tenstorrent / tt-metal

:metal: TT-NN operator library, and TT-Metalium low level kernel programming model.
Apache License 2.0
396 stars 48 forks source link

MobilenetV2 fails with drop in pcc. #11189

Closed keerthana-r-mcw closed 3 days ago

keerthana-r-mcw commented 1 month ago

Describe the bug MobilenetV2 ttnn implementation with old conv api test fails with pcc of 0.67.

To Reproduce Steps to reproduce the behavior:

  1. Checkout keerthanar/mobilenetv2_ttnn.
  2. Use the command to run the test pytest tests/ttnn/integration_tests/mobilenetv2/test_tttnn_mobilenetv2.py.
  3. The test will fails with the pcc of 0.67.
sankarmanoj-tt commented 1 month ago

I think this file mobilenet_v2-b0353104.pth is needed to run this test case. @keerthana-r-mcw Can you share it?

sankarmanoj-tt commented 1 month ago

Why is the old Conv API used? I think the test case consists of the entire network. Do you have test cases for smaller functional blocks of the network?

mywoodstock commented 2 weeks ago

AI: upgrade mobilenetv2 to use latest conv api. @keerthana-r-mcw, assigning back to you. Once the upgrade is done, please retest and reassign if still failing. Thanks. cc: @dvartaniansTT

keerthana-r-mcw commented 1 week ago

https://github.com/tenstorrent/tt-metal/issues/11197 : Port MobilenetV2 model to New Conv API on WH n150

mywoodstock commented 3 days ago

@keerthana-r-mcw Let us know once the port to new conv API is done, then we can revisit. Perhaps we can close this issue in favor of #11197 ?

saichandax commented 3 days ago

@mywoodstock , We are tracking the status of MobilenetV2 in #11197, where we achieve a PCC of 0.94 for real weights. We have two issues #12140 and #12141 for 9 failing Convs. Please note.

We can close this issue for now. Thanks