YOLOv8量化模型无法识别目标

76668800 · 2025 年6 月 25 日 07:08

板子型号为RDKx5，量化后的yolov8_detect_bayese_640x640_nv12_modified.bin 检验识别效果时无法检测到目标。

  **hb_model_modifier.log**

2025-03-26 19:49:00,665 file: tool_utils.py func: tool_utils line No: 77 log will be stored in /yolo_files/yolov8_detect_bayese_640x640_nv12/hb_model_modifier.log
2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 395 storing info of model: yolov8_detect_bayese_640x640_nv12.bin
2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 398 --------input----------
2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 400 name: “images”
type {
elem_type: INT8
dim: 1
dim: 640
dim: 640
dim: 3
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 401 --------output----------
2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 403 name: “output0”
type {
elem_type: FLOAT
dim: 1
dim: 80
dim: 80
dim: 64
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 403 name: “326”
type {
elem_type: FLOAT
dim: 1
dim: 80
dim: 80
dim: 3
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 403 name: “334”
type {
elem_type: FLOAT
dim: 1
dim: 40
dim: 40
dim: 64
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 403 name: “342”
type {
elem_type: FLOAT
dim: 1
dim: 40
dim: 40
dim: 3
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 403 name: “350”
type {
elem_type: FLOAT
dim: 1
dim: 20
dim: 20
dim: 64
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 403 name: “358”
type {
elem_type: FLOAT
dim: 1
dim: 20
dim: 20
dim: 3
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 404 --------value_info----------
2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 406 name: “/model.22/cv2.0/cv2.0.2/Conv_output_0_quantized”
type {
elem_type: INT32
dim: 1
dim: 80
dim: 80
dim: 64
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 406 name: “/model.22/cv3.0/cv3.0.2/Conv_output_0_quantized”
type {
elem_type: INT32
dim: 1
dim: 80
dim: 80
dim: 3
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 406 name: “/model.22/cv2.1/cv2.1.2/Conv_output_0_quantized”
type {
elem_type: INT32
dim: 1
dim: 40
dim: 40
dim: 64
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 406 name: “/model.22/cv3.1/cv3.1.2/Conv_output_0_quantized”
type {
elem_type: INT32
dim: 1
dim: 40
dim: 40
dim: 3
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 406 name: “/model.22/cv2.2/cv2.2.2/Conv_output_0_quantized”
type {
elem_type: INT32
dim: 1
dim: 20
dim: 20
dim: 64
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 406 name: “/model.22/cv3.2/cv3.2.2/Conv_output_0_quantized”
type {
elem_type: INT32
dim: 1
dim: 20
dim: 20
dim: 3
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 407 --------node----------
2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 409 input: “images”
output: “/model.22/cv2.0/cv2.0.2/Conv_output_0_quantized”
output: “/model.22/cv3.0/cv3.0.2/Conv_output_0_quantized”
output: “/model.22/cv2.1/cv2.1.2/Conv_output_0_quantized”
output: “/model.22/cv3.1/cv3.1.2/Conv_output_0_quantized”
output: “/model.22/cv2.2/cv2.2.2/Conv_output_0_quantized”
output: “/model.22/cv3.2/cv3.2.2/Conv_output_0_quantized”
name: “torch_jit_subgraph_0”
op_type: “BPU”

