def feature_extract(data_list, extract_folder):
f = open(data_list, 'r')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
INPUT_SIZE = [112, 112]
backbone_path = './backbone/CurricularFace_Backbone.pth'
backbone_model = IR_101(INPUT_SIZE).to(device)
checkpoint = torch.load(backbone_path,
map_location=lambda storage, loc: storage)
if 'state_dict' not in checkpoint:
backbone_model.load_state_dict(checkpoint)
backbone_model.eval()
data_transform = transforms.Compose([
transforms.Resize((INPUT_SIZE[0], INPUT_SIZE[1])),
transforms.ToTensor(),
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])
for file_path in tqdm.tqdm(f):
file_path = file_path.replace('\n', '')
img_tensor = read_img(file_path, data_transform)
img_tensor = img_tensor.unsqueeze(0).to(device)
with torch.no_grad():
mu, conv_final = backbone_model(img_tensor)
mu = l2_norm(mu)
mu = mu.cpu().data.numpy()
conv_final = conv_final.cpu().data.numpy()
sub_dir = file_path.split('/')[-2]
org_folder = file_path.split('/')[2]
extract_dir = extract_folder.split('/')[2]
# print('org_folder:', org_folder)
# print('extract_dir:', extract_dir)
new_dir = extract_folder + sub_dir
if not os.path.isdir(new_dir):
os.mkdir(new_dir)
new_file_path = file_path.replace(org_folder, extract_dir)
mu_file_path = new_file_path.replace('.jpg', '_mu.npy')
conv_final_file_path = new_file_path.replace('.jpg', '_conv_final.npy')
# print(new_dir)
# print(mu_file_path)
# print(conv_final_file_path)
np.save(mu_file_path, mu)
np.save(conv_final_file_path, conv_final)
f.close()
def __init__(self, device, backbone_name = cfg['BACKBONE_NAME'], INPUT_SIZE = cfg['INPUT_SIZE'], BACKBONE_RESUME_ROOT = cfg['BACKBONE_RESUME_ROOT']):
super().__init__()
BACKBONE_DICT = {'ResNet_50': ResNet_50(INPUT_SIZE),
'ResNet_101': ResNet_101(INPUT_SIZE),
'ResNet_152': ResNet_152(INPUT_SIZE),
'IR_50': IR_50(INPUT_SIZE),
'IR_101': IR_101(INPUT_SIZE),
'IR_152': IR_152(INPUT_SIZE),
'IR_SE_50': IR_SE_50(INPUT_SIZE),
'IR_SE_101': IR_SE_101(INPUT_SIZE),
'IR_SE_152': IR_SE_152(INPUT_SIZE)}
self.device = device
self.embedding = BACKBONE_DICT[backbone_name]
self.embedding.load_state_dict(torch.load(BACKBONE_RESUME_ROOT))
self.embedding = self.embedding.to(device)
def eval(data_path, file_name):
pairs, dirFiles = generate_pair(data_path, file_name)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
B = 1
INPUT_SIZE = [112, 112]
backbone_path = './backbone/CurricularFace_Backbone.pth'
backbone_model = IR_101(INPUT_SIZE).to(device)
checkpoint = torch.load(backbone_path, map_location=lambda storage, loc: storage)
if 'state_dict' not in checkpoint:
backbone_model.load_state_dict(checkpoint)
else:
print('use dict')
pretrained_weights = checkpoint['state_dict']
model_weights = backbone_model.state_dict()
pretrained_weights = {k: v for k, v in pretrained_weights.items() \
if k in model_weights}
model_weights.update(pretrained_weights)
backbone_model.load_state_dict(model_weights)
backbone_model.eval()
# Load model files and config file
config_file = 'config/sphere64_msarcface.py'
config = imp.load_source('config', config_file)
config.batch_format['size'] = 1
network = Network()
network.initialize(config)
network.load_model('log/sphere64_casia_am_PFE/20201028-015531')
data_transform = transforms.Compose([
