rknn-toolkit2
rknn-toolkit2 copied to clipboard
rockchip-linux
yolov5s在RKNN3588上以FP16推理时的精度问题
通过本地的ubuntu服务器对yolov5s模型进行连板调试时，fp16的精度下降了4个点，请问这个现象正常吗？
rknntoolkit2的版本如下：
rknn-toolkit2 version: 2.0.0b0+9bab5682
rk3588的驱动版本如下：
D RKNNAPI: API: 2.0.0b0 (18eacd0 build@2024-03-22T06:07:59)
D RKNNAPI: DRV: rknn_server: 2.0.0b0 (18eacd0 build@2024-03-22T14:07:19)
D RKNNAPI: DRV: rknnrt: 2.0.0b0 (35a6907d79@2024-03-24T10:31:14)
rknn.config的参数如下：
rknn.config(mean_values=[[0, 0, 0]],
std_values=[[255, 255, 255]],
target_platform="rk3588")
精度分析结果如下：
rknn convert start!
I rknn-toolkit2 version: 2.0.0b0+9bab5682
--> Config model
done
--> Loading model
I It is recommended onnx opset 19, but your onnx model opset is 13!
I Model converted from pytorch, 'opset_version' should be set 19 in torch.onnx.export for successful convert!
I Loading : 100%|██████████████████████████████████████████████| 121/121 [00:00<00:00, 39667.87it/s]
done
--> Building model
W build: The dataset='calibrate_dataset.txt' is ignored because do_quantization = False!
I rknn building ...
I rknn buiding done.
done
--> Export rknn model
done
--> Accuracy analysis
adb: unable to connect for root: closed
I target set by user is: rk3588
I Get hardware info: target_platform = rk3588, os = Linux, aarch = aarch64
I Check RK3588 board npu runtime version
I Starting ntp or adb, target is RK3588
I Start adb...
I Connect to Device success!
I NPUTransfer: Starting NPU Transfer Client, Transfer version 2.1.0 (b5861e7@2020-11-23T11:50:36)
D RKNNAPI: ==============================================
D RKNNAPI: RKNN VERSION:
D RKNNAPI: API: 2.0.0b0 (18eacd0 build@2024-03-22T06:07:59)
D RKNNAPI: DRV: rknn_server: 2.0.0b0 (18eacd0 build@2024-03-22T14:07:19)
D RKNNAPI: DRV: rknnrt: 2.0.0b0 (35a6907d79@2024-03-24T10:31:14)
D RKNNAPI: ==============================================
D RKNNAPI: Input tensors:
D RKNNAPI: index=0, name=images, n_dims=4, dims=[1, 640, 640, 3], n_elems=1228800, size=2457600, w_stride = 0, size_with_stride = 0, fmt=NHWC, type=FP16, qnt_type=NONE, zp=0, scale=1.000000
D RKNNAPI: Output tensors:
D RKNNAPI: index=0, name=output1, n_dims=4, dims=[1, 18, 20, 20], n_elems=7200, size=14400, w_stride = 0, size_with_stride = 0, fmt=NCHW, type=FP16, qnt_type=NONE, zp=0, scale=1.000000
D RKNNAPI: index=1, name=output2, n_dims=4, dims=[1, 18, 40, 40], n_elems=28800, size=57600, w_stride = 0, size_with_stride = 0, fmt=NCHW, type=FP16, qnt_type=NONE, zp=0, scale=1.000000
D RKNNAPI: index=2, name=output3, n_dims=4, dims=[1, 18, 80, 80], n_elems=115200, size=230400, w_stride = 0, size_with_stride = 0, fmt=NCHW, type=FP16, qnt_type=NONE, zp=0, scale=1.000000
adb: unable to connect for root: closed
/userdata/dumps/: 89 files pulled. 4.2 MB/s (149901680 bytes in 33.662s)
I Save Tensors to txt: 100%|████████████████████████████████████████| 89/89 [00:03<00:00, 27.86it/s]
I GraphPreparing : 100%|███████████████████████████████████████| 145/145 [00:00<00:00, 15836.22it/s]
I AccuracyAnalysing : 100%|███████████████████████████████████████| 145/145 [00:16<00:00, 8.98it/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.
