Add support for MobileNet v2 in Ethos-U55 and arm baremetal runner#3056
Add support for MobileNet v2 in Ethos-U55 and arm baremetal runner#3056robell wants to merge 11 commits intopytorch:mainfrom
Conversation
robell
commented
Apr 16, 2024
- Builds on TOSA support for mv2
- Currently has accuracy issues for on-target execution.
- Adding support mv2 from examples/models (torchvision mv2)
- Adds ability to use portable Q/DQ ops to enable float signature quantized models (to be improved with int signature when possible to enable integer images rather than int -> float -> quant flow.
- Adds permute for ethos-u backend to enable NCHW to NHWC (which we'll aim to replace with a compiler pass on entry/exit of the subgraph, or the NHWC support when available.
- small fix to enable qantized ops in runtime for the IOQDQ pass, and build/register those kernels/ops in bare metal build
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/3056
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 6f4ce19 with merge base cc12d9b ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
digantdesai
left a comment
There was a problem hiding this comment.
This is great. Thanks @robell.
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
2 similar comments
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
f346e0f to
c2442c4
Compare
|
Sigh.. can you rebase please? |
5996cbf to
0790d93
Compare
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
1 similar comment
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Signed-off-by: Rob Elliott <robert.elliott@arm.com>
Signed-off-by: Rob Elliott <robert.elliott@arm.com>
Currently only enabled in the aot_ compile flow as this requires AoT kernel registration. Signed-off-by: Rob Elliott <robert.elliott@arm.com> Change-Id: I0c41b3a34613c900fc3f350f4f83b41378ac0add
… issues until further fixes * enable further input types and layouts * add optional temp support for reordering NCHW * include portable quantized ops for IOQDQ fallback Optional reordering will be removed once NHWC support is introduced. Signed-off-by: Rob Elliott <robert.elliott@arm.com> Change-Id: Iba2d4954fa14bc937584e59cb2ce82f4cdc1205d
The templated function with this change has an ambigious overload, making it unambiguous by casting up to the templated type in this quantized function
Signed-off-by: Rob Elliott <robert.elliott@arm.com>
Signed-off-by: Rob Elliott <robert.elliott@arm.com>
Signed-off-by: Rob Elliott <robert.elliott@arm.com>
Signed-off-by: Rob Elliott <robert.elliott@arm.com>
Signed-off-by: Rob Elliott <robert.elliott@arm.com>
|
@digantdesai or @mergennachin could we import this now, thanks to the fix in #3449 this now passes when rebased. |
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@digantdesai merged this pull request in 623f0e0. |