mct-nightly 2.2.0.20240902.511__py3-none-any.whl → 2.2.0.20240904.449__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (35) hide show
  1. {mct_nightly-2.2.0.20240902.511.dist-info → mct_nightly-2.2.0.20240904.449.dist-info}/METADATA +6 -6
  2. {mct_nightly-2.2.0.20240902.511.dist-info → mct_nightly-2.2.0.20240904.449.dist-info}/RECORD +35 -26
  3. model_compression_toolkit/__init__.py +1 -1
  4. model_compression_toolkit/gptq/pytorch/quantizer/quantization_builder.py +1 -2
  5. model_compression_toolkit/qat/__init__.py +2 -2
  6. model_compression_toolkit/qat/common/qat_config.py +1 -19
  7. model_compression_toolkit/qat/keras/quantization_facade.py +1 -1
  8. model_compression_toolkit/qat/keras/quantizer/lsq/symmetric_lsq.py +1 -1
  9. model_compression_toolkit/qat/keras/quantizer/lsq/uniform_lsq.py +1 -1
  10. model_compression_toolkit/qat/keras/quantizer/ste_rounding/symmetric_ste.py +1 -1
  11. model_compression_toolkit/qat/keras/quantizer/ste_rounding/uniform_ste.py +1 -1
  12. model_compression_toolkit/qat/pytorch/quantizer/{base_pytorch_qat_quantizer.py → base_pytorch_qat_weight_quantizer.py} +4 -13
  13. model_compression_toolkit/qat/pytorch/quantizer/lsq/symmetric_lsq.py +6 -116
  14. model_compression_toolkit/qat/pytorch/quantizer/lsq/uniform_lsq.py +12 -122
  15. model_compression_toolkit/qat/pytorch/quantizer/quantization_builder.py +8 -7
  16. model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/symmetric_ste.py +6 -84
  17. model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/uniform_ste.py +6 -85
  18. model_compression_toolkit/trainable_infrastructure/__init__.py +9 -3
  19. model_compression_toolkit/trainable_infrastructure/common/base_trainable_quantizer.py +9 -8
  20. model_compression_toolkit/trainable_infrastructure/common/training_method.py +31 -0
  21. model_compression_toolkit/trainable_infrastructure/keras/base_keras_quantizer.py +2 -2
  22. model_compression_toolkit/trainable_infrastructure/keras/quantize_wrapper.py +2 -2
  23. model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/__init__.py +19 -0
  24. model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/base_activation_quantizer.py +22 -0
  25. model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/lsq/__init__.py +14 -0
  26. model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/lsq/symmetric_lsq.py +111 -0
  27. model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/lsq/uniform_lsq.py +106 -0
  28. model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/ste/__init__.py +14 -0
  29. model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/ste/symmetric_ste.py +108 -0
  30. model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/ste/uniform_ste.py +105 -0
  31. model_compression_toolkit/trainable_infrastructure/pytorch/base_pytorch_quantizer.py +7 -14
  32. model_compression_toolkit/{qat/pytorch/quantizer → trainable_infrastructure/pytorch}/quantizer_utils.py +79 -2
  33. {mct_nightly-2.2.0.20240902.511.dist-info → mct_nightly-2.2.0.20240904.449.dist-info}/LICENSE.md +0 -0
  34. {mct_nightly-2.2.0.20240902.511.dist-info → mct_nightly-2.2.0.20240904.449.dist-info}/WHEEL +0 -0
  35. {mct_nightly-2.2.0.20240902.511.dist-info → mct_nightly-2.2.0.20240904.449.dist-info}/top_level.txt +0 -0
{mct_nightly-2.2.0.20240902.511.dist-info → mct_nightly-2.2.0.20240904.449.dist-info}/METADATA RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.1
2
2
  Name: mct-nightly
3
- Version: 2.2.0.20240902.511
3
+ Version: 2.2.0.20240904.449
4
4
  Summary: A Model Compression Toolkit for neural networks
5
5
  Home-page: UNKNOWN
6
6
  License: UNKNOWN
@@ -78,11 +78,11 @@ for hands-on learning. For example:
78
78
  Currently, MCT is being tested on various Python, Pytorch and TensorFlow versions:
79
79
 
80
80
 
81
- | | PyTorch 2.1 | PyTorch 2.2 | PyTorch 2.3 |
82
- |-------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
83
- | Python 3.9 | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch21.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch21.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch22.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch22.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch23.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch23.yml) |
84
- | Python 3.10 | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch21.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch21.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch22.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch22.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch23.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch23.yml) |
85
- | Python 3.11 | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch21.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch21.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch22.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch22.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch23.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch23.yml) |
81
+ | | PyTorch 2.1 | PyTorch 2.2 | PyTorch 2.3 | PyTorch 2.4 |
82
+ |-------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
83
+ | Python 3.9 | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch21.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch21.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch22.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch22.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch23.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch23.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch24.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_pytorch24.yml) |
84
+ | Python 3.10 | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch21.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch21.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch22.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch22.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch23.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch23.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch24.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_pytorch24.yml) |
85
+ | Python 3.11 | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch21.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch21.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch22.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch22.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch23.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch23.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch24.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_pytorch24.yml) |
86
86
 
