Skip to content

Commit 82c1fe9

Browse files
author
AFLALO, Jonathan Isaac
committed
added hardcorenas model
refactoring to support feature extraction
1 parent 5c7d298 commit 82c1fe9

File tree

2 files changed

+150
-0
lines changed

2 files changed

+150
-0
lines changed

‎timm/models/__init__.py‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -29,10 +29,12 @@
2929
from .vovnet import *
3030
from .xception import *
3131
from .xception_aligned import *
32+
from .hardcorenas import *
3233

3334
from .factory import create_model
3435
from .helpers import load_checkpoint, resume_checkpoint, model_parameters
3536
from .layers import TestTimePoolHead, apply_test_time_pool
3637
from .layers import convert_splitbn_model
3738
from .layers import is_scriptable, is_exportable, set_scriptable, set_exportable, is_no_jit, set_no_jit
3839
from .registry import *
40+

‎timm/models/hardcorenas.py‎

Lines changed: 148 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,148 @@
1+
import torch.nn as nn
2+
from .efficientnet_builder import decode_arch_def, resolve_bn_args
3+
from .mobilenetv3 import MobileNetV3, MobileNetV3Features, build_model_with_cfg, default_cfg_for_features
4+
from .layers import hard_sigmoid
5+
from .efficientnet_blocks import resolve_act_layer
6+
from .registry import register_model
7+
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
8+
9+
10+
def _cfg(url='', **kwargs):
11+
return {
12+
'url': url, 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': (1, 1),
13+
'crop_pct': 0.875, 'interpolation': 'bilinear',
14+
'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
15+
'first_conv': 'conv_stem', 'classifier': 'classifier',
16+
**kwargs
17+
}
18+
19+
20+
default_cfgs = {
21+
'hardcorenas_A': _cfg(url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/public/HardCoReNAS/HardCoreNAS_A_Green_38ms_75.9_23474aeb.pth'),
22+
'hardcorenas_B': _cfg(url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/public/HardCoReNAS/HardCoreNAS_B_Green_40ms_76.5_1f882d1e.pth'),
23+
'hardcorenas_C': _cfg(url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/public/HardCoReNAS/HardCoreNAS_C_Green_44ms_77.1_d4148c9e.pth'),
24+
'hardcorenas_D': _cfg(url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/public/HardCoReNAS/HardCoreNAS_D_Green_50ms_77.4_23e3cdde.pth'),
25+
'hardcorenas_E': _cfg(url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/public/HardCoReNAS/HardCoreNAS_E_Green_55ms_77.9_90f20e8a.pth'),
26+
'hardcorenas_F': _cfg(url='https://miil-public-eu.oss-eu-central-1.aliyuncs.com/public/HardCoReNAS/HardCoreNAS_F_Green_60ms_78.1_2855edf1.pth'),
27+
}
28+
29+
def _gen_hardcorenas(pretrained, variant, arch_def, **kwargs):
30+
"""Creates a hardcorenas model
31+
32+
Ref impl: https://github.com/Alibaba-MIIL/HardCoReNAS
33+
Paper: https://arxiv.org/abs/2102.11646
34+
35+
"""
36+
num_features = 1280
