太長不看版
1、將PIL.Image轉換成批訓練的DataLoader
- 爲什麼一批一批進去處理
2、載入網絡（ torch.load('Resnet.pkl') ），並將數據放入網絡，通過 outputs = model(images) 得到預測值，放在對應的對象中
3、按對象的數字順序排序，生成熱力圖保存
耐心看完版
事先準備：
test_path = r'C:\Users\BME419\Desktop\resnet\slide\patch'
background_path = r'E:\WSI\CAMELYON16\Processed\patch-based-classification\raw-data\test\tumor091heatmaps patches\none'
pre_savename = r'C:\Users\BME419\Desktop\resnet\slide\heatmap'
savename = os.path.join(pre_savename, 'heatmap') #保存名字爲'heatmap'
batch_size = 64
classes = ['negative','positive']
global positive_prob
positive_prob = [] #positive_prob類型爲list
def reload_net(model_name): #可選擇四種網絡
if model_name == "VGG":
trainednet = torch.load('VGGnet.pkl')
elif model_name == "Google":
trainednet = torch.load('Google.pkl')
elif model_name == "Res":
trainednet = torch.load('Resnet.pkl')
elif model_name == "Alex":
trainednet = torch.load('Alexnet.pkl')
return trainednet
主函數：
test_data("Res", 224) #內涵調用heatmap_gen()
具體函數 ↓ ↓
def test_data(model_name, input_size): # input_size = 224
# 先轉換成 torch 能識別的 Dataset
testset = torchvision.datasets.ImageFolder(test_path,
transform = transforms.Compose([
transforms.Resize((input_size, input_size)),
# 將圖片縮放到指定大小（h,w）或者保持長寬比並縮放最短的邊到int大小
transforms.ToTensor(),
]))
# 把 dataset 放入 DataLoader
testloader = torch.utils.data.DataLoader(testset, batch_size = batch_size,
shuffle = False, num_workers = 0) # shuffle = False（不打亂）,按順序取patch，否則 = True,隨機取
model = reload_net(model_name) # load模型，函數具體定義見“事先準備”
model.eval() # 把BN和Dropout固定住，不會取平均，而是用訓練好的值
#將testset.imgs從tuple變爲list（用於append網絡產生的概率（outputs））
for j in range(len(testset.imgs)):
testset.imgs[j] = list(testset.imgs[j])
#利用訓練好的網絡預測patch概率
for i, data in enumerate(testloader, 0):
images, labels = data
print(labels)
images = Variable(images, requires_grad=True) # 轉換數據格式用Variable
if torch.cuda.is_available():
images = images.cuda() # 轉換數據格式用Variable
model.cuda()
with torch.no_grad():
outputs = model(images)
outputs = outputs.cpu().numpy() # 將outputs由GPU轉化爲numpy
positive_prob.extend(outputs[:, 1]) # positive_prob（list）:各個patch腫瘤positive的概率
if i <= (len(testset.imgs) / 64-1): # 將testset.imgs與各自概率一一對應
for j in range(64):
testset.imgs[j + 64 * i].append(outputs[j, 1])
else:
for j in range(len(testset.imgs) - 64 * i):
testset.imgs[j + 64 * i].append(outputs[j, 1])
probmin = np.min(positive_prob) # 用於背景顯示
heatmap_gen(positive_prob, testset,input_size,probmin) #調用heatmap_gen（），具體見↓
def heatmap_gen(positive_prob,testset,input_size,probmin):
fig = plt.figure(figsize=(172, 153)) #figsize以英寸爲單位 width=172，height=153
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, wspace=0, hspace=0)
# 讀背景patch
testset2 = torchvision.datasets.ImageFolder(background_path,
transform=transforms.Compose([
transforms.Resize((input_size, input_size)),
# 將圖片縮放到指定大小（h,w）或者保持長寬比並縮放最短的邊到int大小
transforms.ToTensor(), # 把一個取值範圍是[0,255]的PIL.Image 轉換成 Tensor
]))
for n in range(len(testset2.imgs)): # 將背景patch概率減10，並append到testset
testset2.imgs[n] = list(testset2.imgs[n])
testset2.imgs[n].append(probmin-10)
testset.imgs.append(testset2.imgs[n])
#使路徑減成數字，排序
for n in range(len(testset.imgs)):
testset.imgs[n][0] = os.path.basename(testset.imgs[n][0])
testset.imgs.sort(key=lambda x: int(x[0][:-6]))
positive_prob = [None] * len(testset.imgs)
positive_prob = np.array(positive_prob)
# 按照片名從小到大生成positive_prob
for n in range(len(testset.imgs)):
positive_prob[n] = testset.imgs[n][2]
probmin = np.min(positive_prob)
probmax = np.max(positive_prob)
heatmap = ((positive_prob - probmin) / (probmax - probmin + 0.000001)) * 255 # float在[0，1]之間，轉換成0-255
heatmap = heatmap.astype(np.uint8) # 轉成unit8
heatmap = heatmap.reshape(153, 172) # y爲行，x爲列
heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET) # 生成heat map
heatmap = heatmap[:, :, ::-1] # 注意cv2（BGR）和matplotlib(RGB)通道是相反的
plt.imshow(heatmap)
fig.savefig(savename, dpi=10) #dpi指每英寸有多少個像素，save路徑見最上面