You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
277 lines
14 KiB
277 lines
14 KiB
file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
|
|
string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
|
|
|
|
function(_inference_analysis_python_api_int8_test target model_dir data_path filename use_mkldnn)
|
|
py_test(${target} SRCS ${filename}
|
|
ENVS CPU_NUM_THREADS=${CPU_NUM_THREADS_ON_CI}
|
|
FLAGS_use_mkldnn=${use_mkldnn}
|
|
ARGS --infer_model ${model_dir}/model
|
|
--infer_data ${data_path}
|
|
--int8_model_save_path int8_models/${target}
|
|
--warmup_batch_size ${WARMUP_BATCH_SIZE}
|
|
--batch_size 50)
|
|
endfunction()
|
|
|
|
function(inference_analysis_python_api_int8_test target model_dir data_path filename)
|
|
_inference_analysis_python_api_int8_test(${target} ${model_dir} ${data_path} ${filename} False)
|
|
endfunction()
|
|
|
|
function(inference_analysis_python_api_int8_test_custom_warmup_batch_size target model_dir data_dir filename warmup_batch_size)
|
|
set(WARMUP_BATCH_SIZE ${warmup_batch_size})
|
|
inference_analysis_python_api_int8_test(${target} ${model_dir} ${data_dir} ${filename})
|
|
endfunction()
|
|
|
|
function(inference_analysis_python_api_int8_test_mkldnn target model_dir data_path filename)
|
|
_inference_analysis_python_api_int8_test(${target} ${model_dir} ${data_path} ${filename} True)
|
|
endfunction()
|
|
|
|
function(download_quant_data install_dir data_file)
|
|
if (NOT EXISTS ${install_dir}/${data_file})
|
|
inference_download_and_uncompress(${install_dir} ${INFERENCE_URL}/int8 ${data_file})
|
|
endif()
|
|
endfunction()
|
|
|
|
function(download_quant_model install_dir data_file)
|
|
if (NOT EXISTS ${install_dir}/${data_file})
|
|
inference_download_and_uncompress(${install_dir} ${INFERENCE_URL}/int8/QAT_models ${data_file})
|
|
endif()
|
|
endfunction()
|
|
|
|
function(download_quant_fp32_model install_dir data_file)
|
|
if (NOT EXISTS ${install_dir}/${data_file})
|
|
inference_download_and_uncompress(${install_dir} ${INFERENCE_URL}/int8/QAT_models/fp32 ${data_file})
|
|
endif()
|
|
endfunction()
|
|
|
|
function(inference_quant_int8_image_classification_test target quant_model_dir dataset_path)
|
|
py_test(${target} SRCS "${CMAKE_CURRENT_SOURCE_DIR}/quant_int8_image_classification_comparison.py"
|
|
ENVS FLAGS_OMP_NUM_THREADS=${CPU_NUM_THREADS_ON_CI}
|
|
OMP_NUM_THREADS=${CPU_NUM_THREADS_ON_CI}
|
|
FLAGS_use_mkldnn=true
|
|
ARGS --quant_model ${quant_model_dir}
|
|
--infer_data ${dataset_path}
|
|
--batch_size 25
|
|
--batch_num 2
|
|
--acc_diff_threshold 0.1)
|
|
endfunction()
|
|
|
|
|
|
# set batch_size 10 for UT only (avoid OOM). For whole dataset, use batch_size 25
|
|
function(inference_quant2_int8_image_classification_test target quant_model_dir fp32_model_dir dataset_path)
|
|
py_test(${target} SRCS "${CMAKE_CURRENT_SOURCE_DIR}/quant2_int8_image_classification_comparison.py"
|
|
ENVS FLAGS_OMP_NUM_THREADS=${CPU_NUM_THREADS_ON_CI}
|
|
