1. import torch
  2. import torch.nn as nn
  3. import os
  4. from torchvision.models.resnet import *
  5. import numpy as np
  6. import matplotlib.pyplot as plt
  7. import time
  8. import cv2
  9. from visualize import visualize_grid_attention_v2
  11. class SpatialAttention(nn.Module):
  12.     def __init__(self, kernel_size=7):
  13.         super(SpatialAttention, self).__init__()
  15.         assert kernel_size in (3, 7), 'kernel size must be 3 or 7'
  16.         padding = 3 if kernel_size == 7 else 1
  18.         self.conv1 = nn.Conv2d(2, 1, kernel_size, padding=padding, bias=False)
  19.         self.sigmoid = nn.Sigmoid()
  21.     def forward(self, x):
  22.         avg_out = torch.mean(x, dim=1, keepdim=True)
  23.         max_out, _ = torch.max(x, dim=1, keepdim=True)
  24.         x = torch.cat([avg_out, max_out], dim=1)
  25.         x = self.conv1(x)
  26.         return self.sigmoid(x)
  27. def draw_features(width,height,x,savename):
  28.     tic=time.time()
  29.     fig = plt.figure(figsize=(16, 16))
  30.     fig.subplots_adjust(left=0.05, right=0.95, bottom=0.05, top=0.95, wspace=0.05, hspace=0.05)
  31.     for i in range(width*height):
  32.         plt.subplot(height,width, i + 1)
  33.         plt.axis('off')
  34.         img = x[0, i, :, :]
  35.         pmin = np.min(img)
  36.         pmax = np.max(img)
  37.         img = ((img - pmin) / (pmax - pmin + 0.000001))*255  #float?[0?1]??????0-255
  38.         img=img.astype(np.uint8)  #??unit8
  39.         img=cv2.applyColorMap(img, cv2.COLORMAP_JET) #??heat map
  40.         img = img[:, :, ::-1]#??cv2?BGR??matplotlib(RGB)??????
  41.         plt.imshow(img)
  42.         print("{}/{}".format(i,width*height))
  43.     fig.savefig(savename, dpi=100)
  44.     fig.clf()
  45.     plt.close()
  46.     print("time:{}".format(time.time()-tic))
  48. make_pic = False
  49. n_classes = 2
  50. class Resnet50_att(nn.Module):
  51.     def __init__(self,wsl_path,savepath=None):
  52.         super(Resnet50_att, self).__init__()
  53.         print("model: resnet50")
  54.         checkpoint = os.path.join(wsl_path, 'resnet50_wsl.pth')
  55.         model = resnet50()
  56.         state_dict = torch.load(checkpoint)
  57.         model.load_state_dict(state_dict)
  58.         model.fc = torch.nn.Linear(2048, n_classes)
  59.         self.model = model
  60.         self.savepath = savepath
  61.         # ?????????????
  62.         # self.ca = ChannelAttention(self.inplanes)
  63.         self.sa_1 = SpatialAttention()
  64.         # ??????????????????
  65.         self.sa_2 = SpatialAttention()
  68.     def forward(self, x):
  69.         origin_img = x
  70.         att1 = 0
  71.         att2 = 0
  72.         #这段代码仅仅只使用了else后面的内容
  73.         if self.savepath:  # draw features or not
  74.             x = self.model.conv1(x)
  75.             draw_features(8, 8, x.cpu().numpy(), "{}/f1_conv1.png".format(self.savepath))
  77.             x = self.model.bn1(x)
  78.             draw_features(8, 8, x.cpu().numpy(), "{}/f2_bn1.png".format(self.savepath))
  80.             x = self.model.relu(x)
  81.             draw_features(8, 8, x.cpu().numpy(), "{}/f3_relu.png".format(self.savepath))
  83.             x = self.model.maxpool(x)
  84.             draw_features(8, 8, x.cpu().numpy(), "{}/f4_maxpool.png".format(self.savepath))
  86.             x = self.model.layer1(x)
  87.             draw_features(16, 16, x.cpu().numpy(), "{}/f5_layer1.png".format(self.savepath))
  89.             x = self.model.layer2(x)
  90.             draw_features(16, 32, x.cpu().numpy(), "{}/f6_layer2.png".format(self.savepath))
  92.             x = self.model.layer3(x)
  93.             draw_features(32, 32, x.cpu().numpy(), "{}/f7_layer3.png".format(self.savepath))
  95.             x = self.model.layer4(x)
  96.             draw_features(32, 32, x.cpu().numpy()[:, 0:1024, :, :], "{}/f8_layer4_1.png".format(self.savepath))
  97.             draw_features(32, 32, x.cpu().numpy()[:, 1024:2048, :, :], "{}/f8_layer4_2.png".format(self.savepath))
  99.             x = self.model.avgpool(x)
  100.             plt.plot(np.linspace(1, 2048, 2048), x.cpu().numpy()[0, :, 0, 0])
  101.             plt.savefig("{}/f9_avgpool.png".format(self.savepath))
  102.             plt.clf()
  103.             plt.close()
  105.             x = x.view(x.size(0), -1)
  106.             x = self.model.fc(x)
  107.             plt.plot(np.linspace(1, 1000, 1000), x.cpu().numpy()[0, :])
  108.             plt.savefig("{}/f10_fc.png".format(self.savepath))
  109.             plt.clf()
  110.             plt.close()
  111.         else:
  112.             x = self.model.conv1(x)
  113.             x = self.model.bn1(x)
  114.             x = self.model.relu(x)
  115.             att1 = self.sa_1(x)
  116.             x = att1 * x
  117.             # x = self.sa_1(x) * x
  118.             x = self.model.maxpool(x)
  119.             x = self.model.layer1(x)
  120.             # att2 = self.sa_2(x)
  121.             # x = att2 * x
  122.             x = self.model.layer2(x)
  123.             x = self.model.layer3(x)
  124.             x = self.model.layer4(x)
  126.             # x = self.sa_2(x) * x
  127.             x = self.model.avgpool(x)
  128.             x = x.view(x.size(0), -1)
  129.             x = self.model.fc(x)
  130.         #return x
  131.         return x,att1
  132. #预训练权重文件的地址
  133. wsl_path=""
  134. model = Resnet50_att(wsl_path=wsl_path)
  135. inputs = ""
  136. outs，att = model(inputs)
  137. att1 = att.detach().cpu().numpy()
  140. save_path_1 = "/media/zhujunjie/dataset/attVision/att1"
  142. for index,image_id in enumerate(img_ids):
  143.     img_path = os.path.join(dir_train_img, image_id + '.jpg')
  146.     visualize_grid_attention_v2(img_path,
  147.                                 save_path=save_path_1,
  148.                                 attention_mask=att1[index][0],
  149.                                 save_image=True,
  150.                                 save_original_image=True,
  151.                                 quality=100)

