量化训练模型删除指定op转换功能

import paddle.fluid as fluid
from pyramidbox_test import PyramidBox
from paddle.fluid.framework import IrGraph
from paddle.fluid import core
from paddle.fluid.contrib.slim.quantization.quantization_pass import *
import sys
if __name__ == '__main__':
    float_model_path_src = 'faceboxes_V016_float32_compress_0327_float'

    place = fluid.CPUPlace()
    exe = fluid.Executor(place)

    # step1: define infer network without assign op
    main_program = fluid.Program()
    startup_program = fluid.Program()

    with fluid.program_guard(main_program, startup_program):

        print('debug: construct network')
        network = PyramidBox(
            data_shape=[3, 240, 320],
            sub_network=True,
            is_infer=True)
        inference_program, nmsed_out = network.infer(main_program)
        print('nmsed_out name: {}'.format(nmsed_out.name))
    # step2: init variables
    exe.run(startup_program)

    # step3: insert quantzation operators into infer network
    transform_pass = QuantizationTransformPass(
            scope=fluid.executor.global_scope(),
            place=place,
            weight_bits=8,
            activation_bits=8,
            activation_quantize_type='range_abs_max')
    eval_graph = IrGraph(core.Graph(inference_program.desc), for_test=True)
    transform_pass.apply(eval_graph)

    # step4: load float weights into global scope
    fluid.io.load_inference_model(float_model_path_src, exe,
                                   model_filename='model',
                                   params_filename='weights')

    # step5: convert to int8 model
    freeze_pass = QuantizationFreezePass(
                             scope=fluid.executor.global_scope(),
                             place=place,
                             weight_bits=8,
                             activation_bits=8)
    freeze_pass.apply(eval_graph)
    convert_int8_pass = ConvertToInt8Pass(
                               scope=fluid.executor.global_scope(),
                               place=place)
    convert_int8_pass.apply(eval_graph)

    eval_graph.draw('.','eval')

    # step6: save int8 model into filesystem
    program = eval_graph.to_program()
    out_vars = [program.global_block().var('detection_output_0.tmp_0')]
    fluid.io.save_inference_model(
           "./output_int8",
           ['image'],
           out_vars,
           exe,
           main_program=program,
           model_filename='model',
           params_filename='weights',
           export_for_deployment=True)

    # step7: load and check
    program, feed_target_names, fetch_targets = fluid.io.load_inference_model("./output_int8", exe,
                                   model_filename='model',
                                   params_filename='weights')
    print("feed_target_names: {}".format(feed_target_names))
    print("fetch_targets: {}".format(fetch_targets))

    for op in program.global_block().ops:
        print("op: {}".format(op.type))

    for var in program.list_vars():
        if var.persistable:
            print("var: {}".format(var.name))
PaddlePaddle / models

量化训练模型删除指定op转换功能 #2022
