Part Number:TDA4VM
Hi,TIer,
在部署 pretrained model yolov5m6_640_ti_lite_44p1_62p9.onnx(link 如下(1)) 到 TDA4 EVM (8bits 量化)时,基于coco dataset val2017做了精度测试,conf 0.001,nms 0.65,EVM上AP[0.5:0.95]结果0.284,与TI发布页(link 如下(2))中浮点模型的精度0.441相差很大,TIDL 尝试了16bits 量化,EVM上结果为0.311,精度损失依然很大。
(1) https://github.com/TexasInstruments/edgeai-yolov5/blob/master/pretrained_models/models/detection/coco/edgeai-yolov5/yolov5m6_640_ti_lite_44p1_62p9.onnx.link
(2) https://github.com/TexasInstruments/edgeai-yolov5
请问:
1. TI 基于此pretrained yolov5m model是否有做过TDA4 EVM上的精度测试,结果如何?
2. 麻烦帮忙check 下import的配置,是否有待改进的参数?或者针对这个模型,TI的推荐配置?
(1) 测试使用 TI_DEVICE_a72_test_dl_algo_host_rt.out,处理val2017
(2) import configuration file
modelType = 2
numParamBits = 8
numFeatureBits = 8
#quantizationStyle = 3
quantizationStyle = 2
inputNetFile = "../../test/testvecs/models/public/onnx/yolov5m6_640_ti_lite_44p1_62p9.onnx"
outputNetFile = "../../test/testvecs/config/tidl_models/onnx/tidl_net_yolov5m6_640_ti_lite_44p1_62p9.bin"
outputParamsFile = "../../test/testvecs/config/tidl_models/onnx/tidl_io_yolov5m6_640_ti_lite_44p1_62p9_"
inDataNorm = 1
inMean = 0 0 0
inScale = 0.003921568627 0.003921568627 0.003921568627
inDataFormat = 1
inWidth = 640
inHeight = 640
inNumChannels = 3
numFrames = 1
inData = "../../test/testvecs/config/detection_list.txt"
perfSimConfig = ../../test/testvecs/config/import/device_config.cfg
inElementType = 0
metaArchType = 4
metaLayersNamesList = "../../test/testvecs/config/import/public/onnx/yolov5m6_640_ti_lite_metaarch.prototxt"
postProcType = 2
foldPreBnConv2D = 0
#calibrationOption = 7
(3) prototxt file:相对于发布配置,只修改了 confidence_threshold: 0.001
(4) infer configuration file
inFileFormat = 2
numFrames = 5000
netBinFile = "testvecs/config/tidl_models/onnx/tidl_net_yolov5m6_640_ti_lite_44p1_62p9.bin"
ioConfigFile = "testvecs/config/tidl_models/onnx/tidl_io_yolov5m6_640_ti_lite_44p1_62p9_1.bin"
inData = testvecs/config/name_list_acc.txt
outData = testvecs/output/tidl_yolov5m6_640_ti_lite_44p1_62p9_od.bin
inResizeMode = 0
debugTraceLevel = 0
writeTraceLevel = 0
postProcType = 2
Thanks
Regards
Cherry Zhou:
您好,我们已收到您的问题并升级到英文论坛寻求帮助,链接如下,如有答复将尽快回复您。
e2e.ti.com/…/tda4vm-yolov5m-loss-of-accuracy-issue-on-evm
,
Moore Ren:
Cherry,您好!
我看论坛仍然没有回复,麻烦再更新问一下TI是否有相关的精度测试结果及配置参考,谢谢!
,
Cherry Zhou:
十分抱歉,我们这边发邮件催促下工程师,给您带来的不便再次抱歉!
,
Cherry Zhou:
您好,
确保 inference pipeline与training repository的流程完全匹配并不困难。 因此建议您使用此存储库来进行benchmarking。
以下是包含此检测模型设置的配置文件:
https://www.ti2k.com/wp-content/uploads/ti2k/DeyiSupport_DSP_detection.py