torch.quantization conversion utilities, observers for eager mode quantization #22010
Conversation
Differential Revision: D15375695 Differential Version: 85144759
Differential Revision: D15483071 Differential Version: 85154440
Differential Revision: D15375695 Differential Version: 85155178
Differential Revision: D15554224 Differential Version: 85162142
Differential Revision: D15483071 Differential Version: 85187565
Differential Revision: D15483071 Differential Version: 85213062
Differential Revision: D15554224 Differential Version: 85213957
Differential Revision: D15375695 Differential Version: 85219380
Differential Revision: D15554224 Differential Version: 85217753
Differential Revision: D15554183 Differential Version: 85267282
Differential Revision: D15554183 Differential Version: 85267665
Differential Revision: D15483071 Differential Version: 85322492
Differential Revision: D15554224 Differential Version: 85325488
Differential Revision: D15554224 Differential Version: 85330125
Differential Revision: D15483071 Differential Version: 85331722
dzhulgakov
left a comment
dzhulgakov
left a comment
There was a problem hiding this comment.
some comments (I feel I already provided some of these on Raghu's prototype but I'm not sure where it was)
Differential Revision: D15375695 Differential Version: 85439410
Differential Revision: D15483071 Differential Version: 85439430
Differential Revision: D15554224 Differential Version: 85439441
Differential Revision: D15554183 Differential Version: 85439445
Differential Revision: D15554183 Differential Version: 85439488
dzhulgakov
left a comment
dzhulgakov
left a comment
There was a problem hiding this comment.
getting there, but there's still a lot to improve in terms of clarity of code
dzhulgakov
left a comment
dzhulgakov
left a comment
There was a problem hiding this comment.
Yay! It looks good to me. I'll let @raghuramank100 do a pass and answer question about exponential smoothing. But I think it's good to go
raghuramank100
left a comment
raghuramank100
left a comment
There was a problem hiding this comment.
Almost there, a few more comments.
Differential Revision: D15554183 Differential Version: 86050037
| """ | ||
| self.observer(output) | ||
|
|
||
| # TODO(jerryzh): remove_observer? |
There was a problem hiding this comment.
Not sure if we really need it. We can instead show in the tutorial how to do a deepcopy, so that the original float module is still available.
Differential Revision: D15554183 Differential Version: 86079795
raghuramank100
left a comment
raghuramank100
left a comment
There was a problem hiding this comment.
Please see comments, thanks
Differential Revision: D15554183 Differential Version: 86093005
Differential Revision: D15554183 Differential Version: 86094049
raghuramank100
left a comment
raghuramank100
left a comment
There was a problem hiding this comment.
Add an option to prepare: i.e enable/disable add_quant_dequant. This allows for tq.quantize to work for all the test cases: including quantwrapper and manual quantstub insertion.
Differential Revision: D15554183 Differential Version: 86106388
raghuramank100
left a comment
raghuramank100
left a comment
There was a problem hiding this comment.
Just one more test requested. Thanks!
|
@raghuramank100 what test? I have removed add_quant_dequant from default and changed the previous test to call everything explicitly. The quantize api are used in the last few tests(manual and quantwrapper test cases) |
Differential Revision: D15554183 Differential Version: 86168758
|
This pull request has been merged in 5040d52. |
|
This broke lint https://travis-ci.org/pytorch/pytorch/jobs/556423310 |
Stack:
:black_circle: #22010 torch.quantization conversion utilities, observers for eager mode quantization 💚
torch.quantization module with observers and conversion routines
Differential Revision: D15554183