def main():
lfw_dataroot = args.lfw
model_path = args.model_path
far_target = args.far_target
checkpoint = torch.load(model_path)
model = Resnet18Triplet(embedding_dimension=checkpoint['embedding_dimension'])
model.load_state_dict(checkpoint['model_state_dict'])
flag_gpu_available = torch.cuda.is_available()
if flag_gpu_available:
device = torch.device("cuda")
else:
device = torch.device("cpu")
lfw_transforms = transforms.Compose([
transforms.Resize(size=224),
transforms.ToTensor(),
transforms.Normalize(
mean=[0.5157, 0.4062, 0.3550],
std=[0.2858, 0.2515, 0.2433]
)
])
lfw_dataloader = torch.utils.data.DataLoader(
dataset=LFWDataset(
dir=lfw_dataroot,
pairs_path='../datasets/LFW_pairs.txt',
transform=lfw_transforms
),
batch_size=256,
num_workers=2,
shuffle=False
)
model.to(device)
model = model.eval()
with torch.no_grad():
l2_distance = PairwiseDistance(2).cuda()
distances, labels = [], []
progress_bar = enumerate(tqdm(lfw_dataloader))
for batch_index, (data_a, data_b, label) in progress_bar:
data_a, data_b, label = data_a.cuda(), data_b.cuda(), label.cuda()
output_a, output_b = model(data_a), model(data_b)
distance = l2_distance.forward(output_a, output_b) # Euclidean distance
distances.append(distance.cpu().detach().numpy())
labels.append(label.cpu().detach().numpy())
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array([subdist for distance in distances for subdist in distance])
_, _, _, _, _, _, _, tar, far = evaluate_lfw(distances=distances, labels=labels, far_target=far_target)
print("TAR: {:.4f}+-{:.4f} @ FAR: {:.4f}".format(np.mean(tar), np.std(tar), np.mean(far)))
def set_model_architecture(model_architecture, pretrained,
embedding_dimension):
if model_architecture == "resnet18":
model = Resnet18Triplet(embedding_dimension=embedding_dimension,
pretrained=pretrained)
elif model_architecture == "resnet34":
model = Resnet34Triplet(embedding_dimension=embedding_dimension,
pretrained=pretrained)
elif model_architecture == "resnet50":
model = Resnet50Triplet(embedding_dimension=embedding_dimension,
pretrained=pretrained)
elif model_architecture == "resnet101":
model = Resnet101Triplet(embedding_dimension=embedding_dimension,
pretrained=pretrained)
elif model_architecture == "resnet152":
model = Resnet152Triplet(embedding_dimension=embedding_dimension,
pretrained=pretrained)
elif model_architecture == "inceptionresnetv2":
model = InceptionResnetV2Triplet(
embedding_dimension=embedding_dimension, pretrained=pretrained)
elif model_architecture == "mobilenetv2":
model = MobileNetV2Triplet(embedding_dimension=embedding_dimension,
pretrained=pretrained)
print("Using {} model architecture.".format(model_architecture))
return model
def main():
lfw_dataroot = args.lfw
model_path = args.model_path
far_target = args.far_target
batch_size = args.batch_size
flag_gpu_available = torch.cuda.is_available()
if flag_gpu_available:
device = torch.device("cuda")
print('Using GPU')
else:
device = torch.device("cpu")
print('Using CPU')
checkpoint = torch.load(model_path, map_location=device)
model = Resnet18Triplet(
embedding_dimension=checkpoint['embedding_dimension'])
model.load_state_dict(checkpoint['model_state_dict'])
# desiredFaceWidth, height and desiredLeftEye work together to produce centered aligned faces
# desiredLeftEye (x,y) says how the face moves in the heightxwidth window. bigger y moves the face down
# bigger x zooms-out the face and severe values migh flip the face upside down, you wan it in range [0.2;0.36]
desiredFaceHeight = 448 # set non-equal width, height so the tf_resize stretch the faces, select bigger than 224 values so we don't lose too much of the image quality
desiredFaceWidth = 352
desiredLeftEye = (0.28, 0.35)
tf_align = transform_align(
landmark_predictor_weight_path=landmark_predictor_weights,
face_detector_path=face_detector_weights_path,
desiredFaceWidth=desiredFaceWidth,
desiredFaceHeight=desiredFaceHeight,
desiredLeftEye=desiredLeftEye)
lfw_transforms = transforms.Compose([
tf_align,
transforms.Resize(size=(224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.6068, 0.4517, 0.3800],
std=[0.2492, 0.2173, 0.2082])
])
lfw_dataloader = torch.utils.data.DataLoader(
dataset=LFWDataset(dir=lfw_dataroot,
pairs_path='../datasets/LFW_pairs.txt',
transform=lfw_transforms),
batch_size=
batch_size, # default = 256; 160 - allows running under 2GB VRAM
num_workers=2,
shuffle=False)
model.to(device)
model = model.eval()
with torch.no_grad():
l2_distance = PairwiseDistance(p=2)
distances, labels = [], []
progress_bar = enumerate(tqdm(lfw_dataloader))
for batch_index, (data_a, data_b, label) in progress_bar:
data_a = data_a.to(device) # data_a = data_a.cuda()
data_b = data_b.to(device) # data_b = data_b.cuda()
output_a, output_b = model(data_a), model(data_b)
distance = l2_distance.forward(output_a,
output_b) # Euclidean distance
distances.append(distance.cpu().detach().numpy())
labels.append(label.cpu().detach().numpy())
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array(
[subdist for distance in distances for subdist in distance])
_, _, _, _, _, _, _, tar, far = evaluate_lfw(distances=distances,
labels=labels,
far_target=far_target)
print("TAR: {:.4f}+-{:.4f} @ FAR: {:.4f}".format(
np.mean(tar), np.std(tar), np.mean(far)))