# runtime_error: calculate the output error of each layer of the runtime.
# entire: output error of each layer between 'golden' and 'runtime', these errors will accumulate layer by layer.
# single_sim: single-layer output error between 'simulator' and 'runtime', can better reflect the single-layer accuracy of runtime.
layer_name simulator_error runtime_error
entire single entire single_sim
cos euc cos euc cos euc cos euc
-----------------------------------------------------------------------------------------------------------------------------------
[Input] images 1.00000 | 0.0 1.00000 | 0.0 1.00000 | 0.1180 1.00000 | 0.1180
[Conv] /model.0/conv/Conv_output_0 1.00000 | 2.4982 1.00000 | 2.4982
[exSwish] /model.0/act/Mul_output_0 1.00000 | 1.9841 1.00000 | 1.3346 1.00000 | 10.165 1.00000 | 10.132
[Conv] /model.1/conv/Conv_output_0 1.00000 | 2.4134 1.00000 | 1.8541
[exSwish] /model.1/act/Mul_output_0 1.00000 | 1.9273 1.00000 | 1.0880 1.00000 | 13.101 1.00000 | 8.1516
[Conv] /model.2/cv1/conv/Conv_output_0 1.00000 | 0.8336 1.00000 | 0.5864
[exSwish] /model.2/cv1/act/Mul_output_0 1.00000 | 0.4872 1.00000 | 0.3146 1.00000 | 3.0559 1.00000 | 1.9833
[Conv] /model.2/m/m.0/cv1/conv/Conv_output_0 1.00000 | 1.3587 1.00000 | 0.8088
[exSwish] /model.2/m/m.0/cv1/act/Mul_output_0 1.00000 | 0.8585 1.00000 | 0.3834 1.00000 | 4.1053 1.00000 | 1.7504
[Conv] /model.2/m/m.0/cv2/conv/Conv_output_0 1.00000 | 2.4838 1.00000 | 1.0474
[exSwish] /model.2/m/m.0/cv2/act/Mul_output_0 1.00000 | 2.1420 1.00000 | 0.5897 1.00000 | 6.9511 1.00000 | 2.0777
[Add] /model.2/m/m.0/Add_output_0 1.00000 | 2.2869 1.00000 | 0.7157 1.00000 | 7.7536 1.00000 | 0.0
[Conv] /model.2/cv2/conv/Conv_output_0 1.00000 | 1.5840 1.00000 | 1.0400
[exSwish] /model.2/cv2/act/Mul_output_0 1.00000 | 1.0642 1.00000 | 0.5301 1.00000 | 6.9092 1.00000 | 3.4345
[Concat] /model.2/Concat_output_0 1.00000 | 2.5224 1.00000 | 0.6349 1.00000 | 10.385 1.00000 | 0.0
[Conv] /model.2/cv3/conv/Conv_output_0 1.00000 | 3.0151 1.00000 | 1.0589
[exSwish] /model.2/cv3/act/Mul_output_0 1.00000 | 1.7501 1.00000 | 0.4406 0.99999 | 6.0663 1.00000 | 1.6735
[Conv] /model.3/conv/Conv_output_0 1.00000 | 1.1325 1.00000 | 0.3195
[exSwish] /model.3/act/Mul_output_0 1.00000 | 0.5785 1.00000 | 0.1652 0.99999 | 2.2395 1.00000 | 0.3982
[Conv] /model.4/cv1/conv/Conv_output_0 1.00000 | 0.3326 1.00000 | 0.1160
[exSwish] /model.4/cv1/act/Mul_output_0 1.00000 | 0.1531 1.00000 | 0.0573 1.00000 | 0.5484 1.00000 | 0.1548
[Conv] /model.4/m/m.0/cv1/conv/Conv_output_0 1.00000 | 0.7294 1.00000 | 0.2583
[exSwish] /model.4/m/m.0/cv1/act/Mul_output_0 1.00000 | 0.4092 1.00000 | 0.1448 1.00000 | 1.4311 1.00000 | 0.3293