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 409 input: “/model.22/cv2.0/cv2.0.2/Conv_output_0_quantized”
input: “/model.22/cv2.0/cv2.0.2/Conv_x_scale”
output: “output0”
name: “/model.22/cv2.0/cv2.0.2/Conv_output_0_HzDequantize”
op_type: “Dequantize”
attribute {
name: “axis”
type: INT
i: 3
}
attribute {
name: “output_data_format”
type: STRING
s: “NHWC”
}
attribute {
name: “num_args”
type: INT
i: 2
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 409 input: “/model.22/cv3.0/cv3.0.2/Conv_output_0_quantized”
input: “/model.22/cv3.0/cv3.0.2/Conv_x_scale”
output: “326”
name: “/model.22/cv3.0/cv3.0.2/Conv_output_0_HzDequantize”
op_type: “Dequantize”
attribute {
name: “axis”
type: INT
i: 3
}
attribute {
name: “output_data_format”
type: STRING
s: “NHWC”
}
attribute {
name: “num_args”
type: INT
i: 2
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 409 input: “/model.22/cv2.1/cv2.1.2/Conv_output_0_quantized”
input: “/model.22/cv2.1/cv2.1.2/Conv_x_scale”
output: “334”
name: “/model.22/cv2.1/cv2.1.2/Conv_output_0_HzDequantize”
op_type: “Dequantize”
attribute {
name: “axis”
type: INT
i: 3
}
attribute {
name: “output_data_format”
type: STRING
s: “NHWC”
}
attribute {
name: “num_args”
type: INT
i: 2
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 409 input: “/model.22/cv3.1/cv3.1.2/Conv_output_0_quantized”
input: “/model.22/cv3.1/cv3.1.2/Conv_x_scale”
output: “342”
name: “/model.22/cv3.1/cv3.1.2/Conv_output_0_HzDequantize”
op_type: “Dequantize”
attribute {
name: “axis”
type: INT
i: 3
}
attribute {
name: “output_data_format”
type: STRING
s: “NHWC”
}
attribute {
name: “num_args”
type: INT
i: 2
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 409 input: “/model.22/cv2.2/cv2.2.2/Conv_output_0_quantized”
input: “/model.22/cv2.2/cv2.2.2/Conv_x_scale”
output: “350”
name: “/model.22/cv2.2/cv2.2.2/Conv_output_0_HzDequantize”
op_type: “Dequantize”
attribute {
name: “axis”
type: INT
i: 3
}
attribute {
name: “output_data_format”
type: STRING
s: “NHWC”
}
attribute {
name: “num_args”
type: INT
i: 2
}

2025-03-26 19:49:00,667 file: hb_model_modifier.py func: hb_model_modifier line No: 409 input: “/model.22/cv3.2/cv3.2.2/Conv_output_0_quantized”
input: “/model.22/cv3.2/cv3.2.2/Conv_x_scale”
output: “358”
name: “/model.22/cv3.2/cv3.2.2/Conv_output_0_HzDequantize”
op_type: “Dequantize”
attribute {
name: “axis”
type: INT
i: 3
}
attribute {
name: “output_data_format”
type: STRING
s: “NHWC”
}
attribute {
name: “num_args”
type: INT
i: 2
}