OMP_NUM_THREADS=${CPU_NUM_THREADS_ON_CI}
|
|
FLAGS_use_mkldnn=true
|
|
ARGS --quant_model ${quant_model_dir}
|
|
--fp32_model ${fp32_model_dir}
|
|
--infer_data ${dataset_path}
|
|
--batch_size 10
|
|
--batch_num 2
|
|
--acc_diff_threshold 0.1)
|
|
endfunction()
|
|
|
|
# set batch_size 10 for UT only (avoid OOM). For whole dataset, use batch_size 20
|
|
function(inference_quant2_int8_nlp_test target quant_model_dir fp32_model_dir dataset_path labels_path ops_to_quantize)
|
|
py_test(${target} SRCS "${CMAKE_CURRENT_SOURCE_DIR}/quant2_int8_nlp_comparison.py"
|
|
ENVS FLAGS_OMP_NUM_THREADS=${CPU_NUM_THREADS_ON_CI}
|
|
OMP_NUM_THREADS=${CPU_NUM_THREADS_ON_CI}
|
|
FLAGS_use_mkldnn=true
|
|
ARGS --quant_model ${quant_model_dir}
|
|
--fp32_model ${fp32_model_dir}
|
|
--infer_data ${dataset_path}
|
|
--labels ${labels_path}
|
|
--batch_size 10
|
|
--batch_num 2
|
|
--acc_diff_threshold 0.1
|
|
--ops_to_quantize ${ops_to_quantize})
|
|
endfunction()
|
|
|
|
function(download_quant_data install_dir data_file)
|
|
if (NOT EXISTS ${install_dir}/${data_file})
|
|
inference_download_and_uncompress(${install_dir} ${INFERENCE_URL}/int8 ${data_file})
|
|
endif()
|
|
endfunction()
|
|
|
|
function(download_quant_model install_dir data_file)
|
|
if (NOT EXISTS ${install_dir}/${data_file})
|
|
inference_download_and_uncompress(${install_dir} ${INFERENCE_URL}/int8/QAT_models ${data_file})
|
|
endif()
|
|
endfunction()
|
|
|
|
function(save_quant_ic_model_test target quant_model_dir fp32_model_save_path int8_model_save_path)
|
|
py_test(${target} SRCS ${CMAKE_CURRENT_SOURCE_DIR}/save_quant_model.py
|
|
ARGS --quant_model_path ${quant_model_dir}
|
|
--fp32_model_save_path ${fp32_model_save_path}
|
|
--int8_model_save_path ${int8_model_save_path}
|
|
--debug)
|
|
endfunction()
|
|
|
|
function(save_quant_nlp_model_test target quant_model_dir fp32_model_save_path int8_model_save_path ops_to_quantize)
|
|
py_test(${target} SRCS ${CMAKE_CURRENT_SOURCE_DIR}/save_quant_model.py
|
|
ARGS --quant_model_path ${quant_model_dir}
|
|
--fp32_model_save_path ${fp32_model_save_path}
|
|
--int8_model_save_path ${int8_model_save_path}
|
|
--ops_to_quantize ${ops_to_quantize})
|
|
endfunction()
|
|
|
|
function(convert_model2dot_test target model_path save_graph_dir save_graph_name)
|
|
py_test(${target} SRCS ${CMAKE_CURRENT_SOURCE_DIR}/convert_model2dot.py
|
|
ARGS --model_path ${model_path}
|
|
--save_graph_dir ${save_graph_dir}
|
|
--save_graph_name ${save_graph_name})
|
|
endfunction()
|
|
|
|
if(WIN32)
|
|
list(REMOVE_ITEM TEST_OPS test_light_nas)
|
|
list(REMOVE_ITEM TEST_OPS test_post_training_quantization_mnist)
|
|
list(REMOVE_ITEM TEST_OPS test_post_training_quantization_mobilenetv1)
|
|
list(REMOVE_ITEM TEST_OPS test_post_training_quantization_resnet50)
|
|
list(REMOVE_ITEM TEST_OPS test_weight_quantization_mobilenetv1)
|
|
endif()
|
|
|
|
if(LINUX AND WITH_MKLDNN)
|
|
|
|
#### Image classification dataset: ImageNet (small)
|
|