[Conv] /model.4/m/m.0/cv2/conv/Conv_output_0 1.00000 | 0.5763 1.00000 | 0.1975
[exSwish] /model.4/m/m.0/cv2/act/Mul_output_0 1.00000 | 0.2526 1.00000 | 0.0820 1.00000 | 0.8956 1.00000 | 0.2005
[Add] /model.4/m/m.0/Add_output_0 1.00000 | 0.3054 1.00000 | 0.0972 1.00000 | 1.0543 1.00000 | 0.0
[Conv] /model.4/m/m.1/cv1/conv/Conv_output_0 1.00000 | 0.8030 1.00000 | 0.2504
[exSwish] /model.4/m/m.1/cv1/act/Mul_output_0 1.00000 | 0.3734 1.00000 | 0.0868 0.99999 | 1.2211 1.00000 | 0.2090
[Conv] /model.4/m/m.1/cv2/conv/Conv_output_0 1.00000 | 0.9418 1.00000 | 0.2373
[exSwish] /model.4/m/m.1/cv2/act/Mul_output_0 1.00000 | 0.6107 1.00000 | 0.1805 0.99999 | 1.9182 1.00000 | 0.4279
[Add] /model.4/m/m.1/Add_output_0 1.00000 | 0.7068 1.00000 | 0.2072 1.00000 | 2.2366 1.00000 | 0.0
[Conv] /model.4/cv2/conv/Conv_output_0 1.00000 | 0.9380 1.00000 | 0.2877
[exSwish] /model.4/cv2/act/Mul_output_0 1.00000 | 0.6277 1.00000 | 0.1712 0.99999 | 2.7682 1.00000 | 0.4022
[Concat] /model.4/Concat_output_0 1.00000 | 0.9453 1.00000 | 0.1837 0.99999 | 3.5589 1.00000 | 0.0
[Conv] /model.4/cv3/conv/Conv_output_0 1.00000 | 1.0833 1.00000 | 0.3435
[exSwish] /model.4/cv3/act/Mul_output_0 1.00000 | 0.4307 1.00000 | 0.1062 0.99999 | 1.6902 1.00000 | 0.2590
[Conv] /model.5/conv/Conv_output_0 1.00000 | 0.7330 1.00000 | 0.1894
[exSwish] /model.5/act/Mul_output_0 1.00000 | 0.3233 1.00000 | 0.0738 0.99998 | 1.6541 1.00000 | 0.2039
[Conv] /model.6/cv1/conv/Conv_output_0 1.00000 | 0.2583 1.00000 | 0.0809
[exSwish] /model.6/cv1/act/Mul_output_0 1.00000 | 0.1132 1.00000 | 0.0361 0.99999 | 0.5547 1.00000 | 0.0952
[Conv] /model.6/m/m.0/cv1/conv/Conv_output_0 1.00000 | 0.5540 1.00000 | 0.1694
[exSwish] /model.6/m/m.0/cv1/act/Mul_output_0 1.00000 | 0.2677 1.00000 | 0.0821 0.99999 | 1.2722 1.00000 | 0.1834
[Conv] /model.6/m/m.0/cv2/conv/Conv_output_0 1.00000 | 0.3577 1.00000 | 0.1134
[exSwish] /model.6/m/m.0/cv2/act/Mul_output_0 1.00000 | 0.1338 1.00000 | 0.0329 0.99998 | 0.6488 1.00000 | 0.0815
[Add] /model.6/m/m.0/Add_output_0 1.00000 | 0.1792 1.00000 | 0.0528 0.99999 | 0.8114 1.00000 | 0.0
[Conv] /model.6/m/m.1/cv1/conv/Conv_output_0 1.00000 | 0.5771 1.00000 | 0.1766
[exSwish] /model.6/m/m.1/cv1/act/Mul_output_0 1.00000 | 0.2299 1.00000 | 0.0546 0.99998 | 1.0036 1.00000 | 0.1380
[Conv] /model.6/m/m.1/cv2/conv/Conv_output_0 1.00000 | 0.4603 1.00000 | 0.1238
[exSwish] /model.6/m/m.1/cv2/act/Mul_output_0 1.00000 | 0.2284 1.00000 | 0.0528 0.99998 | 1.0113 1.00000 | 0.1356
[Add] /model.6/m/m.1/Add_output_0 1.00000 | 0.2980 1.00000 | 0.0734 0.99999 | 1.3001 1.00000 | 0.0
[Conv] /model.6/m/m.2/cv1/conv/Conv_output_0 1.00000 | 0.5523 1.00000 | 0.1619
