Skip to content

Commit 1f4d189

Browse files
VascoLopesVascoLopes
authored
Add files via upload
1 parent 54a15e9 commit 1f4d189

18 files changed

+4586
-2
lines changed

‎README.md

Lines changed: 45 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,2 +1,45 @@
1-
# LCMNAS
2-
Code for Towards Less Constrained Macro-Neural Architecture Search
1+
# UMNAS
2+
3+
## Instalation:
4+
sudo apt-get install python3-dev graphviz libgraphviz-dev pkg-config
5+
python3 -m venv venv
6+
source venv/bin/activate
7+
pip install wheel torchcontrib pytorchcv
8+
pip install -r requirements.txt
9+
10+
11+
## Get Data sets (informations are anonymized)
12+
Download data.zip and extract it in the main directory of the code from https://drive.google.com/drive/folders/1bnXQfYfNyYKCSX05DSFG8gE---EkfXSP?usp=sharing
13+
14+
## Get Best Architectures Found (informations are anonymized)
15+
Download models.zip (these do not have the autoaugment training) from https://drive.google.com/drive/folders/1bnXQfYfNyYKCSX05DSFG8gE---EkfXSP?usp=sharing
16+
17+
18+
## Experiments (the results get automatically stored in the experiments folder)
19+
### Mixed-performance estimation mechanism: (lambda=0.75)
20+
python3 count.py --generations 50 --population 100 --dataset cifar10 --batch_size 96 —-mixed_training --mixed_fitness_lambda 0.75
21+
22+
### Using only zero-proxy estimation (lambda=1)
23+
python3 count.py --generations 50 --population 100 --dataset cifar10 --batch_size 96 --without_training --mixed_fitness_lambda 1
24+
25+
### Using only regular training (lambda=0)
26+
python3 count.py --generations 50 --population 100 --dataset cifar10 --batch_size 96 --mixed_fitness_lambda 0
27+
28+
### Available datasets: cifar10 cifar100 ImageNet16-120
29+
python3 count.py --generations 50 --population 100 --dataset cifar10 --batch_size 96 —-mixed_training --mixed_fitness_lambda 0.75
30+
31+
python3 count.py --generations 50 --population 100 --dataset cifar100 --batch_size 96 —-mixed_training --mixed_fitness_lambda 0.75
32+
33+
python3 count.py --generations 50 --population 100 --dataset ImageNet16 --batch_size 96 —-mixed_training --mixed_fitness_lambda 0.75
34+
35+
36+
### Train with auto-augment
37+
python aux.py --searched_dataset cifar10 --dataset cifar10 --auto_augment --epochs 1500 --reset_weight --model_path "path_to_model"
38+
39+
40+
### Transfer to other data set
41+
python aux.py --searched_dataset cifar10 --dataset ImageNet16 --reset_weight --model_path "path_to_model"
42+
43+
44+
45+
Thank you for the time spent reviewing our paper.