# The dataset should already be downloaded for INT8v2 unit tests
|
|
set(IMAGENET_DATA_PATH "${INFERENCE_DEMO_INSTALL_DIR}/imagenet/data.bin")
|
|
|
|
#### INT8 image classification python api test
|
|
# Models should be already downloaded for INT8v2 unit tests
|
|
|
|
set(INT8_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/int8v2")
|
|
|
|
#### QUANT & INT8 comparison python api tests
|
|
|
|
set(QUANT_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/quant")
|
|
|
|
### Quant1 for image classification
|
|
|
|
# Quant ResNet50
|
|
set(QUANT_RESNET50_MODEL_DIR "${QUANT_INSTALL_DIR}/ResNet50_quant")
|
|
set(QUANT_RESNET50_MODEL_ARCHIVE "ResNet50_qat_model.tar.gz")
|
|
download_quant_model(${QUANT_RESNET50_MODEL_DIR} ${QUANT_RESNET50_MODEL_ARCHIVE})
|
|
inference_quant_int8_image_classification_test(test_quant_int8_resnet50_mkldnn ${QUANT_RESNET50_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
# Quant ResNet101
|
|
set(QUANT_RESNET101_MODEL_DIR "${QUANT_INSTALL_DIR}/ResNet101_quant")
|
|
set(QUANT_RESNET101_MODEL_ARCHIVE "ResNet101_qat_model.tar.gz")
|
|
download_quant_model(${QUANT_RESNET101_MODEL_DIR} ${QUANT_RESNET101_MODEL_ARCHIVE})
|
|
# inference_quant_int8_image_classification_test(test_quant_int8_resnet101_mkldnn ${QUANT_RESNET101_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
# Quant GoogleNet
|
|
set(QUANT_GOOGLENET_MODEL_DIR "${QUANT_INSTALL_DIR}/GoogleNet_quant")
|
|
set(QUANT_GOOGLENET_MODEL_ARCHIVE "GoogleNet_qat_model.tar.gz")
|
|
download_quant_model(${QUANT_GOOGLENET_MODEL_DIR} ${QUANT_GOOGLENET_MODEL_ARCHIVE})
|
|
inference_quant_int8_image_classification_test(test_quant_int8_googlenet_mkldnn ${QUANT_GOOGLENET_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
# Quant MobileNetV1
|
|
set(QUANT_MOBILENETV1_MODEL_DIR "${QUANT_INSTALL_DIR}/MobileNetV1_quant")
|
|
set(QUANT_MOBILENETV1_MODEL_ARCHIVE "MobileNetV1_qat_model.tar.gz")
|
|
download_quant_model(${QUANT_MOBILENETV1_MODEL_DIR} ${QUANT_MOBILENETV1_MODEL_ARCHIVE})
|
|
inference_quant_int8_image_classification_test(test_quant_int8_mobilenetv1_mkldnn ${QUANT_MOBILENETV1_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
# Quant MobileNetV2
|
|
set(QUANT_MOBILENETV2_MODEL_DIR "${QUANT_INSTALL_DIR}/MobileNetV2_quant")
|
|
set(QUANT_MOBILENETV2_MODEL_ARCHIVE "MobileNetV2_qat_model.tar.gz")
|
|
download_quant_model(${QUANT_MOBILENETV2_MODEL_DIR} ${QUANT_MOBILENETV2_MODEL_ARCHIVE})
|
|
inference_quant_int8_image_classification_test(test_quant_int8_mobilenetv2_mkldnn ${QUANT_MOBILENETV2_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
# Quant VGG16
|
|
set(QUANT_VGG16_MODEL_DIR "${QUANT_INSTALL_DIR}/VGG16_quant")
|
|
set(QUANT_VGG16_MODEL_ARCHIVE "VGG16_qat_model.tar.gz")
|
|
download_quant_model(${QUANT_VGG16_MODEL_DIR} ${QUANT_VGG16_MODEL_ARCHIVE})
|
|
# inference_quant_int8_image_classification_test(test_quant_int8_vgg16_mkldnn ${QUANT_VGG16_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
# Quant VGG19
|
|