[exSwish] /model.6/m/m.2/cv1/act/Mul_output_0 1.00000 | 0.2044 1.00000 | 0.0524 0.99999 | 0.8724 1.00000 | 0.1243
[Conv] /model.6/m/m.2/cv2/conv/Conv_output_0 1.00000 | 0.6084 1.00000 | 0.1569
[exSwish] /model.6/m/m.2/cv2/act/Mul_output_0 1.00000 | 0.3599 1.00000 | 0.0859 0.99998 | 1.5912 1.00000 | 0.2535
[Add] /model.6/m/m.2/Add_output_0 1.00000 | 0.4789 1.00000 | 0.1097 0.99998 | 2.1075 1.00000 | 0.0
[Conv] /model.6/cv2/conv/Conv_output_0 1.00000 | 0.6856 1.00000 | 0.1651
[exSwish] /model.6/cv2/act/Mul_output_0 1.00000 | 0.4227 1.00000 | 0.0984 0.99997 | 2.6074 1.00000 | 0.2393
[Concat] /model.6/Concat_output_0 1.00000 | 0.6388 1.00000 | 0.0984 0.99998 | 3.3526 1.00000 | 0.0
[Conv] /model.6/cv3/conv/Conv_output_0 1.00000 | 0.8111 1.00000 | 0.2263
[exSwish] /model.6/cv3/act/Mul_output_0 1.00000 | 0.3316 1.00000 | 0.0756 0.99997 | 1.9800 1.00000 | 0.1875
[Conv] /model.7/conv/Conv_output_0 1.00000 | 0.5560 1.00000 | 0.1404
[exSwish] /model.7/act/Mul_output_0 1.00000 | 0.2298 1.00000 | 0.0507 0.99994 | 1.8461 1.00000 | 0.1214
[Conv] /model.8/cv1/conv/Conv_output_0 1.00000 | 0.2673 1.00000 | 0.0892
[exSwish] /model.8/cv1/act/Mul_output_0 1.00000 | 0.0746 1.00000 | 0.0210 0.99997 | 0.6137 1.00000 | 0.0458
[Conv] /model.8/m/m.0/cv1/conv/Conv_output_0 1.00000 | 0.4261 1.00000 | 0.1309
[exSwish] /model.8/m/m.0/cv1/act/Mul_output_0 1.00000 | 0.2042 1.00000 | 0.0467 0.99995 | 1.5758 1.00000 | 0.1170
[Conv] /model.8/m/m.0/cv2/conv/Conv_output_0 1.00000 | 0.5180 1.00000 | 0.1237
[exSwish] /model.8/m/m.0/cv2/act/Mul_output_0 1.00000 | 0.2913 1.00000 | 0.0669 0.99994 | 2.5182 1.00000 | 0.1759
[Add] /model.8/m/m.0/Add_output_0 1.00000 | 0.3065 1.00000 | 0.0670 0.99993 | 2.6031 1.00000 | 0.0
[Conv] /model.8/cv2/conv/Conv_output_0 1.00000 | 0.4653 1.00000 | 0.1089
[exSwish] /model.8/cv2/act/Mul_output_0 1.00000 | 0.2483 1.00000 | 0.0540 0.99993 | 2.1787 1.00000 | 0.1265
[Concat] /model.8/Concat_output_0 1.00000 | 0.3945 1.00000 | 0.0571 0.99993 | 3.3945 1.00000 | 0.0
[Conv] /model.8/cv3/conv/Conv_output_0 1.00000 | 0.6968 1.00000 | 0.1682
[exSwish] /model.8/cv3/act/Mul_output_0 1.00000 | 0.3261 1.00000 | 0.0670 0.99991 | 2.8786 1.00000 | 0.1646
[Conv] /model.9/cv1/conv/Conv_output_0 1.00000 | 0.3497 1.00000 | 0.0960
[exSwish] /model.9/cv1/act/Mul_output_0 1.00000 | 0.3225 1.00000 | 0.1103 0.99997 | 2.6514 1.00000 | 0.2577
[MaxPool] /model.9/m/MaxPool_output_0 1.00000 | 0.4414 1.00000 | 0.1297 0.99999 | 3.4128 1.00000 | 0.0
[MaxPool] /model.9/m_1/MaxPool_output_0 1.00000 | 0.4784 1.00000 | 0.1560 0.99999 | 3.6140 1.00000 | 0.0
[MaxPool] /model.9/m_2/MaxPool_output_0 1.00000 | 0.4955 1.00000 | 0.1723 0.99999 | 3.6602 1.00000 | 0.0
