import itertools import os import time import torch import cv2 import numpy as np from backbones import iresnet50,iresnet18,iresnet100 # print(torch.cuda.is_available()) def load_image(img_path): img = cv2.imread(img_path) #print(img) #img2 = np.fliplr(img) #print(img2.shape) img = img.transpose((2, 0, 1)) #img2 = img2.transpose((2, 0, 1)) img = img[np.newaxis, :, :, :] #img2 = img2[np.newaxis, :, :, :] #img = np.concatenate((img,img2)) img = np.array(img, dtype=np.float32) img -= 127.5 img /= 127.5 return img def findEuclideanDistance(source_representation, test_representation): euclidean_distance = source_representation - test_representation euclidean_distance = np.sum(np.multiply(euclidean_distance, euclidean_distance)) euclidean_distance = np.sqrt(euclidean_distance) return euclidean_distance def findCosineDistance(source_representation, test_representation): a = np.matmul(np.transpose(source_representation), test_representation) b = np.sum(np.multiply(source_representation, source_representation)) c = np.sum(np.multiply(test_representation, test_representation)) return 1 - (a / (np.sqrt(b) * np.sqrt(c))) def l2_normalize(x): return x / np.sqrt(np.sum(np.multiply(x, x))) def cosin_metric(x1, x2): return np.dot(x1, x2) / (np.linalg.norm(x1) * np.linalg.norm(x2)) def get_file_list(file_name): file1 = [] file2 = [] code = [] with open(file_name,"r") as f: file_list = f.readlines() for i in file_list: data = i.replace("\n","").split(" ") file1.append(data[0]) file2.append(data[1]) code.append(data[2]) return file1,file2,code def load_list_images(file_list): all = np.zeros((len(file_list),3,112,112),dtype=np.float32) count = 0 for img_path in file_list: img = cv2.imread(img_path) img = img.transpose((2, 0, 1)) img = img[np.newaxis, :, :, :] all[count] = img count = count + 1 all -= 127.5 all /= 127.5 return all def threshold_acc(distance,code): #thresholds = np.arange(24,36,0.1) thresholds = np.arange(0.5,2, 0.1) total = len(code) max_acc = 0 threshold = 0 code_net = [] for i in thresholds: for ds in distance: if ds < i: code_net.append(1) else: code_net.append(0) right = np.sum(code==code_net) #print(threshold, max_acc) if right/total > max_acc: max_acc = right / total threshold = i code_net = [] return threshold,max_acc def reject_and_error(distance,code): total = len(code) threshold = 1.20 code_net = [] reject = 0 error = 0 for ds in distance: if ds < threshold: code_net.append(1) else: code_net.append(0) right = np.sum(code==code_net) max_acc = right / total #print(max_acc,threshold) for i in range(total): if code[i] == 1: if code_net[i] == 0: reject += 1 else: if code_net[i] == 1: error += 1 return reject,error,right # if __name__=='__main__': # img1 = load_image("D:\Download\out\centerface_test\Adolfo_Rodriguez_Saa\Adolfo_Rodriguez_Saa_0002.jpg") # img1 = torch.from_numpy(img1) # #print(img1.shape) # #print(img1) # img2 = load_image("D:\Download\out\centerface_database\Dario_Franchitti\Dario_Franchitti_0001.jpg") # img2 = torch.from_numpy(img2) # model = iresnet100() # model.load_state_dict(torch.load("./model/backbone100.pth",map_location="cpu")) # #print(model) # model.eval() # with torch.no_grad(): # start_time = time.time() # pred1 = model(img1) # #print(pred1.shape) # #print(pred1) # pred2 = model(img2) # pred1 = pred1.numpy() # #print(pred1.shape) # pred2 = pred2.numpy() # distance = findEuclideanDistance(l2_normalize(pred1),l2_normalize(pred2)) # end_time = time.time() # print("EuclideanDistance is :" + str(distance)) # if distance < 1.20: # print("对比两张相同的人脸用时："+str(end_time - start_time)) # else: # print("对比两张不同的人脸用时：" + str(end_time - start_time)) if __name__=='__main__': model = iresnet100() model.load_state_dict(torch.load("./model/backbone100.pth", map_location="cpu")) model.eval() file1, file2, code = get_file_list("pairs.txt") code = np.array(code,dtype=np.int) img = load_list_images(file1) print(img.shape) img = torch.from_numpy(img) img2 = load_list_images(file2) img2 = torch.from_numpy(img2) batch_size = 128 now = 0 number = len(code) distance = [] with torch.no_grad(): start_time = time.time() while now < number: if now + batch_size < number: pred1 = model(img[now:now + batch_size]) pred2 = model(img2[now:now + batch_size]) else: pred1 = model(img[now:]) pred2 = model(img2[now:]) now = now + batch_size pred1 = pred1.numpy() pred2 = pred2.numpy() print("batch"+str(now)) for i in range(len(pred1)): #distance.append(findEuclideanDistance(pred1[i], pred2[i])) distance.append(findEuclideanDistance(l2_normalize(pred1[i]), l2_normalize(pred2[i]))) end_time = time.time() distance = np.array(distance) print("对比两张人脸平均用时：" + str((end_time - start_time) / number)) #print(distance) # threshold,max_acc = threshold_acc(distance,code) # print(threshold,max_acc) reject, error,right = reject_and_error(distance, code) print("正确人脸对数量："+str(right)+" 正确率："+str(right/number)) print("拒识人脸对数量：" + str(reject) + " 拒识率：" + str(2*reject / number)) print("误识人脸对数量：" + str(error) + " 误识率：" + str(2 * error / number)) # sublist = ['zdz','bbb','aaa'] # for item in itertools.combinations(sublist, 2): # for name in item: # print(name) # path = "D:\Download\out\output2\Alan_Trammell\Alan_Trammell_0001.jpg" # path = path.split("\\") # print(path) # print(os.path.join(path[0],path[1],path[2],path[3],path[4]))