‎autoaugment.py

Lines changed: 228 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,228 @@
1+
from PIL import Image, ImageEnhance, ImageOps
2+
import numpy as np
3+
import random
4+
5+
6+
class ImageNetPolicy(object):
7+
""" Randomly choose one of the best 24 Sub-policies on ImageNet.
8+
9+
Example:
10+
>>> policy = ImageNetPolicy()
11+
>>> transformed = policy(image)
12+
13+
Example as a PyTorch Transform:
14+
>>> transform=transforms.Compose([
15+
>>> transforms.Resize(256),
16+
>>> ImageNetPolicy(),
17+
>>> transforms.ToTensor()])
18+
"""
19+
def __init__(self, fillcolor=(128, 128, 128)):
20+
self.policies = [
21+
SubPolicy(0.4, "posterize", 8, 0.6, "rotate", 9, fillcolor),
22+
SubPolicy(0.6, "solarize", 5, 0.6, "autocontrast", 5, fillcolor),
23+
SubPolicy(0.8, "equalize", 8, 0.6, "equalize", 3, fillcolor),
24+
SubPolicy(0.6, "posterize", 7, 0.6, "posterize", 6, fillcolor),
25+
SubPolicy(0.4, "equalize", 7, 0.2, "solarize", 4, fillcolor),
26+
27+
SubPolicy(0.4, "equalize", 4, 0.8, "rotate", 8, fillcolor),
28+
SubPolicy(0.6, "solarize", 3, 0.6, "equalize", 7, fillcolor),
29+
SubPolicy(0.8, "posterize", 5, 1.0, "equalize", 2, fillcolor),
30+
SubPolicy(0.2, "rotate", 3, 0.6, "solarize", 8, fillcolor),
31+
SubPolicy(0.6, "equalize", 8, 0.4, "posterize", 6, fillcolor),
32+
33+
SubPolicy(0.8, "rotate", 8, 0.4, "color", 0, fillcolor),
34+
SubPolicy(0.4, "rotate", 9, 0.6, "equalize", 2, fillcolor),
35+
SubPolicy(0.0, "equalize", 7, 0.8, "equalize", 8, fillcolor),
36+
SubPolicy(0.6, "invert", 4, 1.0, "equalize", 8, fillcolor),
37+
SubPolicy(0.6, "color", 4, 1.0, "contrast", 8, fillcolor),
38+
39+
SubPolicy(0.8, "rotate", 8, 1.0, "color", 2, fillcolor),
40+
SubPolicy(0.8, "color", 8, 0.8, "solarize", 7, fillcolor),
41+
SubPolicy(0.4, "sharpness", 7, 0.6, "invert", 8, fillcolor),
42+
SubPolicy(0.6, "shearX", 5, 1.0, "equalize", 9, fillcolor),
43+
SubPolicy(0.4, "color", 0, 0.6, "equalize", 3, fillcolor),
44+
45+
SubPolicy(0.4, "equalize", 7, 0.2, "solarize", 4, fillcolor),
46+
SubPolicy(0.6, "solarize", 5, 0.6, "autocontrast", 5, fillcolor),
47+
SubPolicy(0.6, "invert", 4, 1.0, "equalize", 8, fillcolor),
48+
SubPolicy(0.6, "color", 4, 1.0, "contrast", 8, fillcolor),
49+
SubPolicy(0.8, "equalize", 8, 0.6, "equalize", 3, fillcolor)
50+
]
51+
52+
53+
def __call__(self, img):
54+
policy_idx = random.randint(0, len(self.policies) - 1)
55+
return self.policies[policy_idx](img)
56+
57+
def __repr__(self):
58+
return "AutoAugment ImageNet Policy"
59+
60+
61+
class CIFAR10Policy(object):
62+
""" Randomly choose one of the best 25 Sub-policies on CIFAR10.
63+
"""
64+
def __init__(self, fillcolor=(128, 128, 128)):
65+
self.policies = [
66+
SubPolicy(0.1, "invert", 7, 0.2, "contrast", 6, fillcolor),
67+
SubPolicy(0.7, "rotate", 2, 0.3, "translateX", 9, fillcolor),
68+
SubPolicy(0.8, "sharpness", 1, 0.9, "sharpness", 3, fillcolor),
69+
SubPolicy(0.5, "shearY", 8, 0.7, "translateY", 9, fillcolor),
70+
SubPolicy(0.5, "autocontrast", 8, 0.9, "equalize", 2, fillcolor),
71+
72+
SubPolicy(0.2, "shearY", 7, 0.3, "posterize", 7, fillcolor),
73+
SubPolicy(0.4, "color", 3, 0.6, "brightness", 7, fillcolor),
74+
SubPolicy(0.3, "sharpness", 9, 0.7, "brightness", 9, fillcolor),
75+
SubPolicy(0.6, "equalize", 5, 0.5, "equalize", 1, fillcolor),
76+
SubPolicy(0.6, "contrast", 7, 0.6, "sharpness", 5, fillcolor),
77+
78+
SubPolicy(0.7, "color", 7, 0.5, "translateX", 8, fillcolor),
79+
SubPolicy(0.3, "equalize", 7, 0.4, "autocontrast", 8, fillcolor),
