[pt] rb pruning #3688
[pt] rb pruning #3688
Conversation
AlexanderDokuchaev
force-pushed
the
ad/rb_prune
branch
from
8ba934d to
3ead4c9
Compare
github-actions
bot
added
the
documentation
| ```bash | ||
| python main.py | ||
| # Or to run Regularization-Based pruning | ||
| python main.py --mode rb |
There was a problem hiding this comment.
the documentation about sparsity is not connected with this sample and contains old configs.
are you going to update in the next PRs?
|
|
||
|
|
||
| @COMPRESSION_MODULES.register() | ||
| class RBPruningMask(nn.Module, StatefulModuleInterface): |
There was a problem hiding this comment.
is it possible to save and load pruned model? Is there test for such functionality?
| assert r == 0.5061728358268738 | ||
| scheduler.step() | ||
| r = float(1 - mask.sum() / 81) | ||
| assert r == 0.1111111044883728 | ||
| scheduler.step() | ||
| r = float(1 - mask.sum() / 81) | ||
| assert r == 0.48148149251937866 |
There was a problem hiding this comment.
shouldn't the ratio monotonously grow?
There was a problem hiding this comment.
No, it's exponential scheduler
|
Overall looks good and functional. I've run the sample and checked actual sparsity in IR .bin file. It's around 68% |
|
sample has some unused imports. formatting of samples is not part of make-precommit routine, is it? |
no, it's works |
2 participants
Changes
Add rb pruning algorithm
Rename prune folder to pruning
Add parameters to select pruning mode in example
Related tickets
173791