[Concat] /model.9/Concat_output_0 1.00000 | 0.8794 1.00000 | 0.2754 0.99999 | 6.7183 1.00000 | 0.0
[Conv] /model.9/cv2/conv/Conv_output_0 1.00000 | 0.6059 1.00000 | 0.2436
[exSwish] /model.9/cv2/act/Mul_output_0 1.00000 | 0.1914 1.00000 | 0.0398 0.99992 | 1.6811 1.00000 | 0.0925
[Conv] /model.10/conv/Conv_output_0 1.00000 | 0.4219 1.00000 | 0.1043
[exSwish] /model.10/act/Mul_output_0 1.00000 | 0.1706 1.00000 | 0.0341 0.99991 | 1.4851 1.00000 | 0.0814
[Resize] /model.11/Resize_output_0 1.00000 | 0.3412 1.00000 | 0.0475 0.99991 | 2.9703 1.00000 | 0.0
[Concat] /model.12/Concat_output_0 1.00000 | 0.4758 1.00000 | 0.0701 0.99994 | 3.5698 1.00000 | 0.0
[Conv] /model.13/cv1/conv/Conv_output_0 1.00000 | 0.4899 1.00000 | 0.1309
[exSwish] /model.13/cv1/act/Mul_output_0 1.00000 | 0.1896 1.00000 | 0.0595 0.99997 | 1.3748 1.00000 | 0.1441
[Conv] /model.13/m/m.0/cv1/conv/Conv_output_0 1.00000 | 0.5131 1.00000 | 0.1695
[exSwish] /model.13/m/m.0/cv1/act/Mul_output_0 1.00000 | 0.2320 1.00000 | 0.0631 0.99996 | 1.6936 1.00000 | 0.1499
[Conv] /model.13/m/m.0/cv2/conv/Conv_output_0 1.00000 | 0.5625 1.00000 | 0.1369
[exSwish] /model.13/m/m.0/cv2/act/Mul_output_0 1.00000 | 0.2800 1.00000 | 0.0610 0.99993 | 2.2030 1.00000 | 0.1513
[Conv] /model.13/cv2/conv/Conv_output_0 1.00000 | 0.5349 1.00000 | 0.1256
[exSwish] /model.13/cv2/act/Mul_output_0 1.00000 | 0.2531 1.00000 | 0.0567 0.99994 | 1.9512 1.00000 | 0.1389
[Concat] /model.13/Concat_output_0 1.00000 | 0.3775 1.00000 | 0.0559 0.99994 | 2.9429 1.00000 | 0.0
[Conv] /model.13/cv3/conv/Conv_output_0 1.00000 | 0.8536 1.00000 | 0.1991
[exSwish] /model.13/cv3/act/Mul_output_0 1.00000 | 0.4066 1.00000 | 0.0833 0.99992 | 3.2888 1.00000 | 0.2034
[Conv] /model.14/conv/Conv_output_0 1.00000 | 0.4692 1.00000 | 0.1020
[exSwish] /model.14/act/Mul_output_0 1.00000 | 0.2757 1.00000 | 0.0680 0.99994 | 2.3128 1.00000 | 0.1682
[Resize] /model.15/Resize_output_0 1.00000 | 0.5514 1.00000 | 0.0886 0.99994 | 4.6257 1.00000 | 0.0
[Concat] /model.16/Concat_output_0 1.00000 | 0.6997 1.00000 | 0.1145 0.99996 | 4.9248 1.00000 | 0.0
[Conv] /model.17/cv1/conv/Conv_output_0 1.00000 | 0.5299 1.00000 | 0.1643
[exSwish] /model.17/cv1/act/Mul_output_0 1.00000 | 0.2125 1.00000 | 0.0811 0.99998 | 1.4724 1.00000 | 0.1977
[Conv] /model.17/m/m.0/cv1/conv/Conv_output_0 1.00000 | 0.5593 1.00000 | 0.2251
[exSwish] /model.17/m/m.0/cv1/act/Mul_output_0 1.00000 | 0.2300 1.00000 | 0.0934 0.99998 | 1.5433 1.00000 | 0.2135
[Conv] /model.17/m/m.0/cv2/conv/Conv_output_0 1.00000 | 0.8545 1.00000 | 0.2604
[exSwish] /model.17/m/m.0/cv2/act/Mul_output_0 1.00000 | 0.5550 1.00000 | 0.1444 0.99996 | 3.8579 1.00000 | 0.3635
