space_quant_k1dwk1.py

# Copyright (c) Alibaba, Inc. and its affiliates.
# The implementation is also open-sourced by the authors, and available at
# https://github.com/alibaba/lightweight-neural-architecture-search.

import bisect
import copy
import os
import random
import sys

from .base import CnnBaseSpace
from .builder import SPACES
from .space_k1kxk1 import Spacek1kxk1
from .mutator import build_mutator 


@SPACES.register_module(module_name = 'space_quant_k1dwk1')
class SpaceQuantk1dwk1(CnnBaseSpace):  
    def __init__(self, name = None, 
                    image_size = 224, 
                    block_num = 2, 
                    exclude_stem = False, 
                    budget_layers=None, 
                    maximum_channel =1280,
                    **kwargs): 

        super().__init__(name, image_size = image_size, 
                        block_num = block_num, 
                        exclude_stem = exclude_stem, 
                        budget_layers = budget_layers, 
                        **kwargs)

        kwargs = dict(stem_mutate_method_list = ['out'] ,
            the_maximum_channel = maximum_channel,
            nbits_ratio = 0.1,
            mutate_method_list = ['out', 'k', 'btn', 'L', 'nbits'],
            search_kernel_size_list = [3, 5],
            search_layer_list = [-2, -1, 1, 2],
            search_channel_list = [2.0, 1.5, 1.25, 0.8, 0.6, 0.5],
            search_nbits_list = [3, 4, 5, 6],
            search_btn_ratio_list = [1.5, 2.0, 2.5, 3.0, 3.5, 4.0],
            budget_layers = budget_layers
        )

        for n in ['ConvKXBNRELU','SuperQuantResK1DWK1']:
            self.mutators[n] = build_mutator(n, kwargs)