Spaces:
Sleeping
Sleeping
| Ticket Name: TDA2: run the user cnn model with TIDL on tda2x | |
| Query Text: | |
| Part Number: TDA2 hi, dear Now although I bought serveral tda2x boards, I have only got part service from the TI tech support in Shanghai of China. I could not get the source code of TIDL. So if I want to run my improved cnn model , I can only run it on the tda2x board not on PC simulator. I have imported the cnn model to the bin format which is suitable for TIDL. Yet the input data need to be resize to 368 X 368，and I should resize it by myself( It has to be the TI method to resize). Even worse, the configuration file (only seg , ssd detection included)on the tda2x board looks like that below: TIDL Configuration parameters ----------------------------- ################################################################################ IMP: Make sure the size of the file names (excluding extension) is not more than 8 characters. ################################################################################ inputWidth=1024 inputHeight=512 inputFile=IN.RGB outputFile=OUT.BIN netFileName=NET.BIN paramFileName=PRM.BIN My cnn net is not for detection or classifcation, but pose recognition. I do not know if I can use this demo cfg file to run my cnn model successfully on the tda2x. If not , could you please be kind enough to offer me new cfg file or make some change for me to run? I mean could you offer me a new in_Data_OD and in_header_OD sized 368 x 368 for my pose recognitation? I will replace the bin to run the video demo. tks, looking forward to your reply. | |
| Responses: | |
| Hi, Did you check below thread where we explained how to run TIDL OD use cases? e2e.ti.com/.../689617 Thanks, Praveen | |
| MY TIDL OD use cases ran sucessfully. Yet the new usr model not validated which is for pose recognition require the input image size 368 X368， we couldnot get the input data in the vision sdk with this size. And could you think of a method to change the OD demo cfg file to run my pose model recognition demo? | |
| Please refer below thread for generating indata and inHeader e2e.ti.com/.../687481 Thanks, Praveen | |
| In fact, in my case only, if I use this case called "chanis_tidl – File IO based use case to validate user model ", it is better to bring up? If so, could you offer me some tips for changing this demo to my own pose mode recog demo?This case is similar to my pose reg demo, and I want to test the performance of the mode only, not for accuracy which maybe need only one image like in.RGB, right? And how to translate my jpg image to the in.RGB format? Good luck, tks! | |
| Beside these facts above, the output of my pose model is not the limited classification but arbitrarily float32 value. So from the link that you offer me before, I do not know if it will work successfully by just commenting the classfications.(this Line : tempoutPutList->objDesc[tempoutPutList->numObjects].objType = (label - 1); need to be changed to " (label - 1)%3 ") Any other change should be made? | |
| Hi, We answered this in below thread, e2e.ti.com/.../723674 No need to create multiple threads for the same topic, we can discuss and follow up in the same thread. Thanks, Praveen | |
| What about the semseg usecase of TIDL？ How to generate the indata and inHeader for semseg case？ | |
| Hi, Please refer to below thread for generating indata and inHeader for TIDL use case, e2e.ti.com/.../660676 Thanks, Praveen | |