13811 - This issue was faced while working on reducing unwnated TM operations in existing pipeline.
Below are the path to run test of vanilla UNet module:
Unet - pytest tests/ttnn/integration_tests/vanilla_unet/test_ttnn_unet.py
Currently,
Torch convTranspose is used.
One torch convolution.
One torch maxpool
one torch sigmoid - (In the model, input to sigmoid has pcc of 0.94. This causes sigmoid output to have pcc ~ 0.0)
Reduced the TM ops such as to_device, ttnn.reshape and from_device to bypass OOM issue and statical buffer issue in the pipeline. However, the pipeline is blocked by #13811 currently.
punithsekar
commented
2 months ago
Model card - https://github.com/tenstorrent/tt-metal/issues/13272
The ttnn implementation of vanilla UNet is in branch punith/vanilla_unet_inference PR #14338 The test file are in path.
Current PCC is 0.93.
Pending issues related to vanilla UNet model:
6326
13324
13336
14882
13811 - This issue was faced while working on reducing unwnated TM operations in existing pipeline.
Below are the path to run test of vanilla UNet module: Unet -
pytest tests/ttnn/integration_tests/vanilla_unet/test_ttnn_unet.pyCurrently, Torch convTranspose is used. One torch convolution. One torch maxpool one torch sigmoid - (In the model, input to sigmoid has pcc of 0.94. This causes sigmoid output to have pcc ~ 0.0)
Reduced the TM ops such as to_device, ttnn.reshape and from_device to bypass OOM issue and statical buffer issue in the pipeline. However, the pipeline is blocked by #13811 currently.