v0.1.0

What's Changed

• Internal clean up. by @copybara-service in #1
• Fix oss unit test in Python 3.9 for calibrator.py. by @copybara-service in #3
• Relax the compatibility check for tensor transformation params. by @copybara-service in #2
• Fix Python3.9 error in calibrator.py by @copybara-service in #5
• Fix comments and docstring formatting in qtyping.py by @copybara-service in #6
• Fix comments and docstring formatting in naive_min_max_quantize.py by @copybara-service in #7
• Add depthwise conv2d support. by @copybara-service in #10
• Add build_pypi.yml in GitHub action workflow. by @copybara-service in #17
• Move nested functions outside of materialize_standard_op by @copybara-service in #12
• Fix a bug in recipe manager. by @copybara-service in #16
• Transform TODOs style by @copybara-service in #14
• Add transpose support by @copybara-service in #19
• Internal test support. by @copybara-service in #22
• Enable int4 DRQ for EmbeddingLookup by @copybara-service in #24
• Fix order in the list returned by materialize_standard_op by @copybara-service in #25
• Support ADD op in quantization tool. by @copybara-service in #20
• Refactor materialize_softmax to materialize_op_with_output_activation_constraint. by @copybara-service in #23
• Add tanh support by @copybara-service in #26
• Trigger github CI to run on pull request. by @copybara-service in #27
• Fix a bug in CONV2D tests. by @copybara-service in #18
• Fix comments and docstring formatting in params_generator.py by @copybara-service in #30
• Switch order of float and quantized for comparison function. by @copybara-service in #29
• No public description by @copybara-service in #34
• Add a skip_checks option to OpQuantizationConfig. by @copybara-service in #32
• Support SUB op in quantization tool. by @copybara-service in #21
• Add BMM op to the SRQ list. by @copybara-service in #38
• Remove full-integer requirement for calibrator by @copybara-service in #35
• Check parameters for compatibility only if they exists by @copybara-service in #36
• Support RSQRT op in quantization tool by @copybara-service in #41
• Format the JSON recipes. by @copybara-service in #46
• Support STRIDED_SLICE op in quantization tool by @copybara-service in #43
• Add split op support by @copybara-service in #44
• Fix a bug in ParamsGenerator. by @copybara-service in #47
• Add i4 EmbeddingLookup to the supported DRQ op list by @copybara-service in #53
• Prevent in-place modification of signature_input_data by @copybara-service in #49
• Fixed a bug to allow input tensors to be quantized for full integer models. by @copybara-service in #51
• Refactor materialize_fc_conv to separate bias materialization. by @copybara-service in #57
• Add conv2d_transpose support. by @copybara-service in #54
• Add conv2d_transpose support. by @copybara-service in #59
• Add a sample default policy for configuration checks, along with the implementation to load it. by @copybara-service in #60
• Fix a bug in transformation_instruction_generator.py by @copybara-service in #61
• Set the max precision of quantization scale when quantize a tensor (1e-6 for int8 and 1e-8 for int16). by @copybara-service in #62
• Expand default config check policy with DRQ op checking and replace hardcoded logic. by @copybara-service in #66
• add frontend support for emulated subchannel by @copybara-service in #56
• Refactor, organize, and add comments to MNIST by @copybara-service in #65
• Add testing functions for ignored inputs/outputs by @copybara-service in #69
• Expand default config check policy with SRQ op checking and replace hardcoded logic. by @copybara-service in #67
• Expand default config check policy with Weight-Only op checking and replace hardcoded logic. by @copybara-service in #68
• Check if input model is already quantized by @copybara-service in #70
• Fix how ignored inputs/outputs are processed by @copybara-service in #63
• Add utils to get input/output/constant tensor names from the model. by @copybara-service in #71
• Fix the default_af32w4float_recipe.json. by @copybara-service in #80
• Internal clean ups. by @copybara-service in #82
• Add ComparisonResult class and process_comparison_result function. by @copybara-service in #72
• Remove the nested folder structure when save the quantized model. by @copybara-service in #83
• Fix for model_validator to check dtype of tensor. by @copybara-service in #84
• Save ComparisonResult object to be used as the input for model explorer. by @copybara-service in #79
• Add conftest.py to fix pytest error. by @copybara-service in #87
• Add some blockwise quantization to default policy. by @copybara-service in #77
• Allow Calibrator to accept custom qsv_update_func. by @copybara-service in #86
• Remove un-needed files. by @copybara-service in #89
• add sample test image for isnet example by @copybara-service in #90
• Remove temporary fix from ParamsGenerator by @copybara-service in #91
• Add a function to update the qsv with minimum min values and maximum max values. by @copybara-service in #88
• skip non-native numpy tensor type when validating by @copybara-service in #94
• Fix a bug for calibration process. by @copybara-service in #92
