看到量化后 cls 输出分布严重异常：

阶段	cls 输出范围	状态
ONNX	`[-229, 82]`	正常
量化后	`[-14, -0.1]`	全为负值且范围极窄

这导致经过 sigmoid 后置信度只有 0.0001-0.47，大量目标被阈值过滤掉。

快速排查方案

先用官方预训练模型验证流程

目的：确认你的量化环境和流程本身没问题。

cd rdk_model_zoo/samples/vision/ultralytics_yolo26/conversion

# 下载官方预训练模型（如果还没有）
wget https://github.com/ultralytics/assets/releases/download/v0.0.0/yolo11n.pt  # 或对应版本

# 导出 ONNX
python export_yolo26_detect_bpu.py \
  --weights yolo11n.pt \
  --output yolo11n.onnx \
  --imgsz 640

# 使用仓库自带的 cal/ 校准数据量化
python mapper.py \
  --onnx ./yolo11n.onnx \
  --cal-images ./cal/ \
  --cal-sample-num 100

# 验证输出分布
./inference.sh ./yolo11n_bayese_640x640_nv12.bin ./cal/frame_000118.jpg

预期正常输出：

output[0]: shape=(1, 80, 80, 9), min=-200, max=80   # cls ✅
output[1]: shape=(1, 80, 80, 4), min=-5, max=30    # box ✅

如果官方模型量化后输出正常，说明你的模型或校准数据有问题。

检查输入格式一致性（最关键）

三者必须完全一致：训练 → 导出 → 量化

# 检查你的训练代码预处理
# 关键问题：
# 1. 用 RGB 还是 BGR？
# 2. 是否除以 255？
# 3. mean/std 是多少？

# 查看 mapper.py 默认配置
grep -A 15 "input_type" mapper.py

常见错误：

训练用 RGB，量化配置用 BGR
训练时已 /255，量化时又重复归一化

校准数据问题排查

你的校准数据 ./cal/ 需要满足：

# 检查数量和格式
ls ./cal/*.jpg | wc -l  # 至少 100 张，建议 200 张

# 检查数据分布
python -c "
import cv2, numpy as np, os
imgs = [cv2.imread(f'./cal/{f}') for f in os.listdir('./cal') if f.endswith('.jpg')][:100]
print(f'数量：{len(imgs)}')
print(f'均值：{np.mean(imgs):.2f}, 标准差：{np.std(imgs):.2f}')
print(f'范围：[{np.min(imgs)}, {np.max(imgs)}]')
"

建议：直接用训练集的一部分（200 张）作为校准数据，不要用测试集。

显式配置量化参数

创建 yolo26_config.yaml：

model:
  model_name: "yolo26n"
  input_type: "rgb"  # ⚠️ 必须与训练一致
  input_type_rt: "rgb"
  normalization: [0.0, 1.0, 0.0, 1.0, 0.0, 1.0]
  mean: [0, 0, 0]
  std: [255, 255, 255]  # 训练时如果除以255，这里对应
  input_shape: [1, 3, 640, 640]

quantization:
  method: "default"
  calibration_dataset: "./cal/"
  calibration_size: 200
  batch_size: 1

compiler:
  target_chip: "x5"
  optimization_level: 2

重新量化：

python mapper.py --config yolo26_config.yaml --onnx ./best26n.onnx

临时验证（降低阈值）

如果急需测试，先降低阈值确认是否是置信度问题：

# 修改推理代码
score_thres = 0.05  # 从 0.25 降到 0.05
nms_thres = 0.5

# 如果检测数量明显增加 → 量化导致置信度偏低

参考资源

下一步建议

先用官方预训练模型 + 自带 cal/ 数据跑通全流程
对比输出范围：官方模型量化后 vs 你的模型量化后
如果仍有问题，请补充：
- 训练时的预处理代码片段
- 完整的量化命令
- 量化前后的输出范围对比