2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 410 --------input_type----------
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 412 7
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 413 --------input_layout----------
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 415 2
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 108 Nodes that will be removed from this model: [‘/model.22/cv2.0/cv2.0.2/Conv_output_0_HzDequantize’, ‘/model.22/cv2.1/cv2.1.2/Conv_output_0_HzDequantize’, ‘/model.22/cv2.2/cv2.2.2/Conv_output_0_HzDequantize’]
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 190 Node ‘/model.22/cv2.0/cv2.0.2/Conv_output_0_HzDequantize’ found, its OP type is ‘Dequantize’
2025-03-26 19:49:00,668 file: model_utils.py func: model_utils line No: 284 replace output name /model.22/cv2.0/cv2.0.2/Conv_output_0_quantized with output0
2025-03-26 19:49:00,668 file: model_utils.py func: model_utils line No: 284 replace output name /model.22/cv2.0/cv2.0.2/Conv_x_scale with output0
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 352 scale: /model.22/cv2.0/cv2.0.2/Conv_x_scale; zero point: 0. node info details are stored in hb_model_modifier log file
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 389 Node /model.22/cv2.0/cv2.0.2/Conv_output_0_HzDequantize info: ‘/model.22/cv2.0/cv2.0.2/Conv_output_0_HzDequantize: scale: [0.0007072120788507164, 0.0006784849101677537, 0.0006074502598494291, 0.000656547665130347, 0.0004839231842197478, 0.0006199857452884316, 0.0004520620859693736, 0.0005251858965493739, 0.00040844891918823123, 0.0003261845849920064, 0.00035047216806560755, 0.0002838772488757968, 0.0002038327365880832, 0.00016048073302954435, 0.000137760114739649, 0.0001566939608892426, 0.0006278204964473844, 0.0006006602197885513, 0.0006131957052275538, 0.0005437280633486807, 0.0005421611131168902, 0.00048679590690881014, 0.0005604421021416783, 0.00043717617518268526, 0.00043430342338979244, 0.0002969350607600063, 0.000360918405931443, 0.00038050516741350293, 0.0003418539999984205, 0.00026820783386938274, 0.00021023106819484383, 0.000268468982540071, 0.0007563095423392951, 0.000685797247570008, 0.000579245388507843, 0.0006581146735697985, 0.0005980486166663468, 0.0004175894137006253, 0.0005703659844584763, 0.00048340088687837124, 0.0003099928726442158, 0.0003791993949562311, 0.0003384589508641511, 0.0002729086554609239, 0.0002074889198411256, 0.00013984936231281608, 0.00012359237007331103, 0.00017301626212429255, 0.0006523691699840128, 0.0005964816664345562, 0.0005358933703973889, 0.0005176123813726008, 0.0005405942210927606, 0.0006225973484106362, 0.0005139561835676432, 0.0003875564143527299, 0.0003572622372303158, 0.0003149548720102757, 0.00022015502327121794, 0.00015356008952949196, 0.00017366913380101323, 0.00017079641111195087, 0.00015630222333129495, 0.0001952145539689809]; zero point: {zero_point}’
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 391 Node ‘/model.22/cv2.0/cv2.0.2/Conv_output_0_HzDequantize’ is removed
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 190 Node ‘/model.22/cv2.1/cv2.1.2/Conv_output_0_HzDequantize’ found, its OP type is ‘Dequantize’
2025-03-26 19:49:00,668 file: model_utils.py func: model_utils line No: 284 replace output name /model.22/cv2.1/cv2.1.2/Conv_output_0_quantized with 334
2025-03-26 19:49:00,668 file: model_utils.py func: model_utils line No: 284 replace output name /model.22/cv2.1/cv2.1.2/Conv_x_scale with 334
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 352 scale: /model.22/cv2.1/cv2.1.2/Conv_x_scale; zero point: 0. node info details are stored in hb_model_modifier log file
