Created by: chuanqi129

1. Build Paddle

Requirements: cmake >= 3.0, python protobuf >= 3.0, patchelf

        git clone  https://github.com/PaddlePaddle/Paddle 
	mkdir build
	cd build
	cmake  .. -DCMAKE_INSTALL_PREFIX=./tmp -DWITH_GPU=OFF -DWITH_MKLDNN=ON -DWITH_TESTING=ON -DWITH_PROFILER=ON -DWITH_MKL=ON -DWITH_INFERENCE_API_TEST=ON -DCMAKE_BUILD_TYPE=Release -DON_INFER=ON -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
  	make -j
	make install -j
	make -j fluid_lib_dist
	pip install --force-reinstall --user ./tmp/opt/paddle/share/wheels/paddlepaddle-*.whl
        or
        pip install paddlepaddle==1.3 --user

2. Run test_calibration.py get resnet50 fp32 and int8 model

        cd $Paddle_home
	FLAGS_use_mkldnn=true python python/paddle/fluid/contrib/tests/test_calibration.py

Then, you will get resnet50 fp32 pretrained model ~/.cache/paddle/dataset/int8/download/resnet50_fp32/model/ and resnet50 int8 model $Paddle_home/calibration_out/

3. Prune and save model

import paddle
import paddle.fluid as fluid
import paddle.fluid.core as core
import argparse


def parse_args():
    """
    Parse input arguments
    """
    parser = argparse.ArgumentParser(description='Prune and save model')
    parser.add_argument('--input', dest='input', help='raw model for pruning and save',
                        default='', type=str)
    parser.add_argument('--output', dest='output', help='pruning and save result model',
                        default='', type=str)
    parser.add_argument('--with_transpiler', dest='with_transpiler', help='with transpiler',
                        default=True, type=bool)
    args = parser.parse_args()
    return args


def prune(args):
    model = args.input
    pruned_dir = args.output

    place = fluid.CPUPlace()
    exe = fluid.Executor(place)
    [test_program, feed_dict,
     fetch_targets] = fluid.io.load_inference_model(model, exe)

    if args.with_transpiler:
        t = fluid.transpiler.InferenceTranspiler()
        t.transpile(test_program, place)

    prune_index = -1
    for op_index, op in enumerate(test_program.current_block().ops):
        if op.type == "softmax" or op.type == "cross_entropy":
            prune_fetch_list = []
            for input_name in op.input_names:
    	        if input_name != 'X':
    	            continue
                prune_fetch_list.append(op.input(input_name)[0])
            prune_index = op_index
            break

    for index in range(len(test_program.current_block().ops) - 1, prune_index - 1, -1):
        test_program.current_block()._remove_op(index)
    label_index = feed_dict.index('label')
    feed_dict.pop(label_index)
    fetch_targets = [test_program.current_block().var(i)
                     for i in prune_fetch_list]
    fluid.io.save_inference_model(pruned_dir, feed_dict, fetch_targets,
                                  exe, test_program, model_filename="model", params_filename="params")


if __name__ == '__main__':
    args = parse_args()
    print(args)
    prune(args)

# save weights as “model” and “params” file
FLAGS_use_mkldnn=true python prune.py --input ~/.cache/paddle/dataset/int8/download/resnet50_fp32/model/ --output resnet50_fp32
FLAGS_use_mkldnn=true python prune.py --input calibration_out/ --output resnet50_int8

4.Run CAPI test application

        cd $Paddle_home/build
	# run capi test application
	./paddle/fluid/inference/tests/api/test_analyzer_resnet50 --infer_model=../resnet50_fp32/ --repeat=100 --gtest_filter=Analyzer_resnet50.profile_mkldnn --batch_size=1 --num_threads=1
       ./paddle/fluid/inference/tests/api/test_analyzer_resnet50 --infer_model=../resnet50_int8/ --repeat=100 --gtest_filter=Analyzer_resnet50.profile_mkldnn --batch_size=1 --num_threads=1
	# you can change the ‘infer_model’ to image classification model you saved.

5. Performance result on SKX-6148

ResNet50 result

Threads Num.		1			2	4
Batch size		1	2	4	1	1
Avg. FPS / per thread	FP32	17.92	9.23 * 2	4.89 * 4	17.02	16.90
	INT8	28.47	14.93 * 2	7.67 * 4	28.23	27.21
	INT8/FP32	1.59	1.62	1.57	1.66	1.61

Mobilenet v1 result

Threads Num.		1			2	4
Batch size		1	2	4	1	1
Avg. FPS / per thread	FP32	91.2	45.46 * 2	24.33 * 4	93.13	87.72
	INT8	155.44	81.15 * 2	41.76 * 4	153.73	152.77
	INT8/FP32	1.70	1.79	1.72	1.65	1.74

Please review on it. @luotao1 @hshen14