37+
act_layer = resolve_act_layer(kwargs, 'hard_swish')
38+
39+
model_kwargs = dict(
40+
block_args=decode_arch_def(arch_def),
41+
num_features=num_features,
42+
stem_size=32,
43+
channel_multiplier=1,
44+
norm_kwargs=resolve_bn_args(kwargs),
45+
act_layer=act_layer,
46+
se_kwargs=dict(act_layer=nn.ReLU, gate_fn=hard_sigmoid, reduce_mid=True, divisor=8),
47+
**kwargs,
48+
)
49+
50+
features_only = False
51+
model_cls = MobileNetV3
52+
if model_kwargs.pop('features_only', False):
53+
features_only = True
54+
model_kwargs.pop('num_classes', 0)
55+
model_kwargs.pop('num_features', 0)
56+
model_kwargs.pop('head_conv', None)
57+
model_kwargs.pop('head_bias', None)
58+
model_cls = MobileNetV3Features
59+
model = build_model_with_cfg(
60+
model_cls, variant, pretrained, default_cfg=default_cfgs[variant],
61+
pretrained_strict=not features_only, **model_kwargs)
62+
if features_only:
63+
model.default_cfg = default_cfg_for_features(model.default_cfg)
64+
return model
65+
66+
67+
@register_model
68+
def hardcorenas_A(pretrained=False, **kwargs):
69+
""" hardcorenas_A """
70+
arch_def = [['ds_r1_k3_s1_e1_c16_nre'], ['ir_r1_k5_s2_e3_c24_nre', 'ir_r1_k5_s1_e3_c24_nre_se0.25'],
71+
['ir_r1_k5_s2_e3_c40_nre', 'ir_r1_k5_s1_e6_c40_nre_se0.25'],
72+
['ir_r1_k5_s2_e6_c80_se0.25', 'ir_r1_k5_s1_e6_c80_se0.25'],
73+
['ir_r1_k5_s1_e6_c112_se0.25', 'ir_r1_k5_s1_e6_c112_se0.25'],
74+
['ir_r1_k5_s2_e6_c192_se0.25', 'ir_r1_k5_s1_e6_c192_se0.25'], ['cn_r1_k1_s1_c960']]
75+
model = _gen_hardcorenas(pretrained=pretrained, variant='hardcorenas_A', arch_def=arch_def, **kwargs)
76+
return model
77+
78+
79+
@register_model
80+
def hardcorenas_B(pretrained=False, **kwargs):
81+
""" hardcorenas_B """
82+
arch_def = [['ds_r1_k3_s1_e1_c16_nre'],
83+
['ir_r1_k5_s2_e3_c24_nre', 'ir_r1_k5_s1_e3_c24_nre_se0.25', 'ir_r1_k3_s1_e3_c24_nre'],
84+
['ir_r1_k5_s2_e3_c40_nre', 'ir_r1_k5_s1_e3_c40_nre', 'ir_r1_k5_s1_e3_c40_nre'],
85+
['ir_r1_k5_s2_e3_c80', 'ir_r1_k5_s1_e3_c80', 'ir_r1_k3_s1_e3_c80', 'ir_r1_k3_s1_e3_c80'],
86+
['ir_r1_k5_s1_e3_c112', 'ir_r1_k3_s1_e3_c112', 'ir_r1_k3_s1_e3_c112', 'ir_r1_k3_s1_e3_c112'],
87+
['ir_r1_k5_s2_e6_c192_se0.25', 'ir_r1_k5_s1_e6_c192_se0.25', 'ir_r1_k3_s1_e3_c192_se0.25'],
88+
['cn_r1_k1_s1_c960']]
89+
model = _gen_hardcorenas(pretrained=pretrained, variant='hardcorenas_B', arch_def=arch_def, **kwargs)
90+
return model
91+
92+
93+
@register_model
94+
def hardcorenas_C(pretrained=False, **kwargs):
95+
""" hardcorenas_C """
96+
arch_def = [['ds_r1_k3_s1_e1_c16_nre'], ['ir_r1_k5_s2_e3_c24_nre', 'ir_r1_k5_s1_e3_c24_nre_se0.25'],
97+
['ir_r1_k5_s2_e3_c40_nre', 'ir_r1_k5_s1_e3_c40_nre', 'ir_r1_k5_s1_e3_c40_nre',
98+
'ir_r1_k5_s1_e3_c40_nre'],
99+
['ir_r1_k5_s2_e4_c80', 'ir_r1_k5_s1_e6_c80_se0.25', 'ir_r1_k3_s1_e3_c80', 'ir_r1_k3_s1_e3_c80'],
100+
['ir_r1_k5_s1_e6_c112_se0.25', 'ir_r1_k3_s1_e3_c112', 'ir_r1_k3_s1_e3_c112', 'ir_r1_k3_s1_e3_c112'],