87
87
 
88
88
 
{mct_nightly-2.2.0.20240902.511.dist-info → mct_nightly-2.2.0.20240904.449.dist-info}/RECORD RENAMED
@@ -1,4 +1,4 @@
1
- model_compression_toolkit/__init__.py,sha256=3gZ7KaDN9YVyZ8h3kQ--DTXrObcp81y40HsqykgSoko,1573
1
+ model_compression_toolkit/__init__.py,sha256=j0NwTQQJFkPKcEOk_ysFr8-24o0scVFQ47S0VGi7HVA,1573
2
2
  model_compression_toolkit/constants.py,sha256=i4wYheBkIdQmsQA-axIpcT3YiSO1USNc-jaNiNE8w6E,3920
3
3
  model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
4
4
  model_compression_toolkit/logger.py,sha256=3DByV41XHRR3kLTJNbpaMmikL8icd9e1N-nkQAY9oDk,4567
@@ -375,7 +375,7 @@ model_compression_toolkit/gptq/pytorch/quantization_facade.py,sha256=TMus5LYJnTn
375
375
  model_compression_toolkit/gptq/pytorch/quantizer/__init__.py,sha256=ZHNHo1yzye44m9_ht4UUZfTpK01RiVR3Tr74-vtnOGI,968
376
376
  model_compression_toolkit/gptq/pytorch/quantizer/base_pytorch_gptq_quantizer.py,sha256=fKg-PNOhGBiL-4eySS9Fyw0GkA76Pq8jT_HbJuJ8iZU,4143
377
377
  model_compression_toolkit/gptq/pytorch/quantizer/quant_utils.py,sha256=OocYYRqvl7rZ37QT0hTzfJnWGiNCPskg7cziTlR7TRk,3893
378
- model_compression_toolkit/gptq/pytorch/quantizer/quantization_builder.py,sha256=uT9N_aBj965hvQfKd67fS1B0SXGnOLVcqa3wW4b2iZE,4566
378
+ model_compression_toolkit/gptq/pytorch/quantizer/quantization_builder.py,sha256=Lf334209uVFXuRKIFqVvq9RyEcv014Bozt1hr_O6XjQ,4447
379
379
  model_compression_toolkit/gptq/pytorch/quantizer/regularization_factory.py,sha256=mDWZERLwtDzqWeJUwHMVyGdlS8wPLjJ3NvZiKBP6BNA,1959
380
380
  model_compression_toolkit/gptq/pytorch/quantizer/soft_rounding/__init__.py,sha256=lNJ29DYxaLUPDstRDA1PGI5r9Fulq_hvrZMlhst1Z5g,697
381
381
  model_compression_toolkit/gptq/pytorch/quantizer/soft_rounding/soft_quantizer_reg.py,sha256=oO7WgsAHMnWoXNm_gTKAAe-Nd79mGL_m677ai-ui424,4132
@@ -394,33 +394,32 @@ model_compression_toolkit/ptq/keras/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4d
394
394
  model_compression_toolkit/ptq/keras/quantization_facade.py,sha256=DAAJPd6pKLgiwoJT-_u2dvVOO4Ox6IgJgfiUbnNRBwQ,10968
395
395
  model_compression_toolkit/ptq/pytorch/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
396
396
  model_compression_toolkit/ptq/pytorch/quantization_facade.py,sha256=xHVTrm9Fyk_j4j8G1Pb97qacN_gn9cGYpsT1HXdTc1A,9305
397
- model_compression_toolkit/qat/__init__.py,sha256=kj2qsZh_Ca7PncsHKcaL5EVT2H8g4hYtvaQ3KFxOkwE,1143
397
+ model_compression_toolkit/qat/__init__.py,sha256=b2mURFGsvaZz_CdAD_w2I4Cdu8ZDN-2iGHMBHTKT5ws,1128
398
398
  model_compression_toolkit/qat/common/__init__.py,sha256=6tLZ4R4pYP6QVztLVQC_jik2nES3l4uhML0qUxZrezk,829
399
- model_compression_toolkit/qat/common/qat_config.py,sha256=zoq0Vb74vCY7WlWD8JH_KPrHDoUHSvMc3gcO53u7L2U,3394
399
+ model_compression_toolkit/qat/common/qat_config.py,sha256=xtfVSoyELGXynHNrw86dB9FU3Inu0zwehc3wLrh7JvY,2918
400
400
  model_compression_toolkit/qat/keras/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
401
- model_compression_toolkit/qat/keras/quantization_facade.py,sha256=i5uRwOWLxsAcLGFAdTGEe-nLaCdcDYz6ojfhByRtIJg,17270
401
+ model_compression_toolkit/qat/keras/quantization_facade.py,sha256=VaZTqK53TOWrXebnJzoHHD99DxOgS4NzHGbmYWaajWA,17274
402
402
  model_compression_toolkit/qat/keras/quantizer/__init__.py,sha256=zmYyCa25_KLCSUCGUDRslh3RCIjcRMxc_oXa54Aui-4,996
403
403
  model_compression_toolkit/qat/keras/quantizer/base_keras_qat_quantizer.py,sha256=hoY3AETaLSRP7YfecZ32tyUUj-X_DHRWkV8nALYeRlY,2202
404
404
  model_compression_toolkit/qat/keras/quantizer/quant_utils.py,sha256=cBULOgWUodcBO1lHevZggdTevuDYI6tQceV86U2x6DA,2543
405
405
  model_compression_toolkit/qat/keras/quantizer/quantization_builder.py,sha256=HD0JIOiqnrpqj5qk6RyzuCsSGZsDUVohdCYSePmJBNQ,5872
406
406
  model_compression_toolkit/qat/keras/quantizer/lsq/__init__.py,sha256=lNJ29DYxaLUPDstRDA1PGI5r9Fulq_hvrZMlhst1Z5g,697
407
- model_compression_toolkit/qat/keras/quantizer/lsq/symmetric_lsq.py,sha256=T9aN_tNh8Hr_zmN5GT54eHl3pSx71U4l7ICIJA2v5hI,12022
408
- model_compression_toolkit/qat/keras/quantizer/lsq/uniform_lsq.py,sha256=VSHyk2-rDRcugT_-4SLiU1s78uY40THt_TM_hR6YbdU,11223
407
+ model_compression_toolkit/qat/keras/quantizer/lsq/symmetric_lsq.py,sha256=MwHo4qUYTm-cZZ9f4bEDU2fcdO1VdLXcrp8MKhJ051k,12043
408
+ model_compression_toolkit/qat/keras/quantizer/lsq/uniform_lsq.py,sha256=lGMJF_8jgHV2Rp97aMIqt7B7Gn7JsEOVbBW55K9tvuI,11244
409
409
  model_compression_toolkit/qat/keras/quantizer/ste_rounding/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
410
- model_compression_toolkit/qat/keras/quantizer/ste_rounding/symmetric_ste.py,sha256=I4KlaGv17k71IyjuSG9M0OlXlD5P0pfvKa6oCyRQ5FE,13517
411
- model_compression_toolkit/qat/keras/quantizer/ste_rounding/uniform_ste.py,sha256=EED6LfqhX_OhDRJ9e4GwbpgNC9vq7hoXyJS2VPvG2qc,10789
410
+ model_compression_toolkit/qat/keras/quantizer/ste_rounding/symmetric_ste.py,sha256=fPAC49mBlB5ViaQT_xHUTC8EvH84OsBX3WAPusqYcM8,13538
411
