Piotr KrzemińskiandMaxim Vafin
e076ed4726
* Support GetAttr with packed params * Apply suggestions from code review * [PT FE] Add quantized types as normal types to decoder * [PT FE] Add decoder dequantize, add dtypes to quantize * [PT FE] Add dequantize example * [PT FE] Implement replacer for quantized nodes * [PT FE] Register replacer for quantize/dequantize * [PT FE] Remove unwanted junk from previous version * [PT FE] Fix building mistakes for frontend * [PT FE] Clang fix * [PT FE] Ease of use upgrade to quantize funcs * [PT FE] Clang format * [PT FE] Introduce new version of quantize/dequantize * [PT FE] Remove unwanted files from new version * [PT FE] Fix style * [PT FE] Add QuantizedPtNode replacer, fix accuracy error * [PT FE] Quantized add * [PT FE] Add improved version of quantize/dequantize with shared_ptrs * [PT FE] Quantized Add Relu * [PT FE] Fix utils shared ptr reference error * [PT FE] Quantized Hardswish & tests * [PT FE] Add quantized_add_relu test * [PT FE] Quantize now takes correct input for operations * [PT FE] Add dtype to tests * [PT FE] Increase matrix size in the tests to cover more edge cases * [PT FE] Upgrade quantize method * [PT FE] Add BFS for dequantize, add quantize_per_channel * [PT FE] Add missing replacer to frontend, improve tests * [PT FE] Rename replacer -> remover, remove unwanted header files * [PT FE] Change function declarations to return ov::Output instead of shared ptr * [PT FE] Add missing context mark node * [PT FE] Remove unknown modifications to ie_c_api * [PT FE] Remove fp16 support, turn off int32 tests * [PT FE] Clang format * [PT FE] Fix quantize_per_tensor * [PT FE] Minor fixes from review * [PT FE] Remove dequantize, remove helpers, replacer now removes nodes instead * [PT FE] Rename Replacer to Remover for dequantize nodes * [PT FE] Clang format * [PT FE] Move comments to header files, minor import fixes * [PT FE] Enable add, add_relu, mul, hardswish with newest quantize ops * [PT FE] Fix clang format * [PT FE] Fix dtype issue * [PT FE] Reenable hardswish tests * [PT FE] Fix building error for quantized ops * [PT FE] Tests now use rand instead of randn * [PT FE] Remove int32 support from tests * [PT FE] Fix quantize_per_channel tests * Apply suggestions from code review Removing sporadic tests from precommit * Apply suggestions from code review * [PT FE] Mark qint32 with skip, remove precommit flag due to sporadic errors * [PT FE] Introduce quantized test for PytorchTestClass * [PT FE] Add missing tensor size in the errors count computation * Update pytorch_layer_test_class.py * [PT FE] Apply suggestions from review - remove contexts, log100 for errors, add precommit flags * Apply suggestions from code review * [PT FE] Fix quantize per channel test after merge * Update tests/layer_tests/pytorch_tests/pytorch_layer_test_class.py * Update pytorch_layer_test_class.py * Update test_quantize.py * Apply suggestions from code review * [PT FE] Fix quantized mul by changing default quantization dtype * Update pytorch_layer_test_class.py * Update tests/layer_tests/pytorch_tests/pytorch_layer_test_class.py * Update src/frontends/pytorch/src/utils_quantize.cpp * Update src/frontends/pytorch/src/utils_quantize.cpp * Update tests/layer_tests/pytorch_tests/test_quantized_mul.py --------- Co-authored-by: Maxim Vafin <maxim.vafin@intel.com>
Layer tests
This folder layer tests framework code and test files.
Getting Started
Pre-requisites
- OpenVINO should be configured as usual.
Setup
- Install requirements:
pip3 install -r requirements.txt - Set up environment variables for layer tests (if you use wheel package path to python api could be removed):
export PYTHONPATH="path_to_openvino"/tests/layer_tests/:"path_to_openvino"/tools/mo:"path to python api" - If there is need to use specific libs it is possible to specify path to them using OV_LIBRARY_PATH env variable
export OV_LIBRARY_PATH="path_to_libs" - To parametrize tests by device and precision (optional)
export TEST_DEVICE="CPU;GPU" export TEST_PRECISION="FP32;FP16"
Run tests
py.test