[Conv] /model.17/cv2/conv/Conv_output_0 1.00000 | 0.5023 1.00000 | 0.1378
[exSwish] /model.17/cv2/act/Mul_output_0 1.00000 | 0.3705 1.00000 | 0.1101 0.99995 | 3.3247 1.00000 | 0.2729
[Concat] /model.17/Concat_output_0 1.00000 | 0.6673 1.00000 | 0.1171 0.99996 | 5.0929 1.00000 | 0.0
[Conv] /model.17/cv3/conv/Conv_output_0 1.00000 | 2.2759 1.00000 | 0.5792
[exSwish] /model.17/cv3/act/Mul_output_0 1.00000 | 1.4949 1.00000 | 0.3410 0.99994 | 11.339 1.00000 | 1.0927
[Conv] /model.18/conv/Conv_output_0 1.00000 | 0.7697 1.00000 | 0.1473
[exSwish] /model.18/act/Mul_output_0 1.00000 | 0.4751 1.00000 | 0.0723 0.99986 | 3.8054 1.00000 | 0.1805
[Concat] /model.19/Concat_output_0 1.00000 | 0.5493 1.00000 | 0.0650 0.99990 | 4.4532 1.00000 | 0.0
[Conv] /model.20/cv1/conv/Conv_output_0 1.00000 | 0.7051 1.00000 | 0.1141
[exSwish] /model.20/cv1/act/Mul_output_0 1.00000 | 0.3838 1.00000 | 0.0557 0.99984 | 3.1648 1.00000 | 0.1376
[Conv] /model.20/m/m.0/cv1/conv/Conv_output_0 1.00000 | 0.6564 1.00000 | 0.1113
[exSwish] /model.20/m/m.0/cv1/act/Mul_output_0 1.00000 | 0.3043 1.00000 | 0.0461 0.99986 | 2.5376 1.00000 | 0.1101
[Conv] /model.20/m/m.0/cv2/conv/Conv_output_0 1.00000 | 0.9716 1.00000 | 0.1299
[exSwish] /model.20/m/m.0/cv2/act/Mul_output_0 1.00000 | 0.6093 1.00000 | 0.0851 0.99981 | 5.2704 1.00000 | 0.2502
[Conv] /model.20/cv2/conv/Conv_output_0 1.00000 | 0.5102 1.00000 | 0.0828
[exSwish] /model.20/cv2/act/Mul_output_0 1.00000 | 0.3326 1.00000 | 0.0576 0.99985 | 3.2475 1.00000 | 0.1576
[Concat] /model.20/Concat_output_0 1.00000 | 0.6942 1.00000 | 0.0676 0.99982 | 6.1906 1.00000 | 0.0
[Conv] /model.20/cv3/conv/Conv_output_0 1.00000 | 1.5858 1.00000 | 0.2464
[exSwish] /model.20/cv3/act/Mul_output_0 1.00000 | 0.9009 1.00000 | 0.1510 0.99986 | 8.1236 1.00000 | 0.5544
[Conv] /model.21/conv/Conv_output_0 1.00000 | 0.6364 1.00000 | 0.0815
[exSwish] /model.21/act/Mul_output_0 1.00000 | 0.3965 1.00000 | 0.0494 0.99974 | 3.5894 1.00000 | 0.1256
[Concat] /model.22/Concat_output_0 1.00000 | 0.4317 1.00000 | 0.0406 0.99980 | 3.8845 1.00000 | 0.0
[Conv] /model.23/cv1/conv/Conv_output_0 1.00000 | 0.6592 1.00000 | 0.0848
[exSwish] /model.23/cv1/act/Mul_output_0 1.00000 | 0.3573 1.00000 | 0.0428 0.99974 | 3.1069 1.00000 | 0.1259
[Conv] /model.23/m/m.0/cv1/conv/Conv_output_0 1.00000 | 0.7364 1.00000 | 0.0879
[exSwish] /model.23/m/m.0/cv1/act/Mul_output_0 0.99999 | 0.4095 1.00000 | 0.0404 0.99964 | 3.4908 1.00000 | 0.1387
[Conv] /model.23/m/m.0/cv2/conv/Conv_output_0 1.00000 | 0.8693 1.00000 | 0.0855
[exSwish] /model.23/m/m.0/cv2/act/Mul_output_0 1.00000 | 0.5731 1.00000 | 0.0634 0.99970 | 4.9156 1.00000 | 0.1682