+ model_compression_toolkit/qat/keras/quantizer/ste_rounding/uniform_ste.py,sha256=6YS0v1qCq5dRqtLKHc2gHaKJWfql84TxtZ7pypaZock,10810
412
412
  model_compression_toolkit/qat/pytorch/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
413
413
  model_compression_toolkit/qat/pytorch/quantization_facade.py,sha256=1eg0jMgFzRLYIFnG9GJnJ8U3W4IOM-4Z27s9Wq-JeOQ,13452
414
414
  model_compression_toolkit/qat/pytorch/quantizer/__init__.py,sha256=xYa4C8pr9cG1f3mQQcBXO_u3IdJN-zl7leZxuXDs86w,1003
415
- model_compression_toolkit/qat/pytorch/quantizer/base_pytorch_qat_quantizer.py,sha256=WQSrtoWmRhyJnABrO6lwUtJruwLFZjBzLxbYh3banYI,2213
416
- model_compression_toolkit/qat/pytorch/quantizer/quantization_builder.py,sha256=sFWGu76PZ9dSRf3L0uZI6YwLIs0biBND1tl76I1piBQ,5721
417
- model_compression_toolkit/qat/pytorch/quantizer/quantizer_utils.py,sha256=nO7IrDRo5b9Asf21WJacE4vf5voD3UzF_oGjBoGusD4,5335
415
+ model_compression_toolkit/qat/pytorch/quantizer/base_pytorch_qat_weight_quantizer.py,sha256=gjzrnBAZr5c_OrDpSjxpQYa_jKImv7ll52cng07_2oE,1813
416
+ model_compression_toolkit/qat/pytorch/quantizer/quantization_builder.py,sha256=lM10cGUkkTDtRyLLdWj5Rk0cgvcxp0uaCseyvrnk_Vg,5752
418
417
  model_compression_toolkit/qat/pytorch/quantizer/lsq/__init__.py,sha256=huHoBUcKNB6BnY6YaUCcFvdyBtBI172ZoUD8ZYeNc6o,696
419
- model_compression_toolkit/qat/pytorch/quantizer/lsq/symmetric_lsq.py,sha256=HihuaMi0P0OQkNZlZAE-QeYlK_4AqcDKV6N405SdgI0,10712
420
- model_compression_toolkit/qat/pytorch/quantizer/lsq/uniform_lsq.py,sha256=fqAI151SUo9OYN4UvQJIcdB6p1r7HVeteqPqHGxv-tI,10355
418
+ model_compression_toolkit/qat/pytorch/quantizer/lsq/symmetric_lsq.py,sha256=VQuS8v-i_dm4koL-gTotoZzeUxveY4dLBuzayUGa7IE,5943
419
+ model_compression_toolkit/qat/pytorch/quantizer/lsq/uniform_lsq.py,sha256=cOxqop4zZbEBL-sfw0diUDd7WJortGwZPnmlL5-3H7k,5590
421
420
  model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
422
- model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/symmetric_ste.py,sha256=4xmLmg7yN2A7iKnifwkWddgJTWMUiIjFilIuorJeK1A,9657
423
- model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/uniform_ste.py,sha256=HshW016iVAMx7iMkUwlONN2P3K4XgDIu-2AnJnBVSGo,8778
421
+ model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/symmetric_ste.py,sha256=rcYI_qCz_f38VDJ6uZDwDdvvqqpv43vnR8-_zZ4j4CY,6229
422
+ model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/uniform_ste.py,sha256=btk1V6-wG7-rkOJwUF4BuKcxpvPEIrlEOg27JtLj-vE,5543
424
423
  model_compression_toolkit/target_platform_capabilities/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
425
424
  model_compression_toolkit/target_platform_capabilities/constants.py,sha256=iJXGy5um7vhC84Me4ld6EHMhy7jPks0T9ItZX23si6s,1519
426
425
  model_compression_toolkit/target_platform_capabilities/immutable.py,sha256=YhROBiXEIB3TU-bAFrnL3qbAsb1yuWPBAQ_CLOJbYUU,1827
@@ -485,22 +484,32 @@ model_compression_toolkit/target_platform_capabilities/tpc_models/tflite_tpc/v1/
485
484
  model_compression_toolkit/target_platform_capabilities/tpc_models/tflite_tpc/v1/tp_model.py,sha256=rxDkISGCxTB2RaVm59zJWxaJKxGgt4uceDgQ_9E_RmI,10033
486
485
  model_compression_toolkit/target_platform_capabilities/tpc_models/tflite_tpc/v1/tpc_keras.py,sha256=-4vNf2Q6c_rgaac19AFO8hG4ANaPfgNPf0kN44mL6TQ,6830
487
486
  model_compression_toolkit/target_platform_capabilities/tpc_models/tflite_tpc/v1/tpc_pytorch.py,sha256=YVJJvqGPBdkKnug99p9bjqtbfecDXZKIB2iWVCe7RUY,5960
488
- model_compression_toolkit/trainable_infrastructure/__init__.py,sha256=DwWh0lXiLNNzqHHNEy-Py6_5OtseNGJDGNV3SYm8rYQ,1224
487
+ model_compression_toolkit/trainable_infrastructure/__init__.py,sha256=uewpvlPkH9mBFt8IxoAgIfz6iEcvWbOImm_fb6_BxD8,1543
489
488
  model_compression_toolkit/trainable_infrastructure/common/__init__.py,sha256=huHoBUcKNB6BnY6YaUCcFvdyBtBI172ZoUD8ZYeNc6o,696
490
- model_compression_toolkit/trainable_infrastructure/common/base_trainable_quantizer.py,sha256=_5hLxzc0SGowa_f4HuVPjwHdy_M4b7UTooCXROxvCPE,7706
489
+ model_compression_toolkit/trainable_infrastructure/common/base_trainable_quantizer.py,sha256=i5ZX0UnSt_XAgxGyyd7ZRHcocuwTh_FxWgGD2qN7zFc,7735
491
490
  model_compression_toolkit/trainable_infrastructure/common/constants.py,sha256=HN120boJxAnEXNrLSj-o_s-VX4o6C-1ap_KZ4840sd0,875
492
491
  model_compression_toolkit/trainable_infrastructure/common/get_quantizer_config.py,sha256=Jxd4IjS_t0FwnA_S_WmZeVbh4VM6Da9ahKGPLp6ZhQo,6983
493
492
  model_compression_toolkit/trainable_infrastructure/common/get_quantizers.py,sha256=KoX-6LJMsRzXy0i72ve4buJ32cGNQVHVLqHJxhv0lPQ,3428
494
493
  model_compression_toolkit/trainable_infrastructure/common/quant_utils.py,sha256=zdiew1jwR7tUKm9XWlHnAPxIZsAdKqbzzC2vH02j5wA,1505
495
494
  model_compression_toolkit/trainable_infrastructure/common/trainable_quantizer_config.py,sha256=My5Wz34jPOyh8z33OTpKnOobRB0cpO_Qgmtsd5lizHo,4791
495
+ model_compression_toolkit/trainable_infrastructure/common/training_method.py,sha256=LUoeJkloowhZKuHTiOfzjmSUn2G-4of11-rbnL-h0P4,1194