80+
SubPolicy(0.4, "translateY", 3, 0.2, "sharpness", 6, fillcolor),
81+
SubPolicy(0.9, "brightness", 6, 0.2, "color", 8, fillcolor),
82+
SubPolicy(0.5, "solarize", 2, 0.0, "invert", 3, fillcolor),
83+
84+
SubPolicy(0.2, "equalize", 0, 0.6, "autocontrast", 0, fillcolor),
85+
SubPolicy(0.2, "equalize", 8, 0.8, "equalize", 4, fillcolor),
86+
SubPolicy(0.9, "color", 9, 0.6, "equalize", 6, fillcolor),
87+
SubPolicy(0.8, "autocontrast", 4, 0.2, "solarize", 8, fillcolor),
88+
SubPolicy(0.1, "brightness", 3, 0.7, "color", 0, fillcolor),
89+
90+
SubPolicy(0.4, "solarize", 5, 0.9, "autocontrast", 3, fillcolor),
91+
SubPolicy(0.9, "translateY", 9, 0.7, "translateY", 9, fillcolor),
92+
SubPolicy(0.9, "autocontrast", 2, 0.8, "solarize", 3, fillcolor),
93+
SubPolicy(0.8, "equalize", 8, 0.1, "invert", 3, fillcolor),
94+
SubPolicy(0.7, "translateY", 9, 0.9, "autocontrast", 1, fillcolor)
95+
]
96+
97+
98+
def __call__(self, img):
99+
policy_idx = random.randint(0, len(self.policies) - 1)
100+
return self.policies[policy_idx](img)
101+
102+
def __repr__(self):
103+
return "AutoAugment CIFAR10 Policy"
104+
105+
106+
class SVHNPolicy(object):
107+
""" Randomly choose one of the best 25 Sub-policies on SVHN.
108+
109+
Example:
110+
>>> policy = SVHNPolicy()
111+
>>> transformed = policy(image)
112+
113+
Example as a PyTorch Transform:
114+
>>> transform=transforms.Compose([
115+
>>> transforms.Resize(256),
116+
>>> SVHNPolicy(),
117+
>>> transforms.ToTensor()])
118+
"""
119+
def __init__(self, fillcolor=(128, 128, 128)):
120+
self.policies = [
121+
SubPolicy(0.9, "shearX", 4, 0.2, "invert", 3, fillcolor),
122+
SubPolicy(0.9, "shearY", 8, 0.7, "invert", 5, fillcolor),
123+
SubPolicy(0.6, "equalize", 5, 0.6, "solarize", 6, fillcolor),
124+
SubPolicy(0.9, "invert", 3, 0.6, "equalize", 3, fillcolor),
125+
SubPolicy(0.6, "equalize", 1, 0.9, "rotate", 3, fillcolor),
126+
127+
SubPolicy(0.9, "shearX", 4, 0.8, "autocontrast", 3, fillcolor),
128+
SubPolicy(0.9, "shearY", 8, 0.4, "invert", 5, fillcolor),
129+
SubPolicy(0.9, "shearY", 5, 0.2, "solarize", 6, fillcolor),
130+
SubPolicy(0.9, "invert", 6, 0.8, "autocontrast", 1, fillcolor),
131+
SubPolicy(0.6, "equalize", 3, 0.9, "rotate", 3, fillcolor),
132+
133+
SubPolicy(0.9, "shearX", 4, 0.3, "solarize", 3, fillcolor),
134+
SubPolicy(0.8, "shearY", 8, 0.7, "invert", 4, fillcolor),
135+
SubPolicy(0.9, "equalize", 5, 0.6, "translateY", 6, fillcolor),
136+
SubPolicy(0.9, "invert", 4, 0.6, "equalize", 7, fillcolor),
137+
SubPolicy(0.3, "contrast", 3, 0.8, "rotate", 4, fillcolor),
138+
139+
SubPolicy(0.8, "invert", 5, 0.0, "translateY", 2, fillcolor),
140+
SubPolicy(0.7, "shearY", 6, 0.4, "solarize", 8, fillcolor),
141+
SubPolicy(0.6, "invert", 4, 0.8, "rotate", 4, fillcolor),
142+
SubPolicy(0.3, "shearY", 7, 0.9, "translateX", 3, fillcolor),
143+
SubPolicy(0.1, "shearX", 6, 0.6, "invert", 5, fillcolor),
144+
145+
SubPolicy(0.7, "solarize", 2, 0.6, "translateY", 7, fillcolor),
146+
SubPolicy(0.8, "shearY", 4, 0.8, "invert", 8, fillcolor),
147+
SubPolicy(0.7, "shearX", 9, 0.8, "translateY", 3, fillcolor),
148+
SubPolicy(0.8, "shearY", 5, 0.7, "autocontrast", 3, fillcolor),
149+
SubPolicy(0.7, "shearX", 2, 0.1, "invert", 5, fillcolor)
150+
]
151+
152+
153+
def __call__(self, img):
154+
policy_idx = random.randint(0, len(self.policies) - 1)
155+
return self.policies[policy_idx](img)
156+
157+
def __repr__(self):
158+
return "AutoAugment SVHN Policy"
159+
160+
161+