[Conv] /model.23/cv2/conv/Conv_output_0 1.00000 | 0.5111 1.00000 | 0.0647
[exSwish] /model.23/cv2/act/Mul_output_0 1.00000 | 0.3930 1.00000 | 0.0593 0.99980 | 3.8416 1.00000 | 0.1424
[Concat] /model.23/Concat_output_0 1.00000 | 0.6949 1.00000 | 0.0571 0.99975 | 6.2387 1.00000 | 0.0
[Conv] /model.23/cv3/conv/Conv_output_0 1.00000 | 0.9849 1.00000 | 0.1291
[exSwish] /model.23/cv3/act/Mul_output_0 1.00000 | 0.4929 1.00000 | 0.0801 0.99983 | 4.6763 1.00000 | 0.3220
[Conv] /model.24/m.2/Conv_output_0 1.00000 | 0.2135 1.00000 | 0.1117
[Sigmoid] output1 1.00000 | 0.0289 1.00000 | 0.0074 0.99999 | 0.2542 1.00000 | 0.0921
[Conv] /model.24/m.1/Conv_output_0 1.00000 | 0.4433 1.00000 | 0.2245
[Sigmoid] output2 1.00000 | 0.0470 1.00000 | 0.0160 0.99999 | 0.4188 1.00000 | 0.1851
[Conv] /model.24/m.0/Conv_output_0 1.00000 | 0.9307 1.00000 | 0.5451
[Sigmoid] output3 1.00000 | 0.0790 1.00000 | 0.0359 1.00000 | 0.6391 1.00000 | 0.3722
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
done
--> Init runtime environment
adb: unable to connect for root: closed
I target set by user is: rk3588
I Get hardware info: target_platform = rk3588, os = Linux, aarch = aarch64
I Check RK3588 board npu runtime version
I Starting ntp or adb, target is RK3588
I Start adb...
I Connect to Device success!
I NPUTransfer: Starting NPU Transfer Client, Transfer version 2.1.0 (b5861e7@2020-11-23T11:50:36)
D RKNNAPI: ==============================================
D RKNNAPI: RKNN VERSION:
D RKNNAPI: API: 2.0.0b0 (18eacd0 build@2024-03-22T06:07:59)
D RKNNAPI: DRV: rknn_server: 2.0.0b0 (18eacd0 build@2024-03-22T14:07:19)
D RKNNAPI: DRV: rknnrt: 2.0.0b0 (35a6907d79@2024-03-24T10:31:14)
D RKNNAPI: ==============================================
D RKNNAPI: Input tensors:
D RKNNAPI: index=0, name=images, n_dims=4, dims=[1, 640, 640, 3], n_elems=1228800, size=2457600, w_stride = 0, size_with_stride = 0, fmt=NHWC, type=FP16, qnt_type=NONE, zp=0, scale=1.000000
D RKNNAPI: Output tensors:
D RKNNAPI: index=0, name=output1, n_dims=4, dims=[1, 18, 20, 20], n_elems=7200, size=14400, w_stride = 0, size_with_stride = 0, fmt=NCHW, type=FP16, qnt_type=NONE, zp=0, scale=1.000000
D RKNNAPI: index=1, name=output2, n_dims=4, dims=[1, 18, 40, 40], n_elems=28800, size=57600, w_stride = 0, size_with_stride = 0, fmt=NCHW, type=FP16, qnt_type=NONE, zp=0, scale=1.000000
D RKNNAPI: index=2, name=output3, n_dims=4, dims=[1, 18, 80, 80], n_elems=115200, size=230400, w_stride = 0, size_with_stride = 0, fmt=NCHW, type=FP16, qnt_type=NONE, zp=0, scale=1.000000
done
我的yolov5s模型是在自制训练集上训练得到的，在pytorch的fp32推理精度是81.8，fp16的推理精度是77.5，我翻阅了官方文档，没有找到解释这个现象的原因。所以fp16相较于fp32在rk3588上通过adb连板推理掉了4个点这个现象正常吗？