496
496
  model_compression_toolkit/trainable_infrastructure/keras/__init__.py,sha256=huHoBUcKNB6BnY6YaUCcFvdyBtBI172ZoUD8ZYeNc6o,696
497
- model_compression_toolkit/trainable_infrastructure/keras/base_keras_quantizer.py,sha256=9_6ztYvXOBB7_PNf1Syi5zPTwLi5xfbYiKd_UYJ6hwo,4113
497
+ model_compression_toolkit/trainable_infrastructure/keras/base_keras_quantizer.py,sha256=tHEI9vkLjBzdeCD7eTgAHuUubmnq8GbWSF7Coun8zzE,4116
498
498
  model_compression_toolkit/trainable_infrastructure/keras/config_serialization.py,sha256=txdWXdZoHazg-3MDPb9P-oXRM92LRn2G_8woEplwKaI,4360
499
499
  model_compression_toolkit/trainable_infrastructure/keras/load_model.py,sha256=DJHibcLo-UCuHV6UPLeVd7dKmPfkGXEiLqCCqvQrISM,3769
500
- model_compression_toolkit/trainable_infrastructure/keras/quantize_wrapper.py,sha256=aBtXxCAbnqn4PUa3wP55M0W5gKIQVGppLtfgFQ48T6s,5585
500
+ model_compression_toolkit/trainable_infrastructure/keras/quantize_wrapper.py,sha256=eVB5FSE3OmTLrhfLUcP2knwN1z2_unQLM-xFEGwdafA,5587
501
501
  model_compression_toolkit/trainable_infrastructure/keras/quantizer_utils.py,sha256=MVwXNymmFRB2NXIBx4e2mdJ1RfoHxRPYRgjb1MQP5kY,1797
502
502
  model_compression_toolkit/trainable_infrastructure/pytorch/__init__.py,sha256=huHoBUcKNB6BnY6YaUCcFvdyBtBI172ZoUD8ZYeNc6o,696
503
- model_compression_toolkit/trainable_infrastructure/pytorch/base_pytorch_quantizer.py,sha256=hbE_KV3IrBl4XZPgat5gMM0j1Nkv6iwPXzYhonmXBAE,2902
503
+ model_compression_toolkit/trainable_infrastructure/pytorch/base_pytorch_quantizer.py,sha256=7ZFf_E8nFao5f38Qk4-GzGxHgrKTHGj-4ohgPzq2Z7k,2304
504
+ model_compression_toolkit/trainable_infrastructure/pytorch/quantizer_utils.py,sha256=1yOXKghUYfw2hmzbqTuNagIXBoM-wR2bP-ul66-mnDw,7767
505
+ model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/__init__.py,sha256=73CXhqqNTvDpsvlJXclrGJq-vsCUYCI64ILu1y2mtvw,1056
506
+ model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/base_activation_quantizer.py,sha256=X6E6mewWQot_aAkz3UxW5X0-Fjl_aMMjs3A-Af5eL6w,972
507
+ model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/lsq/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
508
+ model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/lsq/symmetric_lsq.py,sha256=0UGoFHAR-RP9aFbAOILbM8kAG9OwUJJZ_g3Rz58SGlY,5462
509
+ model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/lsq/uniform_lsq.py,sha256=BPeunWrYNmbduZGXiZKy5t1ubYREX7QqWOXv2Dt85lk,5285
510
+ model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/ste/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
511
+ model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/ste/symmetric_ste.py,sha256=20DEZgn6ZepcjKrATvciaiQNs2VGf5uwF6f6hDJLOVo,5226
512
+ model_compression_toolkit/trainable_infrastructure/pytorch/activation_quantizers/ste/uniform_ste.py,sha256=1XHClqM7EhNvYiH6sqs6OI3JUGPfjW55v2eQotVwy8c,5010
504
513
  model_compression_toolkit/xquant/__init__.py,sha256=vdmr8sQw3jIBLF9ck7qrskPoXzDKtksHWlMOkU1JUnQ,1003
505
514
  model_compression_toolkit/xquant/common/__init__.py,sha256=ycb1Xt7PtixY2Uabr94JGSwBMcct66O8ZMVf3Qa3ud8,719
506
515
  model_compression_toolkit/xquant/common/constants.py,sha256=k-9LOEv1n_m8dV4chX0dNOTWyhhF7S00E0lkUxtO84E,1592
@@ -527,8 +536,8 @@ model_compression_toolkit/xquant/pytorch/model_analyzer.py,sha256=b93o800yVB3Z-i
527
536
  model_compression_toolkit/xquant/pytorch/pytorch_report_utils.py,sha256=bOc-hFL3gdoSM1Th_S2N_-9JJSlPGpZCTx_QLJHS6lg,3388
528
537
  model_compression_toolkit/xquant/pytorch/similarity_functions.py,sha256=CERxq5K8rqaiE-DlwhZBTUd9x69dtYJlkHOPLB54vm8,2354
529
538
  model_compression_toolkit/xquant/pytorch/tensorboard_utils.py,sha256=mkoEktLFFHtEKzzFRn_jCnxjhJolK12TZ5AQeDHzUO8,9767
530
- mct_nightly-2.2.0.20240902.511.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
531
- mct_nightly-2.2.0.20240902.511.dist-info/METADATA,sha256=LJ1lir-dRUUMItJBfMXAiht4X2oC3UE8dxN3Plcv8pE,19718
532
- mct_nightly-2.2.0.20240902.511.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
533
- mct_nightly-2.2.0.20240902.511.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
534
- mct_nightly-2.2.0.20240902.511.dist-info/RECORD,,
539
+ mct_nightly-2.2.0.20240904.449.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
540
+ mct_nightly-2.2.0.20240904.449.dist-info/METADATA,sha256=SeHK4yipNqQZ45k1ilwb4IdW_j6-k20YV1ewTWUnZVg,20813
541
+ mct_nightly-2.2.0.20240904.449.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
542
+ mct_nightly-2.2.0.20240904.449.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
543
+ mct_nightly-2.2.0.20240904.449.dist-info/RECORD,,
model_compression_toolkit/__init__.py CHANGED
@@ -27,4 +27,4 @@ from model_compression_toolkit import data_generation
27
27
  from model_compression_toolkit import pruning
28
28
  from model_compression_toolkit.trainable_infrastructure.keras.load_model import keras_load_quantized_model
29
29
 
