No module named models.Alexというエラーについて

前提・実現したいこと

下記のサイト
https://qiita.com/ohnabe/items/dc915c06ef37a2d30799
を参考にして、adversarial examplesの実行を試みておりますが、
No module named models.Alexというエラーに遭い、行き詰まってしまいました。

当方、python超初心者ですが、解決法をご教授いただけたらと思います。

発生している問題・エラーメッセージ

  File "create_as.py", line 9, in <module>
    import models.Alex
ImportError: No module named models.Alex

ソースコード

python
1#!/usr/bin/env python
2# -*- coding: utf-8 -*-
3
4import numpy as np
5import chainer
6import chainer.functions as F
7from chainer.links.caffe import CaffeFunction
8import pickle
9import models.Alex
10from PIL import Image
11
12from adv_models.fast_gradient import fast_gradient
13from adv_models.iterative_gradient import iterative_gradient
14from adv_models.iterative_least_likely import iterative_least_likely
15
16
17IMAGENET_MEAN_FILE = "../data/ilsvrc_2012_mean.npy"
18INPUT_IMAGE_SIZE = 227
19
20def load_caffemodel(model_path):
21    caffe_model = CaffeFunction(model_path)
22    return caffe_model
23
24
25def save_models(caffe_model, save_model_path):
26    with open(save_model_path, 'wb') as f:
27        pickle.dump(caffe_model, f)
28
29
30def load_models(save_model_path):
31    with open(save_model_path, 'rb') as f:
32        model = pickle.load(f)
33        return model
34
35
36def copy_model(src, dst):
37    assert isinstance(src, chainer.Chain)
38    assert isinstance(dst, chainer.Chain)
39    for child in src.children():
40        if child.name not in dst.__dict__: continue
41        dst_child = dst[child.name]
42        if type(child) != type(dst_child): continue
43        if isinstance(child, chainer.Chain):
44            copy_model(child, dst_child)
45        if isinstance(child, chainer.Link):
46            match = True
47            for a, b in zip(child.namedparams(), dst_child.namedparams()):
48                if a[0] != b[0]:
49                    match = False
50                    break
51                if a[1].data.shape != b[1].data.shape:
52                    match = False
53                    break
54            if not match:
55                print('Ignore %s because of parameter mismatch' % child.name)
56                continue
57            for a, b in zip(child.namedparams(), dst_child.namedparams()):
58                b[1].data = a[1].data
59            print('Copy %s' % child.name)
60
61
62def create_mean_image_array(pic_mean_data_path, size_image):
63    mean_data = np.load(pic_mean_data_path)
64    mean_data = Image.fromarray(mean_data.astype(np.uint8), 'RGB').resize((size_image, size_image))
65    mean_data = np.asarray(mean_data).astype(np.float32)
66    return mean_data
67
68
69def substract_mean_image(target_array, mean_array):
70    # mean_value: 104 B
71    # mean_value: 117 G
72    # mean_value: 123 R
73    result_array = target_array - mean_array
74    return result_array
75
76
77def add_mean_image(target_array, mean_array):
78    result_array = target_array + mean_array
79    return result_array
80
81
82def resize_image(original_image_path):
83    img = Image.open(original_image_path)
84    print("original image format:{} {}".format(img.size, img.mode))
85
86    img_resize = img.resize((INPUT_IMAGE_SIZE, INPUT_IMAGE_SIZE))
87    print("resize image format:{} {}".format(img_resize.size, img_resize.mode))
88    return img_resize
89
90
91def format2chainer(img_data):
92    # RGB to GBR
93    arrayImg = np.asarray(img_data).astype(np.float32)[:, :, ::-1]
94    # HWC to CWH
95    arrayImg = arrayImg.transpose(2, 0, 1)
96    # 3-dimensions to 4-dimensions
97    arrayImg = arrayImg.reshape((1,) + arrayImg.shape)
98    return arrayImg
99
100
101def format2orig(chainer_img):
102    # CWH to HWC
103    #orig_image = chainer_img.transpose(1, 2, 0).astype(np.uint8)
104    orig_image = chainer_img.transpose(1, 2, 0)
105    # BGR to RGB
106    orig_image = orig_image[:,:,::-1]
107    return orig_image
108
109
110def create_label_list(label_file_path):
111    label_d = {}
112    with open(label_file_path, "r") as f:
113        for line in f:
114            line = line.rstrip("\n").strip(" ").split(":")
115            if len(line) == 2:
116               label_d[int(line[0])] = line[1].strip(" ")
117    return label_d
118
119
120def get_result(predict_result, label_d):
121    prob = np.max(predict_result)
122    label_ind = np.argmax(predict_result)
123    label = label_d[label_ind]
124    return prob, label_ind, label
125
126
127if __name__ == '__main__':
128    #model = load_caffemodel("models/bvlc_alexnet.caffemodel")
129    #save_models(model, "models/alexnet.chainermodel")
130
131    caffe_model = load_models("../models/alexnet.chainermodel")
132    chainer_model = models.Alex.Alex()
133
134    # get label dict
135    label_d = create_label_list("../data/imagenet_label.txt")
136
137    # copy caffe_model W, b to chainer_model
138    copy_model(caffe_model, chainer_model)
139
140    # create mean image array
141    mean_image_array = create_mean_image_array(IMAGENET_MEAN_FILE, INPUT_IMAGE_SIZE)
142
143    # predict target_image
144    orig_img = resize_image("../data/panda2.jpeg")
145    orig_img.show()
146
147    orig_array = np.asarray(orig_img)
148    orig_array = substract_mean_image(orig_array, mean_image_array)
149
150    chainer_array = format2chainer(orig_array)
151
152    # apply gradient sign method
153    #adv_array, adv_part_array, orig_result = fast_gradient(chainer_model, chainer_array, eps=8.0)
154
155    # apply iterative gradient sign method
156    #adv_array, adv_part_array, orig_result = iterative_gradient(chainer_model, chainer_array,
157    #                                                                    eps=8.0, alpha=1.0)
158
159    # apply iterative least likely class method
160    adv_array, adv_part_array, orig_result = iterative_least_likely(chainer_model, chainer_array,
161                                                                    eps=8.0, alpha=1.0)
162    least_ind = np.argmin(orig_result)
163    print("least likely category {}".format(label_d[least_ind]))
164
165    # predict original image_result
166    orig_prob, orig_ind, orig_label = get_result(orig_result, label_d)
167    print("predict_original_image: {} predict_prob: {}".format(orig_label.strip(" "), orig_prob))
168
169    # predict adversarial_image
170    predict_result = F.softmax(chainer_model(adv_part_array)).data
171    part_prob, part_label_ind, part_label = get_result(predict_result, label_d)
172    print("predict_adversarial_perturbations: {} predict_prob: {}".format(part_label, part_prob))
173
174    predict_result = F.softmax(chainer_model(adv_array)).data
175    adv_prob, adv_label_ind, adv_label = get_result(predict_result, label_d)
176    print("predict_adversarial_examples: {} predict_prob: {}".format(adv_label, adv_prob))
177    print("original category prob with adv_images {}".format(predict_result[0][orig_ind]))
178
179    # show adv_image
180    adv_array = format2orig(adv_array[0])
181    adv_part_array = format2orig(adv_part_array[0])
182    adv_array = add_mean_image(adv_array, mean_image_array)
183    adv_array = np.clip(adv_array, 0, 255)
184    adv_array = adv_array.astype(np.uint8)
185    Image.fromarray(adv_array, 'RGB').show()
186    Image.fromarray(adv_part_array, 'RGB').show()