提交 9fdb6054 编写于 作者: C ceci3

add multi search_space support

上级 af0eb732
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
from .search_space_base import SearchSpaceBase
from .search_space_registry import SEARCHSPACE
from .base_layer import conv_bn_layer
__all__ = ["CombineSearchSpace"]
class CombineSearchSpace(object):
"""
Combine Search Space.
Args:
configs(list<tuple>): multi config.
"""
def __init__(self, config_lists):
self.lens = len(config_lists)
self.spaces = []
for config_list in config_lists:
key, config = config_list
self.spaces.append(self._get_single_search_space(key, config))
def _get_single_search_space(self, key, config):
"""
get specific model space based on key and config.
Args:
key(str): model space name.
config(dict): basic config information.
return:
model space(class)
"""
cls = SEARCHSPACE.get(key)
space = cls(config['input_size'], config['output_size'],
config['block_num'])
return space
def init_tokens(self):
"""
Combine init tokens.
"""
tokens = []
self.token = []
for space in self.spaces:
tokens.extend(space.init_tokens())
self.token.append(space.init_tokens())
return tokens
def range_table(self):
"""
Combine range table.
"""
range_tables = []
for space in self.spaces:
range_tables.extend(space.range_table())
return range_tables
def token2arch(self, tokens=None):
"""
Combine model arch
"""
if tokens is None:
self.init_tokens()
model_archs = []
for space, token in zip(self.spaces, self.token):
model_archs.append(space.token2arch(token))
def net_arch(input):
for model_arch in model_archs:
input = model_arch(input)
return input
return net_arch
......@@ -52,6 +52,7 @@ class MobileNetV2Space(SearchSpaceBase):
self.scale = scale
self.class_dim = class_dim
def init_tokens(self):
"""
The initial token send to controller.
......@@ -60,7 +61,7 @@ class MobileNetV2Space(SearchSpaceBase):
"""
# original MobileNetV2
# yapf: disable
return [4, # 1, 16, 1
init_token_base = [4, # 1, 16, 1
4, 5, 1, 0, # 6, 24, 1
4, 5, 1, 0, # 6, 24, 2
4, 4, 2, 0, # 6, 32, 3
......@@ -69,6 +70,7 @@ class MobileNetV2Space(SearchSpaceBase):
4, 7, 2, 0, # 6, 160, 3
4, 9, 0, 0] # 6, 320, 1
# yapf: enable
return init_token_base#[:self.tokens_lens]
def range_table(self):
"""
......@@ -76,7 +78,7 @@ class MobileNetV2Space(SearchSpaceBase):
"""
# head_num + 7 * [multiple(expansion_factor), filter_num, repeat, kernel_size]
# yapf: disable
return [7,
range_table_base = [7,
5, 8, 6, 2,
5, 8, 6, 2,
5, 8, 6, 2,
......@@ -85,6 +87,7 @@ class MobileNetV2Space(SearchSpaceBase):
5, 10, 6, 2,
5, 12, 6, 2]
# yapf: enable
return range_table_base#[:self.tokens_lens]
def token2arch(self, tokens=None):
"""
......@@ -127,6 +130,8 @@ class MobileNetV2Space(SearchSpaceBase):
else:
break
self.tokens_lens = 1 + (len(bottleneck_params_list) - 1) * 4
def net_arch(input):
#conv1
# all padding is 'SAME' in the conv2d, can compute the actual padding automatic.
......@@ -137,7 +142,7 @@ class MobileNetV2Space(SearchSpaceBase):
stride=2,
padding='SAME',
act='relu6',
name='conv1_1')
name='mobilenetv2_conv1_1')
# bottleneck sequences
i = 1
......@@ -145,7 +150,7 @@ class MobileNetV2Space(SearchSpaceBase):
for layer_setting in bottleneck_params_list:
t, c, n, s, k = layer_setting
i += 1
input = self.invresi_blocks(
input = self._invresi_blocks(
input=input,
in_c=in_c,
t=t,
......@@ -153,7 +158,7 @@ class MobileNetV2Space(SearchSpaceBase):
n=n,
s=s,
k=k,
name='conv' + str(i))
name='mobilenetv2_conv' + str(i))
in_c = int(c * self.scale)
# if output_size is 1, add fc layer in the end
......@@ -161,8 +166,8 @@ class MobileNetV2Space(SearchSpaceBase):
input = fluid.layers.fc(
input=input,
size=self.class_dim,
param_attr=ParamAttr(name='fc10_weights'),
bias_attr=ParamAttr(name='fc10_offset'))
param_attr=ParamAttr(name='mobilenetv2_fc_weights'),
bias_attr=ParamAttr(name='mobilenetv2_fc_offset'))
else:
assert self.output_size == input.shape[2], \
("output_size must EQUAL to input_size / (2^block_num)."
......@@ -173,7 +178,7 @@ class MobileNetV2Space(SearchSpaceBase):
return net_arch
def shortcut(self, input, data_residual):
def _shortcut(self, input, data_residual):
"""Build shortcut layer.
Args:
input(Variable): input.
......@@ -183,7 +188,7 @@ class MobileNetV2Space(SearchSpaceBase):
"""
return fluid.layers.elementwise_add(input, data_residual)
def inverted_residual_unit(self,
def _inverted_residual_unit(self,
input,
num_in_filter,
num_filters,
......@@ -240,10 +245,10 @@ class MobileNetV2Space(SearchSpaceBase):
name=name + '_linear')
out = linear_out
if ifshortcut:
out = self.shortcut(input=input, data_residual=out)
out = self._shortcut(input=input, data_residual=out)
return out
def invresi_blocks(self, input, in_c, t, c, n, s, k, name=None):
def _invresi_blocks(self, input, in_c, t, c, n, s, k, name=None):
"""Build inverted residual blocks.
Args:
input: Variable, input.
......@@ -257,7 +262,7 @@ class MobileNetV2Space(SearchSpaceBase):
Returns:
Variable, layers output.
"""
first_block = self.inverted_residual_unit(
first_block = self._inverted_residual_unit(
input=input,
num_in_filter=in_c,
num_filters=c,
......@@ -271,7 +276,7 @@ class MobileNetV2Space(SearchSpaceBase):
last_c = c
for i in range(1, n):
last_residual_block = self.inverted_residual_unit(
last_residual_block = self._inverted_residual_unit(
input=last_residual_block,
num_in_filter=last_c,
num_filters=c,
......
......@@ -39,6 +39,6 @@ class SearchSpaceBase(object):
Args:
tokens(list<int>): The tokens which represent a network.
Return:
list<layers>
model arch
"""
raise NotImplementedError('Abstract method.')
......@@ -12,7 +12,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from search_space_registry import SEARCHSPACE
from .combine_search_space import CombineSearchSpace
__all__ = ["SearchSpaceFactory"]
......@@ -21,18 +21,11 @@ class SearchSpaceFactory(object):
def __init__(self):
pass
def get_search_space(self, key, config):
def get_search_space(self, config_lists):
"""
get specific model space based on key and config.
get model spaces based on list(key, config).
Args:
key(str): model space name.
config(dict): basic config information.
return:
model space(class)
"""
cls = SEARCHSPACE.get(key)
space = cls(config['input_size'], config['output_size'],
config['block_num'])
assert isinstance(config_lists, list), "configs must be a list"
return space
return CombineSearchSpace(config_lists)
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册