社区已有多个类似案例，基本都是输入格式不一致或校准数据问题导致的。

VelumZhang · 2026 年4 月 30 日 06:33

我现在没有coco数据集暂时没法使用官方的进行量化操作，我的数据是三通道的灰度图像是否会有影响？校准数据和训练数据是一样的

地瓜小助手_bot · 2026 年4 月 30 日 06:34

关键问题定位

三通道灰度图本身没问题，但需要确保训练→导出→量化三阶段的输入处理完全一致。

核心排查点

确认你的训练输入格式

三通道灰度图有两种常见训练方式：

# 方式A：直接复制灰度到三通道（推荐）
gray = cv2.imread(img_path, cv2.IMREAD_GRAYSCALE)  # [H, W]
rgb_gray = cv2.cvtColor(gray, cv2.COLOR_GRAY2RGB)  # [H, W, 3] R=G=B

# 方式B：用三通道图像但实际是灰度
rgb = cv2.imread(img_path, cv2.IMREAD_COLOR)  # [H, W, 3] 但 R≈G≈B

关键问题：你的训练代码用的是哪种？归一化参数是什么？

# 训练时是否这样处理？
img = img.astype(np.float32) / 255.0  # 归一化到 [0, 1]
# 还是
img = (img - mean) / std  # 使用 mean/std 归一化

量化配置必须与训练一致

根据你的情况，创建专用配置文件 yolo26_gray_config.yaml：

model:
  model_name: "yolo26n_gray"
  # ⚠️ 关键：即使训练用的是灰度图，只要输入是3通道，这里仍然是 rgb
  input_type: "rgb"  # 或 "bgr"，取决于训练时 cv2.imread 的方式
  input_type_rt: "rgb"
  
  # ⚠️ 归一化参数必须与训练时完全一致
  # 如果训练时是 img/255.0：
  normalization: [0.0, 1.0, 0.0, 1.0, 0.0, 1.0]
  mean: [0, 0, 0]
  std: [255, 255, 255]
  
  # 如果训练时用了 mean/std 归一化（如 ImageNet）：
  # mean: [0.485, 0.456, 0.406]
  # std: [0.229, 0.224, 0.225]
  
  input_shape: [1, 3, 640, 640]

quantization:
  method: "default"
  calibration_dataset: "./cal/"  # 你的校准数据
  calibration_size: 200  # 建议≥200张
  batch_size: 1

compiler:
  target_chip: "x5"
  optimization_level: 2

验证校准数据格式

# 检查你的校准数据是否真的是三通道
python -c "
import cv2, numpy as np, os

img_path = './cal/frame001.jpg'  # 替换为你的文件
img = cv2.imread(img_path)
print(f'形状：{img.shape}')  # 应该是 (640, 640, 3)
print(f'通道0-1均值差：{abs(img[:,:,0].mean() - img[:,:,1].mean()):.4f}')
print(f'通道1-2均值差：{abs(img[:,:,1].mean() - img[:,:,2].mean()):.4f}')
print(f'如果是灰度图，三个通道应该几乎相等（差值<1）')

# 检查所有校准图片
imgs = [cv2.imread(f'./cal/{f}') for f in os.listdir('./cal') if f.endswith('.jpg')][:100]
shapes = [img.shape for img in imgs]
print(f'\n校准数据形状分布：{set(shapes)}')
print(f'数量：{len(imgs)}')
"

预期输出（三通道灰度图）：

形状：(640, 640, 3)
通道0-1均值差：0.0000
通道1-2均值差：0.0000

重新量化并验证

cd rdk_model_zoo/samples/vision/ultralytics_yolo26/conversion

# 使用自定义配置重新量化
python mapper.py \
  --config yolo26_gray_config.yaml \
  --onnx ./best26n.onnx \
  --cal-images ./cal/ \
  --cal-sample-num 200

# 验证输出分布
./inference.sh ./best26n_bayese_640x640_nv12.bin ./cal/frame001.jpg

正常输出应该是：

output[0]: shape=(1, 80, 80, 9), min=-200, max=80   # cls ✅
output[1]: shape=(1, 80, 80, 4), min=-5, max=30    # box ✅

常见灰度图量化坑点

问题	症状	解决方案
训练用灰度，量化当RGB	cls 输出异常	确认 `input_type` 与训练一致
归一化重复	输出值范围过小	训练和量化只归一化一次
校准数据通道数不对	量化报错或精度差	确保校准数据是 3 通道
校准数据数量不足	量化不稳定	至少 100 张，建议 200 张

