Compression for Tensorflow #2755
Conversation
| Common base class for all compressors. | ||
|
|
||
| This class is designed for other base classes. | ||
| Algorithms should inherit ``Pruner`` or Quantizer instead. |
There was a problem hiding this comment.
Quantizer -> Quantizer
| if ret is None or ret.get('exclude'): | ||
| return None | ||
| return ret | ||
| The compressed model can only execute in eager mode. |
There was a problem hiding this comment.
is this statement still true?
There was a problem hiding this comment.
Yes, we apply masks by modifying weight variables, which can only be done in eager mode.
There was a problem hiding this comment.
so the performance will be affected if we apply masks on the model? how much performance degradation it would induce?
There was a problem hiding this comment.
better to check the performance degradation using simple test
|
|
||
| If you want to prune the model during training, use callback API instead. |
There was a problem hiding this comment.
which one is callback API?
| The key is weight ``Variable``'s name. The value is a mask ``Tensor`` of weight's shape and dtype. | ||
| If a weight's key does not appear in the return value, that weight will not be pruned. | ||
| Returning ``None`` means the mask is not changed since last time. | ||
| Weight names are globally unique, e.g. `model/conv_1/kernel:0`. |
|
looks good. please add test for it and fix pipeline error |
| }] | ||
| pruner = FPGMPruner(model, config_list) | ||
| pruner.compress() | ||
| ``` |
There was a problem hiding this comment.
i think we should keep Chinese doc unchanged, and let @JunweiSUN to update Chinese doc using crowdin
|
Would you mind telling me why FPGM Pruner is removed? Is former version not work or you guys are work on a new version now? I may have to use filter pruning for CNN while there is almost no tensorflow related toolkit for it. Is there a stable version here?(no need to be latest) Much thanks! |
@tu1258 As model compression framework on NNI has been refactored, we will re-implement the algorithms on the new framework for tensorflow. also, highly recommend external contributors to contribute the compression algorithms. |
(cherry picked from commit 5623dbf)
No description provided.