[VTA] YoloV3 Support #4887
[VTA] YoloV3 Support #4887
Conversation
|
After this patch, I can run Yolov3 tiny on VTA(simulator + FPGA), but it still have some accuracy issue, It may caused by test application parameter usage or VTA existing logic, I am not very sure, at same time please kindly advise if i missing any logic for Yolov3-tiny support in this patch. |
|
FYI, |
|
small frame accuracy issue fixed. |
|
Hi @huajsj, thanks for adding YoloV3 support for VTA. Do you think that along with this PR you can also construct a tutorial similar to this one: https://github.com/apache/incubator-tvm/blob/master/vta/tutorials/frontend/deploy_vision_on_vta.py We could have one tutorial called deploy_classification_on_vta.py, and deploy_detection_on_vta.py to differentiate between object classification vs. detection. In addition, this will exercise compilation support for Yolo to make sure that support does not break. |
|
Hi @tmoreau89 , Thanks for the follow up , sure, would work for a tutorial and upload soon. Regards
|
|
Since the python script deploy_vision_on_vta.py exists in the vta directory already, how about renaming those to deploy_classification.py and deploy_detection.py? |
|
@liangfu that is correct; we'd need to rename |
|
Sorry, I meant removing the '_on_vta' suffix in the file names. |
|
Oh my bad, I read your comment too quickly; I like your suggestion and agree; shortening the filename makes sense! |
|
just add the tutorial which work good in simulator, but not work properly in FPGA, still in the process of trouble shooting. |
|
I'm glad to hear that. For now, I suggest keep an eye on free CMA memory space on the FPGA device. |
|
@huajsj which platform are you using for the FPGA demo? |
Hi @tmoreau89, I am using pynq board for testing. |
Issue: YoloV3 use some operator and logic that not get good support by existing vta logic, like nn.pad, upsample, and 255 output channel. Solution: add related logic to let darknet YoloV3 can running on VTA
|
after apply AutoTVM configure file, now we can run Yolov3-tiny tutorial correctly on PYNQ board, the autotvm configure file in this pull request tlc-pack/tophub#17 |
|
Hi @liangfu, nice to know the tutorial worked in your environment, and thanks for the kindly follow up , all review comments get addressed, please check. |
|
Would you please also rename the files as requested above, see #4887 (comment) |
sure, already changed. |
|
@tmoreau89 , if you have time , could you help to have reivew? thanks. |
|
@tmoreau89 , Thanks for the kindly follow up, 'tlc-pack/tophub#17' is the the default Autotvm configure file for Yolov3 tiny, could you help to review this patch too? thanks. |
|
Thanks @huajsj , I merged tlc-pack/tophub#17 |
|
Thank you @huajsj @cbalint13 @liangfu , the code has been merged! |
* [VTA] YoloV3 Support Issue: YoloV3 use some operator and logic that not get good support by existing vta logic, like nn.pad, upsample, and 255 output channel. Solution: add related logic to let darknet YoloV3 can running on VTA * Fix small(0, or 1 heigh/width) detect frame issue. * add yolov3-tiny turtorial * add os import * address review comments. * rename tutorial file with a short name. * rename deploy_vision_on_vta.py into deploy_classification.py. * address review comment, fix plint eror in deploy_detection.py
* [VTA] YoloV3 Support Issue: YoloV3 use some operator and logic that not get good support by existing vta logic, like nn.pad, upsample, and 255 output channel. Solution: add related logic to let darknet YoloV3 can running on VTA * Fix small(0, or 1 heigh/width) detect frame issue. * add yolov3-tiny turtorial * add os import * address review comments. * rename tutorial file with a short name. * rename deploy_vision_on_vta.py into deploy_classification.py. * address review comment, fix plint eror in deploy_detection.py
* [VTA] YoloV3 Support Issue: YoloV3 use some operator and logic that not get good support by existing vta logic, like nn.pad, upsample, and 255 output channel. Solution: add related logic to let darknet YoloV3 can running on VTA * Fix small(0, or 1 heigh/width) detect frame issue. * add yolov3-tiny turtorial * add os import * address review comments. * rename tutorial file with a short name. * rename deploy_vision_on_vta.py into deploy_classification.py. * address review comment, fix plint eror in deploy_detection.py
* [VTA] YoloV3 Support Issue: YoloV3 use some operator and logic that not get good support by existing vta logic, like nn.pad, upsample, and 255 output channel. Solution: add related logic to let darknet YoloV3 can running on VTA * Fix small(0, or 1 heigh/width) detect frame issue. * add yolov3-tiny turtorial * add os import * address review comments. * rename tutorial file with a short name. * rename deploy_vision_on_vta.py into deploy_classification.py. * address review comment, fix plint eror in deploy_detection.py
* [VTA] YoloV3 Support Issue: YoloV3 use some operator and logic that not get good support by existing vta logic, like nn.pad, upsample, and 255 output channel. Solution: add related logic to let darknet YoloV3 can running on VTA * Fix small(0, or 1 heigh/width) detect frame issue. * add yolov3-tiny turtorial * add os import * address review comments. * rename tutorial file with a short name. * rename deploy_vision_on_vta.py into deploy_classification.py. * address review comment, fix plint eror in deploy_detection.py
Issue:
YoloV3 use some operator and logic that not get good support by
existing vta logic, like nn.pad, upsample, and 255 output channel.
Solution:
add related logic to let darknet YoloV3 can running on VTA
@tmoreau89 ,could you help to do a review? thanks.