search

tenstorrent / tt-metal

:metal: TT-NN operator library, and TT-Metalium low level kernel programming model.
Apache License 2.0
303 stars 26 forks source link

TTNN Unet inference pipeline #8455

Open saichandax opened 1 month ago

saichandax commented 1 month ago

setting up inference pipeline with available PCC Currently - PCC (~0.0)

Expected behaviour - To get PCC > 0.99

To reproduce: go to harini/unet_inference execute demo.py file.

dvartaniansTT commented 1 month ago

@saichandax and @HariniMohan0102 I am a little confused on this one. is this the same model as in #8454? do you mean when you provide a real image input, the pcc goes from 0.8 to 0?

HariniMohan0102 commented 1 month ago

@dvartaniansTT We are using the same model as in #8454 to test inference with a real image input, the pcc drops from 0.8 to ~0.0

Comparing the output tensors of pytorch and reference implementations by passing the same real input image, I get both tensors are equal. But the ttnn predictions are not good with the real input image.

nsmithtt commented 1 month ago

@dvartaniansTT / @saichandax, can you try using the metal PCC visualizer to debug?