This repository has been archived by the owner on Apr 10, 2024. It is now read-only.
/
InceptionV1.py
100 lines (84 loc) · 4.24 KB
/
InceptionV1.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
# Copyright 2018 The Lucid 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, division, print_function
import tensorflow as tf
from lucid.modelzoo.vision_base import Model, _layers_from_list_of_dicts
def _populate_inception_bottlenecks(scope):
"""Add Inception bottlenecks and their pre-Relu versions to the graph."""
graph = tf.get_default_graph()
for op in graph.get_operations():
if op.name.startswith(scope+'/') and 'Concat' in op.type:
name = op.name.split('/')[1]
pre_relus = []
for tower in op.inputs[1:]:
if tower.op.type == 'Relu':
tower = tower.op.inputs[0]
pre_relus.append(tower)
concat_name = scope + '/' + name + '_pre_relu'
_ = tf.concat(pre_relus, -1, name=concat_name)
class InceptionV1(Model):
"""InceptionV1 (or 'GoogLeNet')
This is a (re?)implementation of InceptionV1 from the "Going deeper
with convolutions" paper. Links:
* Official CVPR paper, requires subscription: https://ieeexplore.ieee.org/document/7298594
* Author preprint: https://www.cs.unc.edu/~wliu/papers/GoogLeNet.pdf
* arXiv: https://arxiv.org/abs/1409.4842
The weights were trained at Google and released in an early TensorFlow
tutorial. It is possible the parameters are the original weights
(trained in TensorFlow's predecessor), but we haven't been able to
confirm this.
As far as we can tell, it is exactly the same as the model described in
the original paper, where as the slim and caffe implementations have
minor implementation differences (such as eliding extra classification heads).
"""
model_path = 'gs://modelzoo/vision/other_models/InceptionV1.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_alternate.txt'
synsets_path = 'gs://modelzoo/labels/ImageNet_alternate_synsets.txt'
dataset = 'ImageNet'
image_shape = [224, 224, 3]
image_value_range = (-117, 255-117)
input_name = 'input'
def post_import(self, scope):
_populate_inception_bottlenecks(scope)
InceptionV1.layers = _layers_from_list_of_dicts(InceptionV1(), [
{'tags': ['conv'], 'name': 'conv2d0', 'depth': 64},
{'tags': ['conv'], 'name': 'conv2d1', 'depth': 64},
{'tags': ['conv'], 'name': 'conv2d2', 'depth': 192},
{'tags': ['conv'], 'name': 'mixed3a', 'depth': 256},
{'tags': ['conv'], 'name': 'mixed3b', 'depth': 480},
{'tags': ['conv'], 'name': 'mixed4a', 'depth': 508},
{'tags': ['conv'], 'name': 'mixed4b', 'depth': 512},
{'tags': ['conv'], 'name': 'mixed4c', 'depth': 512},
{'tags': ['conv'], 'name': 'mixed4d', 'depth': 528},
{'tags': ['conv'], 'name': 'mixed4e', 'depth': 832},
{'tags': ['conv'], 'name': 'mixed5a', 'depth': 832},
{'tags': ['conv'], 'name': 'mixed5b', 'depth': 1024},
{'tags': ['conv'], 'name': 'head0_bottleneck', 'depth': 128},
{'tags': ['dense'], 'name': 'nn0', 'depth': 1024},
{'tags': ['dense'], 'name': 'softmax0', 'depth': 1008},
{'tags': ['conv'], 'name': 'head1_bottleneck', 'depth': 128},
{'tags': ['dense'], 'name': 'nn1', 'depth': 1024},
{'tags': ['dense'], 'name': 'softmax1', 'depth': 1008},
{'tags': ['dense'], 'name': 'softmax2', 'depth': 1008},
])
class InceptionV1_adv_finetuned(InceptionV1):
"""adversarially fine-tuned InceptionV1
This model is based on InceptionV1 and has been fine-tuned with
PGD-generated adversarial examples (https://arxiv.org/pdf/1706.06083.pdf).
The PGD-attack was L2-bounded with an epsilon of 255 (1.0 for normalized images).
After fine-tuning, this model achieves a robust top-5 accuracy of ~67%
for eps. 255 L2-bounded adversarial examples compared to ~4% before fine-tuning.
"""
model_path = 'gs://modelzoo/vision/other_models/InceptionV1_adv_finetuned.pb'