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 389 Node /model.22/cv2.1/cv2.1.2/Conv_output_0_HzDequantize info: ‘/model.22/cv2.1/cv2.1.2/Conv_output_0_HzDequantize: scale: [0.0007815019343979657, 0.0007429233519360423, 0.0007009071414358914, 0.0007276447722688317, 0.0006283336551859975, 0.0005057226517237723, 0.0005034308414906263, 0.000345297041349113, 0.000449955667136237, 0.00039800829836167395, 0.00028972100699320436, 0.0003300184034742415, 0.00037547232932411134, 0.00037241660174913704, 0.0003393765655346215, 0.0003674510517157614, 0.0007860855548642576, 0.0007746265619061887, 0.0008120592101477087, 0.0006982333725318313, 0.0007669872138649225, 0.0005859354278072715, 0.0005301684141159058, 0.00038884111563675106, 0.0004423163481988013, 0.0003046176861971617, 0.00032428893609903753, 0.00028341860161162913, 0.0002664211206138134, 0.00029334970167838037, 0.00029678738792426884, 0.0003494986449368298, 0.0009083145414479077, 0.0008792851585894823, 0.0007742445450276136, 0.0006657663034275174, 0.0005771502037532628, 0.0004896800965070724, 0.0004362048639450222, 0.00039991814992390573, 0.00032027828274294734, 0.00036401336546987295, 0.00034892570693045855, 0.00029086691210977733, 0.0002799808862619102, 0.0003030898224096745, 0.00031072914134711027, 0.0003443421155679971, 0.0007761543965898454, 0.0007696609827689826, 0.0007627855520695448, 0.0006925039342604578, 0.0007066366379149258, 0.0006967055378481746, 0.0005951026105321944, 0.000391132925869897, 0.0003886501654051244, 0.00035389125696383417, 0.0003626764810178429, 0.000298315251711756, 0.00032562579144723713, 0.00039877224480733275, 0.0004255098174326122, 0.00048662436893209815]; zero point: {zero_point}’
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 391 Node ‘/model.22/cv2.1/cv2.1.2/Conv_output_0_HzDequantize’ is removed
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 190 Node ‘/model.22/cv2.2/cv2.2.2/Conv_output_0_HzDequantize’ found, its OP type is ‘Dequantize’
2025-03-26 19:49:00,668 file: model_utils.py func: model_utils line No: 284 replace output name /model.22/cv2.2/cv2.2.2/Conv_output_0_quantized with 350
2025-03-26 19:49:00,668 file: model_utils.py func: model_utils line No: 284 replace output name /model.22/cv2.2/cv2.2.2/Conv_x_scale with 350
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 352 scale: /model.22/cv2.2/cv2.2.2/Conv_x_scale; zero point: 0. node info details are stored in hb_model_modifier log file
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 389 Node /model.22/cv2.2/cv2.2.2/Conv_output_0_HzDequantize info: ‘/model.22/cv2.2/cv2.2.2/Conv_output_0_HzDequantize: scale: [0.0013486521784216166, 0.0013626321451738477, 0.0013815461425110698, 0.0015657522017136216, 0.0013174029299989343, 0.001257371506653726, 0.0011216838611289859, 0.0010797440772876143, 0.0008609993383288383, 0.0008560652495361865, 0.0008848473662510514, 0.0006064822664484382, 0.0005966140888631344, 0.0005583748570643365, 0.00041158561361953616, 0.00023066885478328913, 0.0013100018259137869, 0.001300133648328483, 0.0013412510743364692, 0.0014604916796088219, 0.0012746407883241773, 0.001198984682559967, 0.0012146092485636473, 0.000920208403840661, 0.0009785952279344201, 0.0011603343300521374, 0.0007018747855909169, 0.0007347687496803701, 0.0005501513951458037, 0.00039226043736562133, 0.00015655466995667666, 7.971644663484767e-05, 0.0013576979981735349, 0.0014177295379340649, 0.0012721738312393427, 0.0012631278950721025, 0.001185004715807736, 0.0010665865847840905, 0.0011948728933930397, 0.0008392070303671062, 0.0009070508531294763, 0.0009095178684219718, 0.0008922485867515206, 0.0005777000915259123, 0.0005160238943062723, 0.0005102674476802349, 0.00036203907802700996, 0.0002436208596918732, 0.0014596692053601146, 0.001352763967588544, 0.0012581938644871116, 0.0010460277553647757, 0.0012384573929011822, 0.001051784260198474, 0.001134019112214446, 0.0008947155438363552, 0.0009933974361047149, 0.0008309835102409124, 0.0007997343200258911, 0.0006307415897026658, 0.0005661871982738376, 0.0006611684802919626, 0.0004888864350505173, 0.0003015964466612786]; zero point: {zero_point}’
2025-03-26 19:49:00,668 file: hb_model_modifier.py func: hb_model_modifier line No: 391 Node ‘/model.22/cv2.2/cv2.2.2/Conv_output_0_HzDequantize’ is removed
2025-03-26 19:49:00,670 file: hb_model_modifier.py func: hb_model_modifier line No: 130 modified model saved as yolov8_detect_bayese_640x640_nv12_modified.bin

CauchyKesai · 2025 年6 月 25 日 12:56

你好，在算法开发的过程中，遇到各种数值不可控的问题都是正常的，算法开发本身就是需要厚积薄发的领域。算法工具链提供了完整的流程说明，debug工具及流程说明，供您参考。 PTQ流程详解：6.1. PTQ转换原理及流程 — Horizon Open Explorer

精度调优：8.2. PTQ模型精度调优 — Horizon Open Explorer

性能调优：8.1. 模型性能调优 — Horizon Open Explorer

精度debug工具详解：6.2.12. 精度debug工具 — Horizon Open Explorer

Runtime程序编写详解：9. 嵌入式应用开发（runtime）手册 — Horizon Open Explorer

如果将工具链手册所述的所有流程走完仍然不及预期，则说明模型及其权重本身无法量化。特别的，过拟合的模型本身容易出现异常值导致量化表示能力不足。