30
- __version__ = "2.2.0.20240902.000511"
30
+ __version__ = "2.2.0.20240904.000449"
model_compression_toolkit/gptq/pytorch/quantizer/quantization_builder.py CHANGED
@@ -27,7 +27,6 @@ from mct_quantizers.pytorch.quantizers import BasePyTorchInferableQuantizer
27
27
  from model_compression_toolkit.logger import Logger
28
28
  from model_compression_toolkit.trainable_infrastructure.common.get_quantizer_config import \
29
29
  get_trainable_quantizer_weights_config
30
- from model_compression_toolkit.qat.pytorch.quantizer.base_pytorch_qat_quantizer import BasePytorchQATTrainableQuantizer
31
30
  from model_compression_toolkit.trainable_infrastructure.common.get_quantizers import \
32
31
  get_trainable_quantizer_class
33
32
 
@@ -35,7 +34,7 @@ from model_compression_toolkit.trainable_infrastructure.common.get_quantizers im
35
34
  def quantization_builder(n: common.BaseNode,
36
35
  gptq_config: GradientPTQConfig,
37
36
  kernel_attr: str = None
38
- ) -> Tuple[Dict[str, BasePytorchQATTrainableQuantizer], List[BasePyTorchInferableQuantizer]]:
37
+ ) -> Tuple[Dict[str, BasePytorchGPTQTrainableQuantizer], List[BasePyTorchInferableQuantizer]]:
39
38
  """
40
39
  Build quantizers for a node according to its quantization configuration and
41
40
  a global NoOpQuantizeConfig object.
model_compression_toolkit/qat/__init__.py CHANGED
@@ -12,7 +12,7 @@
12
12
  # See the License for the specific language governing permissions and
13
13
  # limitations under the License.
14
14
  # ==============================================================================
15
- from model_compression_toolkit.qat.common.qat_config import QATConfig, TrainingMethod
15
+ from model_compression_toolkit.qat.common.qat_config import QATConfig
16
16
 
17
17
  from model_compression_toolkit.qat.keras.quantization_facade import keras_quantization_aware_training_init_experimental, keras_quantization_aware_training_finalize_experimental
18
- from model_compression_toolkit.qat.pytorch.quantization_facade import pytorch_quantization_aware_training_init_experimental, pytorch_quantization_aware_training_finalize_experimental
18
+ from model_compression_toolkit.qat.pytorch.quantization_facade import pytorch_quantization_aware_training_init_experimental, pytorch_quantization_aware_training_finalize_experimental
model_compression_toolkit/qat/common/qat_config.py CHANGED
@@ -14,10 +14,9 @@
14
14
  # ==============================================================================
15
15
 
16
16
  from typing import Dict
17
- from enum import Enum
18
17
  from model_compression_toolkit.core import common
19
18
  from model_compression_toolkit.core.common.framework_info import FrameworkInfo
20
- from model_compression_toolkit.logger import Logger
19
+ from model_compression_toolkit.trainable_infrastructure import TrainingMethod
21
20
 
22
21
 
23
22
  def is_qat_applicable(node: common.BaseNode,
@@ -38,23 +37,6 @@ def is_qat_applicable(node: common.BaseNode,
38
37
  or node.is_activation_quantization_enabled()
39
38
 
40
39
 
41
-
42
- class TrainingMethod(Enum):
43
- """
44
- An enum for selecting a QAT training method
45
-
46
- STE - Standard straight-through estimator. Includes PowerOfTwo, symmetric & uniform quantizers
47
-
48
- DQA - DNN Quantization with Attention. Includes a smooth quantization introduces by DQA method
49
-
50
- LSQ - Learned Step size Quantization. Includes PowerOfTwo, symmetric & uniform quantizers: https://arxiv.org/pdf/1902.08153.pdf
51
-
52
- """
53
- STE = "STE",
54
- DQA = "DQA",
55
- LSQ = "LSQ"
56
-
57
-
58
40
  class QATConfig:
59
41
  """
60
42
  QAT configuration class.
model_compression_toolkit/qat/keras/quantization_facade.py CHANGED
@@ -24,7 +24,6 @@ from model_compression_toolkit.core.common.mixed_precision.resource_utilization_
24
24
  from model_compression_toolkit.core.common.mixed_precision.mixed_precision_quantization_config import \
25
25
  MixedPrecisionQuantizationConfig
26
26
  from mct_quantizers import KerasActivationQuantizationHolder
27
- from model_compression_toolkit.trainable_infrastructure import KerasTrainableQuantizationWrapper
28
27
  from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework import TargetPlatformCapabilities
29
28
  from model_compression_toolkit.core.runner import core_runner
30
29
  from model_compression_toolkit.ptq.runner import ptq_runner
@@ -34,6 +33,7 @@ if FOUND_TF:
34
33
  from tensorflow.keras.layers import Layer
35
34
  from tensorflow.keras.models import Model
36
35
 
36
+ from model_compression_toolkit.trainable_infrastructure import KerasTrainableQuantizationWrapper
37
37
  from model_compression_toolkit.core.keras.default_framework_info import DEFAULT_KERAS_INFO
38
38
  from model_compression_toolkit.core.keras.keras_implementation import KerasImplementation
39
39
  from model_compression_toolkit.core.keras.keras_model_validation import KerasModelValidation
model_compression_toolkit/qat/keras/quantizer/lsq/symmetric_lsq.py CHANGED
@@ -20,7 +20,7 @@ import tensorflow as tf
20
20
  from tensorflow.python.framework.tensor_shape import TensorShape
21
21
  from model_compression_toolkit.constants import SIGNED
22
22
 
23
- from model_compression_toolkit.qat import TrainingMethod
23
+ from model_compression_toolkit.trainable_infrastructure import TrainingMethod
24
24
 
25
25
  from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
26
26
  from model_compression_toolkit.trainable_infrastructure import KerasTrainableQuantizationWrapper
model_compression_toolkit/qat/keras/quantizer/lsq/uniform_lsq.py CHANGED
@@ -18,7 +18,7 @@ from tensorflow.python.framework.tensor_shape import TensorShape
18
18
  from model_compression_toolkit.constants import RANGE_MIN, RANGE_MAX
19
19
  from model_compression_toolkit.trainable_infrastructure.common.constants import FQ_MIN, FQ_MAX
20
20
  from model_compression_toolkit.trainable_infrastructure import KerasTrainableQuantizationWrapper
21
- from model_compression_toolkit.qat import TrainingMethod
21
+ from model_compression_toolkit.trainable_infrastructure import TrainingMethod
22
22
 
