• State
  • Locked Locked
  • Replies 2 replies
  • Subscribers 98 subscribers
  • Views 343 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

TDA4VM: PSDK TIDL8.1: Demo model pelee floating-point mode numParamBits=32 infference don't detect the target

Cherry Zhou
Mastermind 22235 points
Part Number: TDA4VM


Hi team,

Here's an issue from the customer may need your help:

Issue: The correct detection results are obtained from the import and inference of the fixed-point mode numParamBits=8 based on the TIDL demo model pelee && host emulation, configured as follows. However, when changing into the floating-point mode numParamBits=32 to import and infection, no target was detected and the result was 0. The configuration is as follows 1 and 2, and the floating-point mode import has a warning 3:

1. import config file 

modelType          = 0
numParamBits = 8
inputNetFile       = "../../test/testvecs/models/public/caffe/peele/pelee_voc/deploy.prototxt"
inputParamsFile    = "../../test/testvecs/models/public/caffe/peele/pelee_voc/pelee_304x304_acc7094.caffemodel"
outputNetFile      = "../../test/testvecs/config/tidl_models/caffe/tidl_net_peele_300.bin"
outputParamsFile   = "../../test/testvecs/config/tidl_models/caffe/tidl_io_peele_300_"
inDataNorm  = 1
inMean = 104 117 123
inScale = 0.0170000009239 0.0170000009239 0.0170000009239
inDataFormat = 0
inData  =   "../../test/testvecs/config/detection_list.txt"
postProcType = 2
perfSimConfig = ../../test/testvecs/config/import/device_config.cfg
inWidth  = 1024
inHeight = 512
inNumChannels = 3
foldMaxPoolInConv2D = 0
foldPreBnConv2D = 0

2. inference config file

inFileFormat    = 2
postProcType = 2
#numFrames   = 272
numFrames   = 1
netBinFile      = "testvecs/config/tidl_models/caffe/tidl_net_peele_300.bin"
ioConfigFile    = "testvecs/config/tidl_models/caffe/tidl_io_peele_300_1.bin"
outData =   "testvecs/output/pelee.bin"
inData  =   "testvecs/config/detection_list.txt"
debugTraceLevel = 2
writeTraceLevel = 1
reservedCtrl = 2

3.  warning in float mode import

WARNING: [TIDL_E_DATAFLOW_INFO_NULL] ti_cnnperfsim.out fails to allocate memory in MSMC. Please look into perfsim log. This model can only be used on PC emulation, it will get fault on target.

Could you help check this case? Thanks.

Best Regards,

Cherry

  • 0
    TI__Mastermind 22235 points

    Hi,

    The customer has found some clues from source code.

    sysMems[TIDL_SYSMEM_L2_SCRATCH] is limited to a small value, that results in algDetLyrParams->topMAllClasses is -16, so the detection is actually skipped and nothing is detected.
    But they didn't find why algHandle->intAlgHandle->sysMems is limited to a small value.
    Could you help check it?
    Thanks and regards,
    Cherry
  • 0 in reply to Cherry Zhou
    TI__Guru 56820 points

    Hi Cherry,

        I noticed the version mentioned in the title itself. We fixed one bug which seems related to this issue in SDK 8.2. Can you try this network with the same?

    Regards,

    ANshu