清洗数据

  1.     for step, batch in enumerate(valloader):
  2.         if step % 1000 == 0:
  3.             logger.info('Valid step {} of {}'.format(step, len(valloader)))
  4.         inputs = batch["image"]
  5.         inputs = inputs.to(device, dtype=torch.float)
  7.         outputs = model(inputs,meta_inputs)
  8.         sfx = nn.Softmax(dim=-1)(outputs)
  9.         sfx_cpu = sfx.detach().cpu().numpy()
  10.         # sfx = sfx.detach().cpu().numpy()
  11.         # ???0??????????1??????
  12.         predict_per = sfx.argmax(dim=1)
  13.         prob = sfx.max(dim=-1, keepdim=False)[1].detach().cpu().numpy()
  14.         truel = batch['target'].detach().cpu().numpy()
  15.         #筛选出错误分类，p1-p0>0表示将负例分类成正例，反之，就是正例分类成正例
  16.         for i in range(len(predict_per)):
  17.             if predict_per[i] != truel[i]:
  18.                 a = {'image_id': batch['image_id'][i], 'p1-p0': sfx_cpu[i][1] - sfx_cpu[i][0]}
  19.                 fake_predict = fake_predict.append(a, ignore_index=True)
  21.         prob_all.extend(prob)
  22.         lable_all.extend(truel)
  23.     probability_outfile = os.path.join(WORK_DIR,
  24.                                        'probability/errprob_attUResnet50_ep{}_fold{}.csv'.format(epoch, fold))
  25.     fake_predict.to_csv(probability_outfile, index=False)

将输出的概率文件用excel打开，按照p1-p0列进行排序，选择最大和最小的一部分数据，复制到新的表中，存为csv文件: delImg.csv


# 删除csv中的部分记录

dir_train_img = '/media/zhujunjie/dataset/baMelanoma_data'
path_img = '/media/zhujunjie/dataset/baMelanoma_data/train'
train_csv_path = '/media/zhujunjie/dataset/train.csv'
del_csv_path = '/media/zhujunjie/dataset/delImg.csv'
del_train = pd.read_csv(del_csv_path)
trndf = pd.read_csv(train_csv_path)
print('del_csv_img shape {} {}'.format(*del_train.shape))
print('trn_csv_img shape {} {}'.format(*trndf.shape))
del_pngs = [img_id for img_id in del_train['image_id']]
print('Count of del_pngs : {}'.format(len(del_pngs)))
#删除操作
trndf = trndf.set_index('image_id')
trndf.drop(index=del_pngs,axis=0,inplace=True)
#恢复索引
trndf = trndf.reset_index()
#保存删除了部分数据的新标签文件
print('new_csv_img shape {} {}'.format(*trndf.shape))
trndf.to_csv(os.path.join(dir_train_img,"new_train.csv"), index=False)