23
23
  from mct_quantizers import mark_quantizer, QuantizationMethod, QuantizationTarget
24
24
  from mct_quantizers.keras.quantizers import \
model_compression_toolkit/qat/keras/quantizer/ste_rounding/symmetric_ste.py CHANGED
@@ -21,7 +21,7 @@ from tensorflow.python.framework.tensor_shape import TensorShape
21
21
  from model_compression_toolkit.constants import SIGNED
22
22
  from model_compression_toolkit.trainable_infrastructure.common.constants import FQ_MIN, FQ_MAX
23
23
 
24
- from model_compression_toolkit.qat import TrainingMethod
24
+ from model_compression_toolkit.trainable_infrastructure import TrainingMethod
25
25
 
26
26
  from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
27
27
  from model_compression_toolkit.trainable_infrastructure import KerasTrainableQuantizationWrapper
model_compression_toolkit/qat/keras/quantizer/ste_rounding/uniform_ste.py CHANGED
@@ -18,7 +18,7 @@ from tensorflow.python.framework.tensor_shape import TensorShape
18
18
  from model_compression_toolkit.constants import RANGE_MIN, RANGE_MAX
19
19
  from model_compression_toolkit.trainable_infrastructure.common.constants import FQ_MIN, FQ_MAX
20
20
  from model_compression_toolkit.trainable_infrastructure import KerasTrainableQuantizationWrapper
21
- from model_compression_toolkit.qat import TrainingMethod
21
+ from model_compression_toolkit.trainable_infrastructure import TrainingMethod
22
22
 
23
23
  from mct_quantizers import mark_quantizer, QuantizationMethod, QuantizationTarget
24
24
  from mct_quantizers.keras.quantizers import \
model_compression_toolkit/qat/pytorch/quantizer/{base_pytorch_qat_quantizer.py → base_pytorch_qat_weight_quantizer.py} RENAMED
@@ -24,23 +24,14 @@ from model_compression_toolkit.trainable_infrastructure.pytorch.base_pytorch_qua
24
24
 
25
25
  if FOUND_TORCH:
26
26
 
27
- class BasePytorchQATTrainableQuantizer(BasePytorchTrainableQuantizer):
27
+ class BasePytorchQATWeightTrainableQuantizer(BasePytorchTrainableQuantizer):
28
28
  """
29
- A base class for trainable Keras quantizer for QAT.
29
+ A base class for trainable PyTorch weights quantizer for QAT.
30
30
  """
31
-
32
- def __init__(self,
33
- quantization_config: Union[TrainableQuantizerWeightsConfig, TrainableQuantizerActivationConfig]):
34
- """
35
- Initializes BasePytorchQATTrainableQuantizer object.
36
-
37
- Args:
38
- quantization_config: quantizer config class contains all the information about a quantizer configuration.
39
- """
40
- super().__init__(quantization_config)
31
+ pass
41
32
 
42
33
  else: # pragma: no cover
43
- class BasePytorchQATTrainableQuantizer(BasePytorchTrainableQuantizer):
34
+ class BasePytorchQATWeightTrainableQuantizer(BasePytorchTrainableQuantizer):
44
35
  def __init__(self,
45
36
  quantization_config: Union[TrainableQuantizerWeightsConfig, TrainableQuantizerActivationConfig]):
46
37
  super().__init__(quantization_config)
model_compression_toolkit/qat/pytorch/quantizer/lsq/symmetric_lsq.py CHANGED
@@ -18,56 +18,27 @@ import numpy as np
18
18
  import torch
19
19
  import torch.nn as nn
20
20
 
21
- from model_compression_toolkit.qat import TrainingMethod
22
21
  from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
23
22
  from mct_quantizers import PytorchQuantizationWrapper
24
23
  from model_compression_toolkit.qat.common import THRESHOLD_TENSOR
25
24
  from model_compression_toolkit import constants as C
26
- from model_compression_toolkit.qat.pytorch.quantizer.base_pytorch_qat_quantizer import BasePytorchQATTrainableQuantizer
25
+ from model_compression_toolkit.qat.pytorch.quantizer.base_pytorch_qat_weight_quantizer import BasePytorchQATWeightTrainableQuantizer
27
26
  from mct_quantizers.common.base_inferable_quantizer import mark_quantizer, QuantizationTarget
28
27
 
29
28
  from model_compression_toolkit.core.pytorch.utils import to_torch_tensor
30
- from model_compression_toolkit.qat.pytorch.quantizer.quantizer_utils import ste_round, grad_scale
29
+ from model_compression_toolkit.trainable_infrastructure import TrainingMethod
30
+ from model_compression_toolkit.trainable_infrastructure.pytorch.quantizer_utils import symmetric_lsq_quantizer
31
31
  from mct_quantizers.pytorch.quantizers import \
32
- WeightsPOTInferableQuantizer, WeightsSymmetricInferableQuantizer, ActivationPOTInferableQuantizer, \
33
- ActivationSymmetricInferableQuantizer
32
+ WeightsPOTInferableQuantizer, WeightsSymmetricInferableQuantizer
34
33
  from model_compression_toolkit.trainable_infrastructure.common.trainable_quantizer_config import \
35
- TrainableQuantizerWeightsConfig, TrainableQuantizerActivationConfig
34
+ TrainableQuantizerWeightsConfig
36
35
  from model_compression_toolkit.trainable_infrastructure.common.base_trainable_quantizer import VariableGroup
37
36
 