set(QUANT_VGG19_MODEL_DIR "${QUANT_INSTALL_DIR}/VGG19_quant")
|
|
set(QUANT_VGG19_MODEL_ARCHIVE "VGG19_qat_model.tar.gz")
|
|
download_quant_model(${QUANT_VGG19_MODEL_DIR} ${QUANT_VGG19_MODEL_ARCHIVE})
|
|
# inference_quant_int8_image_classification_test(test_quant_int8_vgg19_mkldnn ${QUANT_VGG19_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
### Quant2 for image classification
|
|
|
|
# Quant2 ResNet50 with input/output scales in `fake_quantize_moving_average_abs_max` operators,
|
|
# with weight scales in `fake_dequantize_max_abs` operators
|
|
set(QUANT2_RESNET50_MODEL_DIR "${QUANT_INSTALL_DIR}/ResNet50_quant2")
|
|
set(QUANT2_RESNET50_MODEL_ARCHIVE "ResNet50_qat_perf.tar.gz")
|
|
download_quant_model(${QUANT2_RESNET50_MODEL_DIR} ${QUANT2_RESNET50_MODEL_ARCHIVE})
|
|
set(FP32_RESNET50_MODEL_DIR "${INT8_INSTALL_DIR}/resnet50")
|
|
inference_quant2_int8_image_classification_test(test_quant2_int8_resnet50_mkldnn ${QUANT2_RESNET50_MODEL_DIR}/ResNet50_qat_perf/float ${FP32_RESNET50_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
# Quant2 ResNet50 with input/output scales in `fake_quantize_range_abs_max` operators and the `out_threshold` attributes,
|
|
# with weight scales in `fake_dequantize_max_abs` operators
|
|
set(QUANT2_RESNET50_RANGE_MODEL_DIR "${QUANT_INSTALL_DIR}/ResNet50_quant2_range")
|
|
set(QUANT2_RESNET50_RANGE_MODEL_ARCHIVE "ResNet50_qat_range.tar.gz")
|
|
download_quant_model(${QUANT2_RESNET50_RANGE_MODEL_DIR} ${QUANT2_RESNET50_RANGE_MODEL_ARCHIVE})
|
|
inference_quant2_int8_image_classification_test(test_quant2_int8_resnet50_range_mkldnn ${QUANT2_RESNET50_RANGE_MODEL_DIR}/ResNet50_qat_range ${FP32_RESNET50_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
# Quant2 ResNet50 with input/output scales in `fake_quantize_range_abs_max` operators and the `out_threshold` attributes,
|
|
# with weight scales in `fake_channel_wise_dequantize_max_abs` operators
|
|
set(QUANT2_RESNET50_CHANNELWISE_MODEL_DIR "${QUANT_INSTALL_DIR}/ResNet50_quant2_channelwise")
|
|
set(QUANT2_RESNET50_CHANNELWISE_MODEL_ARCHIVE "ResNet50_qat_channelwise.tar.gz")
|
|
download_quant_model(${QUANT2_RESNET50_CHANNELWISE_MODEL_DIR} ${QUANT2_RESNET50_CHANNELWISE_MODEL_ARCHIVE})
|
|
inference_quant2_int8_image_classification_test(test_quant2_int8_resnet50_channelwise_mkldnn ${QUANT2_RESNET50_CHANNELWISE_MODEL_DIR}/ResNet50_qat_channelwise ${FP32_RESNET50_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
# Quant2 MobileNetV1
|
|
set(QUANT2_MOBILENETV1_MODEL_DIR "${QUANT_INSTALL_DIR}/MobileNetV1_quant2")
|
|
set(QUANT2_MOBILENETV1_MODEL_ARCHIVE "MobileNet_qat_perf.tar.gz")
|
|
download_quant_model(${QUANT2_MOBILENETV1_MODEL_DIR} ${QUANT2_MOBILENETV1_MODEL_ARCHIVE})
|
|
set(FP32_MOBILENETV1_MODEL_DIR "${INT8_INSTALL_DIR}/mobilenetv1")
|
|
inference_quant2_int8_image_classification_test(test_quant2_int8_mobilenetv1_mkldnn ${QUANT2_MOBILENETV1_MODEL_DIR}/MobileNet_qat_perf/float ${FP32_MOBILENETV1_MODEL_DIR}/model ${IMAGENET_DATA_PATH})
|
|
