Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26782
At least we should be consistent on top-level APIs and prepare/convert/etc.
Logic is inplace=False by default but top-level APIs take care of doing fewer copies.
Also renames always-inplace methods like add_observer to have underscore in the end.
One fix for MinMaxObserver was triggered by deepcopy surfacing that we were accidentally keeping autograd around
Test Plan: Imported from OSS
Differential Revision: D17595956
Pulled By: dzhulgakov
fbshipit-source-id: 801f9f5536b553f24c7a660064dd6fce685edd65
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26709
Polishes implementation from #25975. Primarily, we use NoopObserver to communicate that weights need to be quantized to float16. The very top-level API (quantize_dynamic) stays the same with `dtype` argument but the implementation follows the common flow.
One can argue that dynamic fp16 quantization doesn't really fit into the 'observer' mechanism. It's in fact not ideal, but it's better to have the same flow than branching on both dtype and qconfig.
Test Plan: Imported from OSS
Differential Revision: D17544103
Pulled By: dzhulgakov
fbshipit-source-id: 6af3f18c35929a1a53ea734079c005f656e4925f
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26492
Previous definition of observers was quite clumsy - with things like `default_observer()()`. This PR strips a way a lot of craft and allows to pass just class names directly. In order to override default arguments either `functools.partial` can be used or convenient wrapper `MyObserver.with_args(x=1)` is provided.
Also rename `QConfig_dynamic` to `QConfigDynamic` because it violates the naming convention.
Test Plan: Imported from OSS
Differential Revision: D17521265
Pulled By: dzhulgakov
fbshipit-source-id: ba9df19b368641acf4093c43df9990796284fd9e
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26574
Since we also have `quantized::linear`, `quantize_linear` sounds
confusing, so we plan to rename it before the branch cut
Test Plan:
ci
Imported from OSS
Differential Revision: D17514876
fbshipit-source-id: 01d9005e6ec8cb9950b9d8bba122109c389641d3
Summary:
Mainly want to resolve comments from https://github.com/pytorch/pytorch/pull/25830.
Overall, we want to provide a recording observer for recording the runtime tensor values of activation path in order to debug the numerical accuracy loss offline.
According to the feedback from https://github.com/pytorch/pytorch/issues/25830, it might be better to record all the observers in a dict and query the dict to get corresponding tensor values. hx89 is working on how to insert the recording observers into model under debug.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26413
Differential Revision: D17506502
Pulled By: llyfacebook
fbshipit-source-id: 3ab90dc78920e7ec3fa572c2a07327a9991c530a
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25975
We would like to add the FP16 weight support for the dynamic quantized LSTM.
Test Plan:
buck test mode/dev caffe2/test:quantization -- 'test_quantized_rnn \(test_quantization\.PostTrainingDynamicQuantTest\)' --print-passing-details
```
[jianyuhuang@devvm794.ftw3.facebook.com: ~/fbsource/fbcode/caffe2/test] $ buck test mode/dev caffe2/test:quantization
-- 'test_quantized_rnn \(test_quantization\.PostTrainingDynamicQuantTest\)' --print-passing-details
Building: finished in 13.4 sec (100%) 8134/8134 jobs, 81 updated
Total time: 13.9 sec
Trace available for this run at /tmp/testpilot.20190910-210241.2092790.log
TestPilot test runner for Facebook. See https://fburl.com/testpilot for details.
Testpilot build revision c86e65add357582accb6ec0be23b92c8a2c510bd fbpkg ca46e8f5b26c451a8b0b2462c11bb61d at Mon Sep 9
22:16:37 2019 by twsvcscm from /usr/local/fbprojects/packages/testinfra.testpilot/696/t.par
Discovering tests
Running 1 tests
Started new test run: https://our.intern.facebook.com/intern/testinfra/testrun/1125900050322971
✓ caffe2/test:quantization - test_quantized_rnn (test_quantization.PostTrainingDynamicQuantTest) 0.183 1/1 (passed)
Test output:
> test_quantized_rnn (test_quantization.PostTrainingDynamicQuantTest) ... ok
>
> ----------------------------------------------------------------------
> Ran 1 test in 0.184s
>
> OK
Finished test run: https://our.intern.facebook.com/intern/testinfra/testrun/1125900050322971
Summary (total time 4.35s):
PASS: 1
FAIL: 0
SKIP: 0
FATAL: 0
TIMEOUT: 0
OMIT: 0
```
Differential Revision: D17299116
fbshipit-source-id: 7fe91ece25867f2c0496f1b63fb1041e6b815166
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25887
ghstack-source-id: 90383258
Add per channel observer to compute the qparams for each channel.