38
37
 
39
- def symmetric_lsq_quantizer(x: nn.Parameter,
40
- thresholds: nn.Parameter,
41
- num_bits: int,
42
- sign: bool,
43
- min_int: int,
44
- max_int: int,
45
- scale_factor: float) -> Union[nn.Parameter, torch.Tensor]:
46
- """
47
- Symmetric quantizer according to LSQ algorithm: https://arxiv.org/pdf/1902.08153.pdf
48
- Args:
49
- x: input to quantize
50
- thresholds: thresholds of quantization levels
51
- num_bits: number of bits for quantization
52
- sign: whether x is signed or not
53
- min_int: min clipping integer value
54
- max_int: max clipping integer value
55
- scale_factor: grad scale of LSQ algorithm
56
- Returns:
57
- A quantized tensor
58
- """
59
- delta = thresholds / (2 ** (num_bits - int(sign)))
60
- delta_scaled = grad_scale(delta, scale_factor)
61
- rounded = ste_round(x / delta_scaled)
62
- clipped = torch.clip(rounded, min=min_int, max=max_int)
63
- quantized = delta_scaled * clipped
64
- return quantized
65
-
66
-
67
38
  @mark_quantizer(quantization_target=QuantizationTarget.Weights,
68
39
  quantization_method=[QuantizationMethod.POWER_OF_TWO, QuantizationMethod.SYMMETRIC],
69
40
  identifier=TrainingMethod.LSQ)
70
- class LSQWeightQATQuantizer(BasePytorchQATTrainableQuantizer):
41
+ class LSQWeightQATQuantizer(BasePytorchQATWeightTrainableQuantizer):
71
42
  """
72
43
  Trainable constrained quantizer to quantize layer's weights.
73
44
  """
@@ -145,84 +116,3 @@ class LSQWeightQATQuantizer(BasePytorchQATTrainableQuantizer):
145
116
  threshold=threshold_values.tolist(),
146
117
  per_channel=self.quantization_config.weights_per_channel_threshold,
147
118
  channel_axis=self.quantization_config.weights_channels_axis)
148
-
149
-
150
-
151
- @mark_quantizer(quantization_target=QuantizationTarget.Activation,
152
- quantization_method=[QuantizationMethod.POWER_OF_TWO, QuantizationMethod.SYMMETRIC],
153
- identifier=TrainingMethod.LSQ)
154
- class LSQActivationQATQuantizer(BasePytorchQATTrainableQuantizer):
155
- """
156
- Trainable constrained quantizer to quantize layer activations.
157
- """
158
-
159
- def __init__(self, quantization_config: TrainableQuantizerActivationConfig):
160
- """
161
- Initialize a LSQActivationQATQuantizer object with parameters to use
162
- for symmetric or power of two quantization.
163
-
164
- Args:
165
- quantization_config: trainable quantizer config class
166
- """
167
- super().__init__(quantization_config)
168
- self.power_of_two = quantization_config.activation_quantization_method == QuantizationMethod.POWER_OF_TWO
169
- self.sign = quantization_config.activation_quantization_params['is_signed']
170
- self.threshold_values = np.array([quantization_config.activation_quantization_params[C.THRESHOLD]])
171
- self.num_bits = quantization_config.activation_n_bits
172
- n_pos_bits = self.num_bits - int(self.sign)
173
- self.min_int = -int(self.sign) * (2 ** n_pos_bits)
174
- self.max_int = (2 ** n_pos_bits) - 1
175
-
176
- def initialize_quantization(self,
177
- tensor_shape: torch.Size,
178
- name: str,
179
- layer: PytorchQuantizationWrapper):
180
- """
181
- Add quantizer parameters to the quantizer parameters dictionary
182
-
183
- Args:
184
- tensor_shape: tensor shape of the quantized tensor.
185
- name: Tensor name.
186
- layer: Layer to quantize.
187
- """
188
- layer.register_parameter(name, nn.Parameter(to_torch_tensor(self.threshold_values), requires_grad=True))
189
-
190
- # save the quantizer added parameters for later calculations
191
- self.add_quantizer_variable(THRESHOLD_TENSOR, layer.get_parameter(name), VariableGroup.QPARAMS)
192
-
193
- def __call__(self,
194
- inputs: torch.Tensor,
195
- training: bool = True) -> torch.Tensor:
196
- """
197
- Quantize a tensor.
198
- Args:
199
- inputs: Input tensor to quantize.
200
- training: Whether the graph is in training mode.
201
-
202
- Returns:
203
- The quantized tensor.
204
- """
205
-
206
- thresholds = self.get_quantizer_variable(THRESHOLD_TENSOR)
207
- n_channels = inputs.shape[1]
208
- scale_factor = 1.0 / np.sqrt(self.max_int * n_channels)
209
- inputs_quantized = symmetric_lsq_quantizer(inputs, thresholds, self.num_bits, self.sign, self.min_int, self.max_int, scale_factor)
210
- return inputs_quantized
211
-
212
- def convert2inferable(self) -> Union[ActivationPOTInferableQuantizer, ActivationSymmetricInferableQuantizer]:
213
- """
214
- Convert quantizer to inferable quantizer.
215
-
216
- Returns:
217
- A pytorch inferable quanizer object.
218
- """
219
- threshold_values = self.get_quantizer_variable(THRESHOLD_TENSOR).cpu().detach().numpy()
220
- if self.power_of_two:
221
- pot_threshold = np.power(2.0, np.ceil(np.log2(threshold_values)))
222
- return ActivationPOTInferableQuantizer(num_bits=self.num_bits,
223
- threshold=pot_threshold.tolist(),
224
- signed=self.sign)
225
- else:
226
- return ActivationSymmetricInferableQuantizer(num_bits=self.num_bits,
227
- threshold=threshold_values.tolist(),
228
- signed=self.sign)
model_compression_toolkit/qat/pytorch/quantizer/lsq/uniform_lsq.py CHANGED
@@ -12,66 +12,32 @@
12
12
  # See the License for the specific language governing permissions and
13
13
  # limitations under the License.
14
14
  # ==============================================================================
15
- from typing import Union
16
15
  import numpy as np
17
16
  import torch
18
17
  import torch.nn as nn
19
18
 