class SubPolicy(object):
162+
def __init__(self, p1, operation1, magnitude_idx1, p2, operation2, magnitude_idx2, fillcolor=(128, 128, 128)):
163+
ranges = {
164+
"shearX": np.linspace(0, 0.3, 10),
165+
"shearY": np.linspace(0, 0.3, 10),
166+
"translateX": np.linspace(0, 150 / 331, 10),
167+
"translateY": np.linspace(0, 150 / 331, 10),
168+
"rotate": np.linspace(0, 30, 10),
169+
"color": np.linspace(0.0, 0.9, 10),
170+
"posterize": np.round(np.linspace(8, 4, 10), 0).astype(np.int),
171+
"solarize": np.linspace(256, 0, 10),
172+
"contrast": np.linspace(0.0, 0.9, 10),
173+
"sharpness": np.linspace(0.0, 0.9, 10),
174+
"brightness": np.linspace(0.0, 0.9, 10),
175+
"autocontrast": [0] * 10,
176+
"equalize": [0] * 10,
177+
"invert": [0] * 10
178+
}
179+
180+
# from https://stackoverflow.com/questions/5252170/specify-image-filling-color-when-rotating-in-python-with-pil-and-setting-expand
181+
def rotate_with_fill(img, magnitude):
182+
rot = img.convert("RGBA").rotate(magnitude)
183+
return Image.composite(rot, Image.new("RGBA", rot.size, (128,) * 4), rot).convert(img.mode)
184+
185+
func = {
186+
"shearX": lambda img, magnitude: img.transform(
187+
img.size, Image.AFFINE, (1, magnitude * random.choice([-1, 1]), 0, 0, 1, 0),
188+
Image.BICUBIC, fillcolor=fillcolor),
189+
"shearY": lambda img, magnitude: img.transform(
190+
img.size, Image.AFFINE, (1, 0, 0, magnitude * random.choice([-1, 1]), 1, 0),
191+
Image.BICUBIC, fillcolor=fillcolor),
192+
"translateX": lambda img, magnitude: img.transform(
193+
img.size, Image.AFFINE, (1, 0, magnitude * img.size[0] * random.choice([-1, 1]), 0, 1, 0),
194+
fillcolor=fillcolor),
195+
"translateY": lambda img, magnitude: img.transform(
196+
img.size, Image.AFFINE, (1, 0, 0, 0, 1, magnitude * img.size[1] * random.choice([-1, 1])),
197+
fillcolor=fillcolor),
198+
"rotate": lambda img, magnitude: rotate_with_fill(img, magnitude),
199+
# "rotate": lambda img, magnitude: img.rotate(magnitude * random.choice([-1, 1])),
200+
"color": lambda img, magnitude: ImageEnhance.Color(img).enhance(1 + magnitude * random.choice([-1, 1])),
201+
"posterize": lambda img, magnitude: ImageOps.posterize(img, magnitude),
202+
"solarize": lambda img, magnitude: ImageOps.solarize(img, magnitude),
203+
"contrast": lambda img, magnitude: ImageEnhance.Contrast(img).enhance(
204+
1 + magnitude * random.choice([-1, 1])),
205+
"sharpness": lambda img, magnitude: ImageEnhance.Sharpness(img).enhance(
206+
1 + magnitude * random.choice([-1, 1])),
207+
"brightness": lambda img, magnitude: ImageEnhance.Brightness(img).enhance(
208+
1 + magnitude * random.choice([-1, 1])),
209+
"autocontrast": lambda img, magnitude: ImageOps.autocontrast(img),
210+
"equalize": lambda img, magnitude: ImageOps.equalize(img),
211+
"invert": lambda img, magnitude: ImageOps.invert(img)
212+
}
213+
214+
# self.name = "{}_{:.2f}_and_{}_{:.2f}".format(
215+
# operation1, ranges[operation1][magnitude_idx1],
216+
# operation2, ranges[operation2][magnitude_idx2])
217+
self.p1 = p1
218+
self.operation1 = func[operation1]
219+
self.magnitude1 = ranges[operation1][magnitude_idx1]
220+
self.p2 = p2
221+
self.operation2 = func[operation2]
222+
self.magnitude2 = ranges[operation2][magnitude_idx2]
223+
224+
225+
def __call__(self, img):
226+
if random.random() < self.p1: img = self.operation1(img, self.magnitude1)
227+
if random.random() < self.p2: img = self.operation2(img, self.magnitude2)
228+
return img

0 commit comments

Comments
 (0)