Test Plan:
buck test mode/dev caffe2/test:quantization -- 'test_per_channel_minmax_observer'
buck test mode/dev caffe2/test:quantization -- 'test_per_channel_minmax_observer_scriptable'
Differential Revision: D17137226
fbshipit-source-id: 0b1c93e3cbcda86f5c4e30f7cd94c670f2665063
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/24022
In histogram observer add an approximation for L2 error minimization for selecting min/max.
By selecting new min/max, we filter out outliers in input distribution.
This follows the implementation of NormMinimization::NonlinearQuantizationParamsSearch in caffe2/quantization/server/norm_minimization.cc
ghstack-source-id: 90298789
Test Plan: buck test mode/dev caffe2/test:quantization -- 'test_histogram_observer'
Differential Revision: D16713239
fbshipit-source-id: 82631ba47974e25689c9c66bc3088117090e26d4
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23959
Add histogram observer that records the running histogram of tensor values along with min/max values.
ghstack-source-id: 90076996
Test Plan:
Added a test test_histogram_observer
buck test mode/dev caffe2/test:quantization -- 'test_histogram_observer'
buck test mode/dev caffe2/test:quantization -- 'test_observer_scriptable'
Differential Revision: D16692835
fbshipit-source-id: 0f047d3349cb9770fad4a2b6cb346c51d9e99cd4
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25976
As recommended in https://github.com/pytorch/pytorch/pull/25877/files#r322956051:
> We should move more of these toward using BytesIO. Using files in tests is generally considered bad practice because it introduces syscalls and dependencies on the execution environment, and thus can cause test flakiness/instability.
ghstack-source-id: 89929947
Test Plan: CI
Differential Revision: D17310441
fbshipit-source-id: ba97cce4224225df45ff44062f1bc8ebefb25922
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25667
Relax scale and zero-point for activations to ensure that fbgemm implementations of conv and linear do not saturate due to 16 bit intermediate accumulation.
Add test to verify precision of numerics of quantized model with updated observer. This test catches errors in
handling layouts for quantized ops in addition to saturation/quantization errors.
ghstack-source-id: 89587942
Test Plan:
buck test caffe2/test:quantized -- 'test_float_quant_compare \(test_quantized_models\.ModelNumerics\)' --print-passing-details
Passes when SQNR > 35 dB
buck test caffe2/test:quantization -- 'test_minmax_observer \(test_quantization\.ObserverTest\)' --print-passing-details
Passes with additional coverage for observer changes
Differential Revision: D17140498
fbshipit-source-id: 42c58e726bb0b0f51890590ee2525428f9a8d24e
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25157
Add the dynamic quantized LSTM module.
TODO (separate PRs):
- Serialization.
- Bias can be Null.
ghstack-source-id: 89443731
Test Plan:
buck test mode/dev caffe2/test:quantization -- 'test_quantized_rnn \(test_quantization\.PostTrainingDynamicQuantTest\)' --print-passing-details
```
[jianyuhuang@devvm2816.prn3.facebook.com: ~/fbsource/fbcode/caffe2/test] $ buck test mode/dev caffe2/test:quantization -- 'test_quantized_rnn \(test_q
uantization\.PostTrainingDynamicQuantTest\)' --print-passing-details
Action graph will be rebuilt because files have been added or removed.
Parsing buck files: finished in 1.4 sec
Building: finished in 4.0 sec (100%) 8122/8122 jobs, 2 updated
Total time: 5.5 sec
Trace available for this run at /tmp/testpilot.20190902-164918.1275502.log
TestPilot test runner for Facebook. See https://fburl.com/testpilot for details.