|
|
### Quant2 for NLP
|
|
|
|
set(NLP_DATA_ARCHIVE "Ernie_dataset.tar.gz")
|
|
set(NLP_DATA_DIR "${INFERENCE_DEMO_INSTALL_DIR}/Ernie_dataset")
|
|
set(NLP_DATA_PATH "${NLP_DATA_DIR}/Ernie_dataset/1.8w.bs1")
|
|
set(NLP_LABLES_PATH "${NLP_DATA_DIR}/Ernie_dataset/label.xnli.dev")
|
|
download_quant_data(${NLP_DATA_DIR} ${NLP_DATA_ARCHIVE})
|
|
|
|
set(QUANT2_NLP_OPS_TO_QUANTIZE "fc,reshape2,transpose2,matmul,elementwise_add")
|
|
|
|
# Quant2 Ernie
|
|
set(QUANT2_ERNIE_MODEL_ARCHIVE "ernie_qat.tar.gz")
|
|
set(QUANT2_ERNIE_MODEL_DIR "${QUANT_INSTALL_DIR}/Ernie_quant2")
|
|
download_quant_model(${QUANT2_ERNIE_MODEL_DIR} ${QUANT2_ERNIE_MODEL_ARCHIVE})
|
|
set(FP32_ERNIE_MODEL_ARCHIVE "ernie_fp32_model.tar.gz")
|
|
set(FP32_ERNIE_MODEL_DIR "${QUANT_INSTALL_DIR}/Ernie_float")
|
|
download_quant_fp32_model(${FP32_ERNIE_MODEL_DIR} ${FP32_ERNIE_MODEL_ARCHIVE})
|
|
inference_quant2_int8_nlp_test(test_quant2_int8_ernie_mkldnn ${QUANT2_ERNIE_MODEL_DIR}/Ernie_qat/float ${FP32_ERNIE_MODEL_DIR}/ernie_fp32_model ${NLP_DATA_PATH} ${NLP_LABLES_PATH} ${QUANT2_NLP_OPS_TO_QUANTIZE})
|
|
|
|
### Save FP32 model or INT8 model from Quant model
|
|
|
|
set(QUANT2_INT8_RESNET50_SAVE_PATH "${QUANT_INSTALL_DIR}/ResNet50_quant2_int8")
|
|
set(QUANT2_FP32_RESNET50_SAVE_PATH "${QUANT_INSTALL_DIR}/ResNet50_quant2_fp32")
|
|
save_quant_ic_model_test(save_quant2_model_resnet50 ${QUANT2_RESNET50_MODEL_DIR}/ResNet50_qat_perf/float ${QUANT2_FP32_RESNET50_SAVE_PATH} ${QUANT2_INT8_RESNET50_SAVE_PATH})
|
|
|
|
set(QUANT2_INT8_ERNIE_SAVE_PATH "${QUANT_INSTALL_DIR}/Ernie_quant2_int8")
|
|
set(QUANT2_FP32_ERNIE_SAVE_PATH "${QUANT_INSTALL_DIR}/Ernie_quant2_fp32")
|
|
save_quant_nlp_model_test(save_quant2_model_ernie ${QUANT2_ERNIE_MODEL_DIR}/Ernie_qat/float ${QUANT2_FP32_ERNIE_SAVE_PATH} ${QUANT2_INT8_ERNIE_SAVE_PATH} ${QUANT2_NLP_OPS_TO_QUANTIZE})
|
|
|
|
# Convert Quant2 model to dot and pdf files
|
|
set(QUANT2_INT8_ERNIE_DOT_SAVE_PATH "${QUANT_INSTALL_DIR}/Ernie_quant2_int8_dot_file")
|
|
convert_model2dot_test(convert_model2dot_ernie ${QUANT2_ERNIE_MODEL_DIR}/Ernie_qat/float ${QUANT2_INT8_ERNIE_DOT_SAVE_PATH} "Ernie_quant2_int8")
|
|
|
|
endif()
|
|
|
|
# Since the tests for Quant & INT8 comparison support only testing on Linux
|
|
# with MKL-DNN, we remove it here to not test it on other systems.
|
|
list(REMOVE_ITEM TEST_OPS
|
|
test_mkldnn_int8_quantization_strategy
|
|
quant_int8_image_classification_comparison
|
|
quant_int8_nlp_comparison)
|
|
|
|
#TODO(wanghaoshuang): Fix this unitest failed on GCC8.
|
|
LIST(REMOVE_ITEM TEST_OPS test_auto_pruning)
|
|
LIST(REMOVE_ITEM TEST_OPS test_filter_pruning)
|
|
|
|
foreach(src ${TEST_OPS})
|
|
py_test(${src} SRCS ${src}.py)
|
|
endforeach()
|
|
|
|
# setting timeout value for old unittests
|
|
if(NOT WIN32)
|
|
set_tests_properties(test_post_training_quantization_mobilenetv1 PROPERTIES TIMEOUT 250)
|
|
set_tests_properties(test_post_training_quantization_resnet50 PROPERTIES TIMEOUT 200)
|
|
endif()
|