101+
['ir_r1_k5_s2_e6_c192_se0.25', 'ir_r1_k5_s1_e6_c192_se0.25', 'ir_r1_k3_s1_e3_c192_se0.25'],
102+
['cn_r1_k1_s1_c960']]
103+
model = _gen_hardcorenas(pretrained=pretrained, variant='hardcorenas_C', arch_def=arch_def, **kwargs)
104+
return model
105+
106+
107+
@register_model
108+
def hardcorenas_D(pretrained=False, **kwargs):
109+
""" hardcorenas_D """
110+
arch_def = [['ds_r1_k3_s1_e1_c16_nre'], ['ir_r1_k5_s2_e3_c24_nre_se0.25', 'ir_r1_k5_s1_e3_c24_nre_se0.25'],
111+
['ir_r1_k5_s2_e3_c40_nre_se0.25', 'ir_r1_k5_s1_e4_c40_nre_se0.25', 'ir_r1_k3_s1_e3_c40_nre_se0.25'],
112+
['ir_r1_k5_s2_e4_c80_se0.25', 'ir_r1_k3_s1_e3_c80_se0.25', 'ir_r1_k3_s1_e3_c80_se0.25',
113+
'ir_r1_k3_s1_e3_c80_se0.25'],
114+
['ir_r1_k3_s1_e4_c112_se0.25', 'ir_r1_k5_s1_e4_c112_se0.25', 'ir_r1_k3_s1_e3_c112_se0.25',
115+
'ir_r1_k5_s1_e3_c112_se0.25'],
116+
['ir_r1_k5_s2_e6_c192_se0.25', 'ir_r1_k5_s1_e6_c192_se0.25', 'ir_r1_k5_s1_e6_c192_se0.25',
117+
'ir_r1_k3_s1_e6_c192_se0.25'], ['cn_r1_k1_s1_c960']]
118+
model = _gen_hardcorenas(pretrained=pretrained, variant='hardcorenas_D', arch_def=arch_def, **kwargs)
119+
return model
120+
121+
122+
@register_model
123+
def hardcorenas_E(pretrained=False, **kwargs):
124+
""" hardcorenas_E """
125+
arch_def = [['ds_r1_k3_s1_e1_c16_nre'], ['ir_r1_k5_s2_e3_c24_nre_se0.25', 'ir_r1_k5_s1_e3_c24_nre_se0.25'],
126+
['ir_r1_k5_s2_e6_c40_nre_se0.25', 'ir_r1_k5_s1_e4_c40_nre_se0.25', 'ir_r1_k5_s1_e4_c40_nre_se0.25',
127+
'ir_r1_k3_s1_e3_c40_nre_se0.25'], ['ir_r1_k5_s2_e4_c80_se0.25', 'ir_r1_k3_s1_e6_c80_se0.25'],
128+
['ir_r1_k5_s1_e6_c112_se0.25', 'ir_r1_k5_s1_e6_c112_se0.25', 'ir_r1_k5_s1_e6_c112_se0.25',
129+
'ir_r1_k5_s1_e3_c112_se0.25'],
130+
['ir_r1_k5_s2_e6_c192_se0.25', 'ir_r1_k5_s1_e6_c192_se0.25', 'ir_r1_k5_s1_e6_c192_se0.25',
131+
'ir_r1_k3_s1_e6_c192_se0.25'], ['cn_r1_k1_s1_c960']]
132+
model = _gen_hardcorenas(pretrained=pretrained, variant='hardcorenas_E', arch_def=arch_def, **kwargs)
133+
return model
134+
135+
136+
@register_model
137+
def hardcorenas_F(pretrained=False, **kwargs):
138+
""" hardcorenas_F """
139+
arch_def = [['ds_r1_k3_s1_e1_c16_nre'], ['ir_r1_k5_s2_e3_c24_nre_se0.25', 'ir_r1_k5_s1_e3_c24_nre_se0.25'],
140+
['ir_r1_k5_s2_e6_c40_nre_se0.25', 'ir_r1_k5_s1_e6_c40_nre_se0.25'],
141+
['ir_r1_k5_s2_e6_c80_se0.25', 'ir_r1_k5_s1_e6_c80_se0.25', 'ir_r1_k3_s1_e3_c80_se0.25',
142+
'ir_r1_k3_s1_e3_c80_se0.25'],
143+
['ir_r1_k3_s1_e6_c112_se0.25', 'ir_r1_k5_s1_e6_c112_se0.25', 'ir_r1_k5_s1_e6_c112_se0.25',
144+
'ir_r1_k3_s1_e3_c112_se0.25'],
145+
['ir_r1_k5_s2_e6_c192_se0.25', 'ir_r1_k5_s1_e6_c192_se0.25', 'ir_r1_k3_s1_e6_c192_se0.25',
146+
'ir_r1_k3_s1_e6_c192_se0.25'], ['cn_r1_k1_s1_c960']]
147+
model = _gen_hardcorenas(pretrained=pretrained, variant='hardcorenas_F', arch_def=arch_def, **kwargs)
148+
return model

0 commit comments

Comments
 (0)