Testpilot build revision b61bc0e3b71033578eddfe0a28b0739bc685663f fbpkg 3b1c1aed1c534c0cb161a981eca6e2f0 at Sun Sep 1 20:58:52 2019 by twsvcscm from /usr/local/fbprojects/packages/testinfra.testpilot/690/t.par
Discovering tests
Running 1 tests
Started new test run: https://our.intern.facebook.com/intern/testinfra/testrun/2251799823877227
✓ caffe2/test:quantization - test_quantized_rnn (test_quantization.PostTrainingDynamicQuantTest) 1.048 1/1 (passed)
Test output:
> test_quantized_rnn (test_quantization.PostTrainingDynamicQuantTest) ... ok
>
> ----------------------------------------------------------------------
> Ran 1 test in 1.049s
>
> OK
Finished test run: https://our.intern.facebook.com/intern/testinfra/testrun/2251799823877227
Summary (total time 5.53s):
PASS: 1
FAIL: 0
SKIP: 0
FATAL: 0
TIMEOUT: 0
OMIT: 0
```
Differential Revision: D16955662
fbshipit-source-id: 61cf1a74913105fa02e44b3941813eabac0006b5
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23753
Add intrinsic(fused) module mappings in quantize.py to enable mapping fused modules
in both QAT and post PTQ
Differential Revision: D16820749
fbshipit-source-id: 07de76a4f09b44bde8b193c103eac02c22b875b6
Summary:
We want to use the Module type as the key for the qconfig_dict for the module replacement during the quantization.
Before this Diff, to dynamic quantize the BERT model, we have to specify each layer:
```
qconfig_dict = {
'encoder.layer.0.attention.self.query': default_qconfig,
'encoder.layer.0.attention.self.key': default_qconfig,
'encoder.layer.0.attention.self.value': default_qconfig,
'encoder.layer.0.attention.output.dense': default_qconfig,
'encoder.layer.0.intermediate.dense': default_qconfig,
'encoder.layer.0.output.dense': default_qconfig,
'encoder.layer.1.attention.self.query': default_qconfig,
'encoder.layer.1.attention.self.key': default_qconfig,
'encoder.layer.1.attention.self.value': default_qconfig,
'encoder.layer.1.attention.output.dense': default_qconfig,
'encoder.layer.1.intermediate.dense': default_qconfig,
'encoder.layer.1.output.dense': default_qconfig,
...
}
```
After this Diff, we only need the following
```
qconfig_dict = {
torch.nn.Linear : default_qconfig
}
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23212
ghstack-source-id: 88287091
Reviewed By: zafartahirov
Differential Revision: D16436542
fbshipit-source-id: 11fbe68ee460560c1a7cdded63581eb7a00e5a89
Summary:
- ~~Add a unit test for the Dynamic Quantized Linear operator (```torch.fbgemm_linear_quantize_weight```, ```torch.fbgemm_pack_quantized_matrix```, and ```torch.fbgemm_linear_int8_weight```) in ```test_quantized.py```.~~ Move this to D16404027 for a separate review.
- Add the Dynamic Quantized Linear module in ```torch/nn/quantized/modules/linear.py```. ~~This is in a rudimentary stage. Will add more functions later~~.
- Add the torch.quantize logic (prepare, eval, convert) for dynamic quantization.
- Add a unit test for the Dynamic Quantized Linear module in ```test_nn_quantized.py```.
- Add a unit test for the Model-level Quantization API
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23128
ghstack-source-id: 88257232
Differential Revision: D16258664
fbshipit-source-id: 4be3ac39ee27c088b341c741d3f09f51d5a23ef0
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23635
It appears it is the same complexity to add new modules using a base class and using a generation script.
Test Plan: Imported from OSS
Differential Revision: D16593364
Pulled By: zafartahirov
fbshipit-source-id: 852dcf41f3dfa2a89152042b8e61d0b6defa8feb
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23891
This adds an initial set of testing coverage for quantization that checks if the modules can be scripted. Testing for tracing and serialization is forthcoming
Test Plan: Imported from OSS
Differential Revision: D16698045
Pulled By: jamesr66a
fbshipit-source-id: 96d80d938b816220af72359165a7b96d998a30c9
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23858
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23718
Changes:
- Enable tests for quantization test files in `run_tests.py`
- Remove `__future__` imports from `torch/nn/qat/modules/__init__.py`, since `unicode_literals` messes up imports on python2 because the elements in `__all__` will be Unicode and not string
- Skip PostTrainingQuantTests if the build doesn't have FBGEMM (only a small subset of targets in tests) or if testing under UBSAN (the suppression file doesn't seem to work)
Test Plan: Imported from OSS
Reviewed By: ZolotukhinM
Differential Revision: D16639467
Pulled By: jamesr66a
fbshipit-source-id: 532766797c216976dd7e07d751f768ff8e0fc207
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23465
We decided not to allow user to use qconfig_dict to do quantization
since that API is not robust.