• Add an option to use reference kernel for numerical validation. by @copybara-service in #96
• Rename default_af32w8int_recipe.json to dynamic_wi8_afp32_recipe.json by @copybara-service in #98
• fix shape issue for emulated subchannel quantization by @copybara-service in #101
• Remove suffix _quantized in the result tflite file. by @copybara-service in #103
• Add tensor name checks in params generator. by @copybara-service in #100
• No allocate_tensor for _check_is_float_model by @copybara-service in #105
• Removes the interpreter dependency from quantizer initialization process. by @copybara-service in #109
• Add input/output operators to the subgraph. by @copybara-service in #102
• Ignore Lookup index when quantizing embedding_lookup op. by @copybara-service in #110
• Add getting_started colab to OSS (dynamic quantization for MNIST model) by @copybara-service in #111
• Improve the op coverage documentation by @copybara-service in #113
• Remove obsolete check_weight_only_config, check_drq_config, check_srq_config functions by @copybara-service in #97
• Elaborate getting_started colab by @copybara-service in #116
• Add end to end test for input output op quantization. by @copybara-service in #114
• Add colab with torch conversion and quantization by @copybara-service in #118
• Move Colabs and documentation to top-level folder by @copybara-service in #117
• Add test_data folder for selective_quantization_isnet colab by @copybara-service in #119
• Remove weight-only quantization from selective_quantization_isnet colab by @copybara-service in #122
• Internal document changes. by @copybara-service in #123
• Add FP16 tfl.embedding_lookup support by @copybara-service in #121
• Small cleanup of the torch Colab. by @copybara-service in #125
• Add export_model method to the quantization by @copybara-service in #124
• Utilize export_model method in Colabs by @copybara-service in #126
• Some cleanup of the getting_started Colab. by @copybara-service in #128
• Improvements to the selective quantization Colab by @copybara-service in #129
• Add links to open in Colab and view source on GitHub. by @copybara-service in #132
• Link to the Colabs from the README. by @copybara-service in #130
• Additional cleanup of the docs by @copybara-service in #131
• Some more edits to the top-level README. by @copybara-service in #134
• Small cleanup of docs by @copybara-service in #136
• Add Logistic OP to AI Edge Quantizer. by @copybara-service in #133
• Reduce memory usage in Quantizer by @copybara-service in #138
• Extend interpreter utils to support multi-signature models. by @copybara-service in #137
• Add LOGISTIC to the list of supported ops in README. by @copybara-service in #142
• Add getting started colab to nightly testing by @copybara-service in #141
• Fix nightly testing workflow for getting started colab by @copybara-service in #145
• Add the quantized two signature model for testing purposes. by @copybara-service in #143
• Refactor ComparisonResult to support multiple signatures. by @copybara-service in #144
• Fix the error in the selective quantization colab. by @copybara-service in #149
• Add model size reduction information to ComparisonResult. by @copybara-service in #150
• Add weight_only support for bmm op. by @copybara-service in #156
• Add EMBEDDING_LOOKUP to weightonly_wi4_afp32 policy. by @copybara-service in #157
• Add INT4 Conv to supported policy by @copybara-service in #154
• Reduce quantization time for large models by @copybara-service in #159
• Implement multi-signature calibration by @copybara-service in #160
• Check if calibration result has sufficient QSV before quantization. by @copybara-service in #161
• Add recipe for dynamic legacy compatibility by @copybara-service in #163
• Add dynamic_legacy_wi8_afp32 recipe to recipe module by @copybara-service in #164
• Add static quantization for EMBEDDING_LOOKUP in Quantizer by @copybara-service in #169
• Add full integer quantization for SELECT_V2 in Quantizer by @copybara-service in #171
• Update README with newly added ops. by @copybara-service in #172
• Enable packing only for the 4 bits case. by @copybara-service in #173
• Enable calibration and model validation with XNNPACK by @copybara-service in #175
• Fix selective quantization Colab by @copybara-service in #176
• Enable quantization when op may contain tensors with zero-size array by @copybara-service in #178
• add quantize tensor with Blockwise data into transformation by @copybara-service in #177
• Refactor min/max quantize files by @copybara-service in #179
• front end change for native blockwise quantization by @copybara-service in #180
• Support py3.12 for ai_edge_quantizer. by @copybara-service in #187
• Support py3.12 for ai_edge_quantizer. by @copybara-service in #190
• Register dequantized weight recovery algorithm for FullyConnected and EmbeddingLookup. by @copybara-service in #188
• Internal test cleanups. by @copybara-service in #185
• Add full integer quantization for DYNAMIC_UPDATE_SLICE in Quantizer by @copybara-service in #191
• Internal test cleanups. by @copybara-service in #189

Full Changelog: https://github.com/google-ai-edge/ai-edge-quantizer/commits/v0.1.0