Optimizations and WAs to support HPU execution for Detr-Resnet-50 #1334
Conversation
|
This PR builds on #1155 , which is meant for Eager mode and adds changes necessary for Lazy mode execution. Some of the review feedback for the former PR is addressed here. More details below. [x] pls. rebase/sync on top of main OH. |
|
|
|
Test Result: |
sandeep-maddipatla
force-pushed
the
detr-hpu
branch
from
a31ae42
to
0237edd
Compare
- Add capability to ignore targets that have an out-of-range ID - This helps to pad target objects to avoid graph recompilation and yet not affect the loss computation in training.
| """ | ||
| Copied from https://github.com/huggingface/transformers/tree/v4.40.2 | ||
| https://github.com/huggingface/transformers/blob/4fdf58afb72b0754da30037fc800b6044e7d9c99/src/transformers/models/detr/modeling_detr.py#L2287 | ||
| The modications are: |
There was a problem hiding this comment.
| The modications are: | |
| The modifications are: |
|
|
||
| # Compute the classification cost. Contrary to the loss, we don't use the NLL, | ||
| # but approximate it in 1 - proba[target class]. | ||
| # The 1 is a constant that doesn't change the matching, it can be ommitted. |
There was a problem hiding this comment.
| # The 1 is a constant that doesn't change the matching, it can be ommitted. | |
| # The 1 is a constant that doesn't change the matching, it can be omitted. |
There was a problem hiding this comment.
@splotnikv , please take a look if you're covering for Sandeep
Done. I don't have access right to update this PR, so created a new one. See #1404. |
Modifications to the Detr transformer including WA's and Optimizations to run the Detr-Resnet-50 model in eager and lazy modes on the HPU.
Fixes # (issue)
Before submitting