Differential Revision: D16611504
fbshipit-source-id: b0d1d311b32c990a165c480f50e9ce3d68b785b5
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23003
torch.quantization.fuse_module and torch.nn._intrinsic convRelu and LinearRelu
Fusion function to combine specific modules: (conv,bn) and (conv,bn,relu).
In all cases, replace modules in place. The first module is replaced with the _intrinsic fused module and the remaining modules are replaced by nn.Identity.
Support both training and eval. For training, the modules are "fused" with a sequential container. This is to allow for further module swaps for quantization aware training.
Also add: torch.nn._intrinsic for convRelu and LinearRelu.
TODO: Add tests for _intrinsic modules.
Conv BN fusion code is based on DsKhudia's implementation
Differential Revision: D16199720
fbshipit-source-id: 95fb9ffe72b361d280313b2ec57de2acd4f9dda2
Summary:
Add support for quantization aware training in eager mode
Modifications to Post training flow:
## Prepare
* Fusion: e.g. (Conv, Bn) → ConvBn (float)
* Swapping: To insert fake_quant to weight, we need to swap the float modules that has weight with different qat modules, e.g. Conv → torch.nn.qat.Conv , ConvBn → torch.nn._intrinsic.qat.ConvBn
```
* previously we were thinking about modify the weight in forward_pre hook and change it back in forward_hook:
* def forward_pre_hook(self, input):
self.float_weight = self.weight
self.weight = self.fake_quantize(self.float_weight)
def forward_hook(self, input):
self.weight = self.float_weight
```
* Assignments to self.weight are needed because we can’t change forward function and in forward function they are using self.weight.
* But we will need to keep two copies of weight in this case, so it’s probably better to just swap the module
* So we want to just swap Conv to torch.nn.qat.Conv and Linear to torch.nn.qat.Linear
* qat modules will have fake_quant for output and weights inserted in forward function
## Convert
* flow should be identical to ptq, but the swapping dictionary is slightly different since modules are changed in prepare step.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23082
ghstack-source-id: 86824650
Differential Revision: D16379374
fbshipit-source-id: 7d16d1acd87025065a24942ff92abf18e9fc8070
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22732
Add support for quantization aware training in eager mode
Modifications to Post training flow:
## Prepare
* Fusion: e.g. (Conv, Bn) → ConvBn (float)
* Swapping: To insert fake_quant to weight, we need to swap the float modules that has weight with different qat modules, e.g. Conv → torch.nn.qat.Conv , ConvBn → torch.nn._intrinsic.qat.ConvBn
```
* previously we were thinking about modify the weight in forward_pre hook and change it back in forward_hook:
* def forward_pre_hook(self, input):
self.float_weight = self.weight
self.weight = self.fake_quantize(self.float_weight)
def forward_hook(self, input):
self.weight = self.float_weight
```
* Assignments to self.weight are needed because we can’t change forward function and in forward function they are using self.weight.
* But we will need to keep two copies of weight in this case, so it’s probably better to just swap the module
* So we want to just swap Conv to torch.nn.qat.Conv and Linear to torch.nn.qat.Linear
* qat modules will have fake_quant for output and weights inserted in forward function
## Convert
* flow should be identical to ptq, but the swapping dictionary is slightly different since modules are changed in prepare step.
Reviewed By: zafartahirov
Differential Revision: D16199356
fbshipit-source-id: 62aeaf47c12c62a87d9cac208f25f7592e245d6c
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22706
Moved the models used for quantization test from the test_quantization.py file to common_quantization.py
Reviewed By: jerryzh168
Differential Revision: D16189865
fbshipit-source-id: 409b43454b6b3fe278ac16b1affb9085d6ed6835
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22694
Move quantization and quantized utility functions for testing to common_quantized.py and common_quantization.py. Addditionally, add a quantized test case base class which contains common methods for checking the results of quantization on modules. As a consequence of the move, fixed the import at the top of test_quantized.py, and test_quantization to use the new utility
Reviewed By: jerryzh168
Differential Revision: D16172012
fbshipit-source-id: 329166af5555fc829f26bf1383d682c25c01a7d9