快速验证步骤

# 1. 检查校准数据
python -c "import cv2; img=cv2.imread('./cal/frame001.jpg'); print(f'Shape: {img.shape}, Dtype: {img.dtype}')"

# 2. 检查训练预处理代码（找到你的训练脚本）
grep -A 5 "transform\|normalize" your_train_script.py

# 3. 用当前配置重新量化
python mapper.py --onnx ./best26n.onnx --cal-images ./cal/ --cal-sample-num 200

# 4. 验证输出
./inference.sh ./best26n_bayese_640x640_nv12.bin ./cal/frame001.jpg

如果问题仍未解决

请补充以下信息：

训练预处理代码片段（如何读取和处理图像）
校准数据检查结果（shape、通道值）
量化后的输出范围（inference.sh 的完整输出）

三通道灰度图量化是可行的，社区有成功案例，关键是输入格式配置要与训练时完全一致。

Marcelo6151 · 2026 年4 月 30 日 07:02

coco数据集只需要准备20张用于校准即可

目前量化yaml 校准数据集准备推理都是以 RGB转NV12进行准备的，这块可能需要按照灰度的输入修改mapper和推理脚本了

model zoo目前提供的量化方案为使用原始官方预训练模型和对应coco数据集方案

VelumZhang · 2026 年4 月 30 日 09:01

使用coco数据集量化导出无异常，可能就是不支持灰度图，即使是三通道的。

地瓜小助手_bot · 2026 年4 月 30 日 09:02

问题确认

@VelumZhang 你的判断基本正确：rdk_model_zoo 默认流程确实是为 RGB 彩色图像设计的，对于灰度图（即使复制成三通道）需要特殊配置。

但三通道灰度图本身是可以量化的，关键在于校准数据的生成方式和量化配置参数需要调整。

灰度图量化解决方案

方案 A：修改校准数据生成方式（推荐）

model_zoo 的 mapper.py 默认会将校准图片从 BGR/RGB 转换为 NV12 格式，这个转换过程对灰度图不友好。

解决步骤：

1. 预处理校准数据为 NV12 格式

# 创建预处理脚本：preprocess_gray_cal.py
import cv2
import os
import numpy as np

def rgb2nv12(rgb_img):
    """RGB 转 NV12"""
    yuv = cv2.cvtColor(rgb_img, cv2.COLOR_RGB2YUV_I420)
    y = yuv[0]
    u = yuv[1]
    v = yuv[2]
    # NV12 格式：Y 平面 + UV 交错平面
    uv = cv2.merge([u, v])
    uv = uv.reshape(u.shape[0] * 2, u.shape[1])
    nv12 = np.vstack([y, uv])
    return nv12

cal_dir = './cal_raw'  # 原始灰度图目录
output_dir = './cal_nv12'
os.makedirs(output_dir, exist_ok=True)

for fname in os.listdir(cal_dir):
    if fname.endswith('.jpg') or fname.endswith('.png'):
        # 读取灰度图
        gray = cv2.imread(os.path.join(cal_dir, fname), cv2.IMREAD_GRAYSCALE)
        # 复制成三通道（与训练时一致）
        rgb_gray = cv2.cvtColor(gray, cv2.COLOR_GRAY2RGB)
        # 转 NV12
        nv12 = rgb2nv12(rgb_gray)
        # 保存为二进制或 npy 格式
        np.save(os.path.join(output_dir, fname.replace('.jpg', '.npy')), nv12)
        print(f'Processed: {fname}')

2. 修改量化配置

# yolo26_gray_config.yaml
model:
  model_name: "yolo26n_gray"
  input_type: "rgb"  # 保持 rgb，因为模型输入是 3 通道
  input_type_rt: "nv12"  # 推理时输入 NV12
  normalization: [0.0, 1.0, 0.0, 1.0, 0.0, 1.0]
  mean: [0, 0, 0]
  std: [255, 255, 255]
  input_shape: [1, 3, 640, 640]

quantization:
  method: "default"
  calibration_dataset: "./cal_nv12/"  # 使用预处理后的 NV12 数据
  calibration_size: 100
  batch_size: 1
  # 关键：跳过某些敏感层
  skip_layers: []

compiler:
  target_chip: "x5"
  optimization_level: 2

3. 修改 `mapper.py` 调用方式

# 可能需要修改 mapper.py 中的输入数据加载逻辑
# 或者使用 hb_mapper 直接命令：

hb_mapper makertbin \
  --model-type onnx \
  --config yolo26_gray_config.yaml \
  --march bayes-e \
  --output-dir ./quantized_gray

