[onert] Introduce fine tuning feature for training #12951
Draft
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
7 tasks
mbencer
changed the title
mbencer
commented
May 7, 2024
| @@ -1695,7 +1742,7 @@ NNFW_STATUS nnfw_session::train_export_circle(const char *path) | |||
|
|
|||
| auto subg = subgs->Get(0); // Get 1st subgraph | |||
| if (!idx.valid() || idx.value() >= subg->tensors()->size()) | |||
| throw std::runtime_error("Trainable tensor index is out of range"); | |||
| return; | |||
There was a problem hiding this comment.
It's just a workaround for a problem reproduced even on master to enable exporting trained circle to a file (used in my test scripts to compare weights before and after training).
My assumption is that the trained graph can have more tensors because of the tensors consumed by a Loos function, but not sure how it worked before.
BTW:
auto model = ::circle::GetModel(mmapfile.buf());
if (!model)
throw std::runtime_error("Failed to get model from circle");
auto subgs = model->subgraphs();
if (!subgs || subgs->size() != 1)
throw std::runtime_error("Circle does not has valid subgraph or has multiple subgraphs");
auto subg = subgs->Get(0); // Get 1st subgraph
can be move outside iterateTrainableTensors but I'll try to deal with it in a separate PR.
jyoungyun
reviewed
May 8, 2024
| } | ||
|
|
||
| const auto ir_op_index = onert::ir::OperationIndex{op_index}; | ||
| auto &options = _coptions[0]; |
There was a problem hiding this comment.
How about using TrainingInfo instead of _coptions?
There was a problem hiding this comment.
@jyoungyun I like this idea, introduced ;)
This commits extend TrainableOperation API with methods to disable/enable weights update of some particular trainable operations. In trainable operations is stored also status about contribution in backward propagation phase. ONE-DCO-1.0-Signed-off-by: Mateusz Bencer <m.bencer@partner.samsung.com>
Co-authored-by: Jiyoung Giuliana Yun <wldudyun10@gmail.com>
ONE-DCO-1.0-Signed-off-by: Mateusz Bencer <m.bencer@partner.samsung.com>
mbencer
force-pushed
the
mbencer/FineTuning
branch
from
a0fcf8d
to
6ac060a
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR adds fine tuning feature to onert. The new methods to nnfw API were added. onert_train tool was also extended to support this feature via
--frozen_ops_idxargument.TEST ENVIRONMENT:
epoch/batch size/frozen opsconfig to measure time and accuracyThe test methodology was described here and the results are collected in tests report
Issue: #12386