search

tenstorrent / tt-metal

:metal: TT-NN operator library, and TT-Metalium low level kernel programming model.
Apache License 2.0
475 stars 75 forks source link

ttnn conv output shape mismatch #11050

Open Sudharsan-V opened 3 months ago

Sudharsan-V commented 3 months ago

Describe the bug While working on the Blazepose model, we observed that the ttnn conv2d operation fails to produce the expected output shape for four different input configurations. Instead, the channel size of the output tensor produced by the conv2d operation is rounded to the nearest multiple of 32.

ttnn output shape: ([1, 32, 1, 64])   Expected shap: ([1, 2, 8, 8])
ttnn output shape: ([1, 32, 1, 64])  Expected shape: ([1, 6, 8, 8])
ttnn output shape: ([1, 32, 1, 256])  Expected shape: ([1, 24, 16, 16])
ttnn output shape: ([1, 96, 1, 64])   Expected shape: ([1, 72, 8, 8])

To Reproduce Steps to reproduce the behavior:

  1. pick the test from the commit
  2. Run pytest<path to file>
Sudharsan-V commented 3 months ago

cc: @boris-drazic