方案 B：使用单通道模型（更彻底）

如果你的应用确实只需要灰度图，可以考虑修改模型结构为单通道输入：

1. 修改模型输入层

# 训练时修改模型配置
# yolov26n.yaml 中修改第一层：
# 从：
# - [[-1, 1, Conv, [64, 3, 2]]]  # 输入 3 通道
# 改为：
# - [[-1, 1, Conv, [64, 1, 2]]]  # 输入 1 通道

2. 量化配置调整

model:
  input_type: "gray"  # 如果工具链支持
  input_shape: [1, 1, 640, 640]  # 单通道

注意：此方案需要修改模型结构并重新训练，工作量较大。

方案 C：临时验证（快速确认问题）

先用以下方法快速确认是否是灰度图导致的问题：

# 1. 从训练集中选 20 张彩色图作为校准数据
mkdir -p ./cal_color
# 复制 20 张彩色图到 ./cal_color/

# 2. 用彩色校准数据量化你的灰度模型
python mapper.py \
  --onnx ./best26n.onnx \
  --cal-images ./cal_color/ \
  --cal-sample-num 20

# 3. 验证输出分布
./inference.sh ./best26n_bayese_640x640_nv12.bin ./cal_color/frame001.jpg

# 如果输出正常（cls: [-200, 80]），说明确实是灰度校准数据的问题

关键结论

问题	原因	解决方案
灰度图量化失败	model_zoo 默认 RGB→NV12 转换对灰度图不友好	预处理校准数据为 NV12 格式
cls 输出异常	校准数据分布与训练数据差异大	使用与训练时完全一致的预处理流程
工具链不支持	默认配置针对彩色图像优化	修改 yaml 配置或校准数据生成方式

建议操作顺序

先用方案 C 快速验证：用 20 张彩色图校准，确认是否是灰度图问题
如果确认是灰度图问题：使用方案 A 预处理校准数据为 NV12
如果仍无法解决：考虑方案 B（单通道模型），但需要重新训练
同时建议：在社区发起一个灰度图量化的讨论帖，可能会有官方支持

参考资源

如果需要更详细的灰度图量化配置，建议在社区发起专门的技术讨论帖，@官方技术支持获取更准确的配置参数。

Marcelo6151 · 2026 年5 月 6 日 02:35

你好，支持灰度，需要自己改一下逻辑，目前提供的版本是没有做灰度的方案

rdk x5使用model_zoo 量化部署yolo26，几乎无法使用

问题分析

排查步骤

步骤 1：检查输入格式一致性

步骤 2：使用官方校准数据集测试

步骤 3：检查量化配置参数

步骤 4：临时验证（降低阈值）

步骤 5：尝试跳过 YOLO Head 量化

快速验证清单

参考资源

建议

问题分析

快速排查方案

先用官方预训练模型验证流程

检查输入格式一致性（最关键）

校准数据问题排查

显式配置量化参数

临时验证（降低阈值）

参考资源

下一步建议

关键问题定位

核心排查点

确认你的训练输入格式

量化配置必须与训练一致

验证校准数据格式

重新量化并验证

常见灰度图量化坑点

快速验证步骤

如果问题仍未解决

问题确认

灰度图量化解决方案

方案 A：修改校准数据生成方式（推荐）

1. 预处理校准数据为 NV12 格式

2. 修改量化配置

3. 修改 mapper.py 调用方式

方案 B：使用单通道模型（更彻底）

1. 修改模型输入层

2. 量化配置调整

方案 C：临时验证（快速确认问题）

关键结论

建议操作顺序

参考资源

3. 修改 `mapper.py` 调用方式