CIPL/interpretable_reasoning.py at master · cwangrun/CIPL · GitHub

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
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
import os
import torch
import torch.utils.data
import re
from PIL import Image
import cv2
import matplotlib.pyplot as plt
from helpers import makedir
import model
from utlis.utlis_func import *
from matplotlib import cm
import pandas as pd


Labels_dict = {
    "No Finding": 14,
    "Atelectasis": 0,
    "Cardiomegaly": 1,
    "Effusion": 2,
    "Infiltrate": 3,
    "Mass": 4,
    "Nodule": 5,
    "Pneumonia": 6,
    "Pneumothorax": 7,
    "Consolidation": 8,
    "Edema": 9,
    "Emphysema": 10,
    "Fibrosis": 11,
    "Pleural_Thickening": 12,
    "Hernia": 13,
}


def distance_2_similarity_exp(distances):
    return torch.exp(-distances / 256.0)  # 128.0


def ind2name(jjj):
    save_name = str(jjj + 1)
    if jjj < 9:
        save_name = '0' + save_name
    else:
        save_name = save_name
    return save_name


def get_one_heatmap(similarity_map, gt_box, img_np):

    original_img_size2, original_img_size1 = img_np.shape[0], img_np.shape[1]

    proto_act_img_j = similarity_map.squeeze().detach().cpu().numpy()

    upsampled_act_img_j = cv2.resize(proto_act_img_j, dsize=(original_img_size2, original_img_size1), interpolation=cv2.INTER_CUBIC)
    rescaled_act_img_j = (upsampled_act_img_j - np.amin(upsampled_act_img_j)) / (np.amax(upsampled_act_img_j) - np.amin(upsampled_act_img_j))

    proto_bound_j = gt_box

    heatmap = cv2.applyColorMap(np.uint8(255 * rescaled_act_img_j), cv2.COLORMAP_JET)
    heatmap = np.float32(heatmap) / 255
    heatmap = heatmap[..., ::-1]

    original_img_j = np.array(img_np).astype(float) / 1.0
    overlayed_original_img_j = 0.7 * original_img_j + 0.3 * heatmap
    img_bgr_uint8 = cv2.cvtColor(np.uint8(255 * overlayed_original_img_j), cv2.COLOR_RGB2BGR)

    cv2.rectangle(img_bgr_uint8, (proto_bound_j[2], proto_bound_j[0]), (proto_bound_j[3] - 1, proto_bound_j[1] - 1), color=(100, 255, 100), thickness=2)

    img_rgb_uint8 = img_bgr_uint8[..., ::-1]
    img_rgb_float = np.float32(img_rgb_uint8) / 255

    return img_rgb_float, rescaled_act_img_j


from settings import base_architecture, img_size, prototype_shape, num_classes, \
    prototype_activation_function, add_on_layers_type, root_dir

base_architecture_type = re.match('^[a-z]*', base_architecture).group(0)


# construct the model
ppnet = model.construct_CIPL(base_architecture=base_architecture,
                             pretrained=True, img_size=img_size,
                             prototype_shape=prototype_shape,
                             num_classes=num_classes,
                             prototype_activation_function=prototype_activation_function,
                             add_on_layers_type=add_on_layers_type)
ppnet = ppnet.cuda()

checkpoint_path = "14nopush0.8232.pth"

ppnet.load_state_dict(torch.load(checkpoint_path))
model = torch.nn.DataParallel(ppnet)
model.eval()

for p in model.module.parameters():
    p.requires_grad = False
connection_weight_allclass = model.module.last_layer.weight.squeeze()

transform = transforms.Compose([
            transforms.Resize([img_size, img_size]),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
            ])
transform_ori = transforms.Compose([
            transforms.Resize([img_size, img_size]),
            transforms.ToTensor(),
            ])

gr_path = os.path.join(root_dir, "BBox_List_2017.csv")
df_box = pd.read_csv(gr_path)

save_dir_main = './output_vis'
makedir(save_dir_main)


image_name = '00028027_000.png'
target_name = "Mass"


if __name__ == '__main__':

    img_idx = df_box.loc[df_box['Image Index'] == image_name].index
    box_x = int(df_box.iloc[img_idx]['Bbox [x'] / 2.0)
    box_y = int(df_box.iloc[img_idx]['y'] / 2.0)
    box_w = int(df_box.iloc[img_idx]['w'] / 2.0)
    box_h = int(df_box.iloc[img_idx]['h]'] / 2.0)
    gt_box = (box_y, box_y + box_h, box_x, box_x + box_w)
    label_gt = Labels_dict[target_name]

    img_PIL = Image.open(os.path.join(root_dir, 'data', image_name)).convert('RGB')
    img_input = transform(img_PIL)

    with torch.no_grad():
        img_input = img_input.unsqueeze(0).cuda()
        target = torch.tensor(label_gt).cuda()

        output_all, min_distances_all, similarities_all = model.module.forward_infer(img_input)

        # compute predictions
        output_fg = output_all[:, 0:num_classes-1]
        output_bg = output_all[:, -1].unsqueeze(1).repeat(1, num_classes-1)  # no finding is the last group of prototypes
        output_new = torch.stack((output_bg, output_fg), dim=-1)
        prob = torch.softmax(output_new.squeeze(0), dim=1)[label_gt][1]
        print(image_name, 'prob:', round(prob.item(), 3))

        img_np = np.array(transform_ori(img_PIL)).transpose(1, 2, 0).astype(np.float32)
        img_np = (img_np - img_np.min()) / (img_np.max() - img_np.min())

        original_img_size1, original_img_size2 = img_np.shape[0], img_np.shape[1]
        num_proto_per_class = model.module.num_prototypes // model.module.num_classes  # 50
        proto_index_start = num_proto_per_class * label_gt
        proto_index_end = num_proto_per_class * label_gt + num_proto_per_class

        similarity_maps = similarities_all[0, proto_index_start:proto_index_end]
        min_distances = min_distances_all[0, proto_index_start:proto_index_end]
        connection_weight = connection_weight_allclass[label_gt, proto_index_start:proto_index_end]

        # _, sorted_index = torch.sort(min_distances * connection_weight, descending=False)
        _, sorted_index = torch.sort(min_distances, descending=False)
        similarity = distance_2_similarity_exp(min_distances)

        for iii, index in enumerate(sorted_index.cpu().numpy()):
            score = np.around(similarity[index].cpu().numpy(), decimals=3)
            weight = np.around(connection_weight[index].cpu().numpy(), decimals=3)
            overlap_map, heatmap = get_one_heatmap(similarity_maps[index], gt_box, img_np)
            save_dir = os.path.join(save_dir_main, target_name)
            makedir(save_dir)
            save_path = os.path.join(save_dir, image_name.replace('.png', '_' + ind2name(iii) + '_overlap.jpg'))
            plt.imsave(save_path, overlap_map, vmin=0.0, vmax=1.0, pil_kwargs={'quality': 99})

        save_path = os.path.join(save_dir, image_name.replace('.png', '_ori.png'))
        plt.imsave(save_path, img_np, vmin=0.0, vmax=1.0, cmap=cm.gray, pil_kwargs={'quality': 99})