20
- from model_compression_toolkit.constants import RANGE_MAX, RANGE_MIN
21
- from model_compression_toolkit.trainable_infrastructure.common.constants import FQ_MIN, FQ_MAX
22
-
23
- from model_compression_toolkit.qat import TrainingMethod
24
- from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
25
19
  from mct_quantizers import QuantizationTarget, PytorchQuantizationWrapper
26
- from model_compression_toolkit import constants as C
27
-
28
- from model_compression_toolkit.qat.pytorch.quantizer.base_pytorch_qat_quantizer import BasePytorchQATTrainableQuantizer
29
20
  from mct_quantizers import mark_quantizer
30
- from model_compression_toolkit.qat.pytorch.quantizer.quantizer_utils import ste_round, grad_scale
31
- from model_compression_toolkit.core.pytorch.utils import to_torch_tensor
32
21
  from mct_quantizers.pytorch.quantizers import \
33
- WeightsUniformInferableQuantizer, ActivationUniformInferableQuantizer
34
- from model_compression_toolkit.trainable_infrastructure.common.trainable_quantizer_config import \
35
- TrainableQuantizerWeightsConfig, TrainableQuantizerActivationConfig
22
+ WeightsUniformInferableQuantizer
23
+
24
+ from model_compression_toolkit.constants import RANGE_MAX, RANGE_MIN
25
+ from model_compression_toolkit.trainable_infrastructure.common.constants import FQ_MIN, FQ_MAX
26
+ from model_compression_toolkit.trainable_infrastructure import TrainingMethod
27
+ from model_compression_toolkit.trainable_infrastructure.pytorch.quantizer_utils import uniform_lsq_quantizer
36
28
  from model_compression_toolkit.trainable_infrastructure.common.base_trainable_quantizer import VariableGroup
37
- from model_compression_toolkit.qat.pytorch.quantizer.quantizer_utils import adjust_range_to_include_zero
29
+ from model_compression_toolkit.trainable_infrastructure.common.trainable_quantizer_config import \
30
+ TrainableQuantizerWeightsConfig
31
+ from model_compression_toolkit.target_platform_capabilities.target_platform import QuantizationMethod
32
+ from model_compression_toolkit.core.pytorch.utils import to_torch_tensor
38
33
  from model_compression_toolkit.core.common.quantization.quantizers.quantizers_helpers import fix_range_to_include_zero
39
-
40
-
41
-
42
- def uniform_lsq_quantizer(x: nn.Parameter,
43
- min_range: nn.Parameter,
44
- max_range: nn.Parameter,
45
- num_bits: int,
46
- min_int: int,
47
- max_int: int,
48
- scale_factor: float) -> Union[nn.Parameter, torch.Tensor]:
49
- """
50
- Uniform quantizer according to LSQ algorithm: https://arxiv.org/pdf/1902.08153.pdf
51
- Args:
52
- x: input to quantize
53
- min_range: min range of quantization values
54
- max_range: min range of quantization values
55
- num_bits: number of bits for quantization
56
- min_int: min clipping integer value
57
- max_int: max clipping integer value
58
- scale_factor: grad scale of LSQ algorithm
59
- Returns:
60
- A quantized tensor
61
- """
62
- a, b = adjust_range_to_include_zero(min_range, max_range, num_bits)
63
- delta = (b - a) / (2 ** num_bits - 1)
64
- delta_scaled = grad_scale(delta, scale_factor)
65
- rounded = ste_round((x - a) / delta_scaled)
66
- clipped = torch.clip(rounded, min=min_int, max=max_int)
67
- quantized = delta_scaled * clipped + a
68
- return quantized
34
+ from model_compression_toolkit.qat.pytorch.quantizer.base_pytorch_qat_weight_quantizer import BasePytorchQATWeightTrainableQuantizer
69
35
 
70
36
 
71
37
  @mark_quantizer(quantization_target=QuantizationTarget.Weights,
72
38
  quantization_method=[QuantizationMethod.UNIFORM],
73
39
  identifier=TrainingMethod.LSQ)
74
- class LSQUniformWeightQATQuantizer(BasePytorchQATTrainableQuantizer):
40
+ class LSQUniformWeightQATQuantizer(BasePytorchQATWeightTrainableQuantizer):
75
41
  """
76
42
  Trainable constrained quantizer to quantize layer's weights.
77
43
  """
@@ -145,79 +111,3 @@ class LSQUniformWeightQATQuantizer(BasePytorchQATTrainableQuantizer):
145
111
  max_range=max_range.tolist(),
146
112
  per_channel=self.quantization_config.weights_per_channel_threshold,
147
113
  channel_axis=self.quantization_config.weights_channels_axis)
148
-
149
-
150
- @mark_quantizer(quantization_target=QuantizationTarget.Activation,
151
- quantization_method=[QuantizationMethod.UNIFORM],
152
- identifier=TrainingMethod.LSQ)
153
- class LSQUniformActivationQATQuantizer(BasePytorchQATTrainableQuantizer):
154
- """
155
- Trainable constrained quantizer to quantize layer activations.
156
- """
157
-
158
- def __init__(self, quantization_config: TrainableQuantizerActivationConfig):
159
- """
160
- Initialize a LSQUniformActivationQATQuantizer object with parameters to use
161
- for uniform quantization.
162
-
163
- Args:
164
- quantization_config: trainable quantizer config class
165
- """
166
- super().__init__(quantization_config)
167
- self.num_bits = self.quantization_config.activation_n_bits
168
- self.min_int = 0
169
- self.max_int = 2 ** self.num_bits - 1
170
- self.min_range = np.array([quantization_config.activation_quantization_params[C.RANGE_MIN]])
171
- self.max_range = np.array([quantization_config.activation_quantization_params[C.RANGE_MAX]])
172
-
173
- def initialize_quantization(self,
174
- tensor_shape: torch.Size,
175
- name: str,
176
- layer: PytorchQuantizationWrapper):
177
- """
178
- Add quantizer parameters to the quantizer parameters dictionary
179
-
180
- Args:
181
- tensor_shape: tensor shape of the quantized tensor.
182
- name: Tensor name.
183
- layer: Layer to quantize.
184
- """
185
- layer.register_parameter(name+"_"+FQ_MIN, nn.Parameter(to_torch_tensor(self.min_range), requires_grad=True))
186
- layer.register_parameter(name+"_"+FQ_MAX, nn.Parameter(to_torch_tensor(self.max_range), requires_grad=True))
187
-
188
- # Save the quantizer parameters for later calculations
189
- self.add_quantizer_variable(FQ_MIN, layer.get_parameter(name+"_"+FQ_MIN), VariableGroup.QPARAMS)
190
- self.add_quantizer_variable(FQ_MAX, layer.get_parameter(name+"_"+FQ_MAX), VariableGroup.QPARAMS)
191
-
192
- def __call__(self,
193
- inputs: torch.Tensor,
194
- training: bool = True) -> torch.Tensor:
195
- """
196
- Quantize a tensor.
197
- Args:
198
- inputs: Input tensor to quantize.
199
- training: Whether the graph is in training mode.
200
-
201
- Returns:
202
- The quantized tensor.
203
- """
204
- min_range = self.get_quantizer_variable(FQ_MIN)
205
- max_range = self.get_quantizer_variable(FQ_MAX)
206
- n_channels = inputs.shape[1]
207
- scale_factor = 1.0 / np.sqrt(self.max_int * n_channels)
208
- inputs_quantized = uniform_lsq_quantizer(inputs, min_range, max_range, self.num_bits, self.min_int, self.max_int, scale_factor)
209
- return inputs_quantized
210
-
211
- def convert2inferable(self) -> ActivationUniformInferableQuantizer:
212
- """
213
- Convert quantizer to inferable quantizer.
214
-
215
- Returns:
216
- A pytorch inferable quanizer object.
217
- """
218
- min_range = self.get_quantizer_variable(FQ_MIN).cpu().detach().numpy()
219
- max_range = self.get_quantizer_variable(FQ_MAX).cpu().detach().numpy()
220
- min_range, max_range = fix_range_to_include_zero(min_range, max_range, self.num_bits)
221
- return ActivationUniformInferableQuantizer(num_bits=self.num_bits,
222
- min_range=min_range.tolist(),
223
- max_range=max_range.tolist())