transforms.Resize((INPUT_SIZE[0], INPUT_SIZE[1])),
transforms.ToTensor(),
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])
embeddings_org_list = []
embeddings_mu_list = []
embeddings_sigma_list = []
for i, (filename) in tqdm.tqdm(enumerate((dirFiles))):
start_time = time.time()
img_tensor = read_img(filename, data_transform)
img_tensor = img_tensor.unsqueeze(0).to(device)
with torch.no_grad():
mu, conv_final = backbone_model(img_tensor)
mu = l2_norm(mu)
mu = mu.cpu().data.numpy()
conv_final = conv_final.cpu().data.numpy()
embeddings_org_list.append(mu.reshape(-1))
mu, sigma_sq = network.extract_feature(mu, conv_final)
# print('mu:', mu.shape)
# print('sigma_sq:', sigma_sq.shape)
embeddings_mu_list.append(mu.reshape(-1))
embeddings_sigma_list.append(sigma_sq.reshape(-1))
embeddings_org = np.array(embeddings_org_list)
accuracy, threshold = evaluate(embeddings_org, pairs, nrof_folds=10)
print('Org Accuracy: %1.5f' % accuracy)
print('threshold: %1.5f' % threshold)
embeddings_mu = np.array(embeddings_mu_list)
embeddings_sigma = np.array(embeddings_sigma_list)
accuracy, threshold = evaluate(embeddings_mu, pairs, nrof_folds=10, compare_func=pair_MLS_score,
sigma_sq=embeddings_sigma)
print('MLS Accuracy: %1.5f' % accuracy)
print('threshold: %1.5f' % threshold)
NUM_CLASS = len(train_loader.dataset.classes)
print("Number of Training Classes: {}".format(NUM_CLASS))
lfw, cfp_ff, cfp_fp, agedb, calfw, cplfw, vgg2_fp, lfw_issame, cfp_ff_issame, cfp_fp_issame, agedb_issame, calfw_issame, cplfw_issame, vgg2_fp_issame = get_val_data(DATA_ROOT)
#======= model & loss & optimizer =======#
if BACKBONE_NAME == 'ResNet_50':
BACKBONE = ResNet_50(INPUT_SIZE)
# 'ResNet_101': resnet101(INPUT_SIZE),
# 'ResNet_152': resnet152(INPUT_SIZE),
elif BACKBONE_NAME == 'IR_50':
BACKBONE = IR_50(INPUT_SIZE)
elif BACKBONE_NAME == 'IR_101':
BACKBONE = IR_101(INPUT_SIZE)
elif BACKBONE_NAME == 'IR_152':
BACKBONE = IR_152(INPUT_SIZE)
elif BACKBONE_NAME == 'IR_SE_50':
BACKBONE = IR_SE_50(INPUT_SIZE)
elif BACKBONE_NAME == 'IR_SE_101':
BACKBONE = IR_SE_101(INPUT_SIZE)
elif BACKBONE_NAME == 'IR_SE_152':
BACKBONE = IR_SE_152(INPUT_SIZE)
elif BACKBONE_NAME == 'ShuffleNet':
BACKBONE = shufflenet(cfg=cfg)
elif BACKBONE_NAME == 'ShuffleNetV2':
BACKBONE = shufflenetv2(cfg=cfg)
elif BACKBONE_NAME == 'Mobilenet':
BACKBONE = mobilenet()
elif BACKBONE_NAME == 'Mobilenetv2':
num_workers=NUM_WORKERS,
drop_last=DROP_LAST)
NUM_CLASS = len(train_loader.dataset.classes)
print("Number of Training Classes: {}".format(NUM_CLASS))
lfw, cfp_ff, cfp_fp, agedb, calfw, cplfw, vgg2_fp, lfw_issame, cfp_ff_issame, cfp_fp_issame, agedb_issame, calfw_issame, cplfw_issame, vgg2_fp_issame = get_val_data(
DATA_ROOT)
# ======= model & loss & optimizer =======#
BACKBONE_DICT = {
'ResNet_50': ResNet_50(INPUT_SIZE),
'ResNet_101': ResNet_101(INPUT_SIZE),
'ResNet_152': ResNet_152(INPUT_SIZE),
'IR_50': IR_50(INPUT_SIZE),
'IR_101': IR_101(INPUT_SIZE),
'IR_152': IR_152(INPUT_SIZE),
'IR_SE_50': IR_SE_50(INPUT_SIZE),
'IR_SE_101': IR_SE_101(INPUT_SIZE),
'IR_SE_152': IR_SE_152(INPUT_SIZE)
}
BACKBONE = BACKBONE_DICT[BACKBONE_NAME]
print("=" * 60)
print(BACKBONE)
print("{} Backbone Generated".format(BACKBONE_NAME))
print("=" * 60)
HEAD_DICT = {
'ArcFace':
ArcFace(in_features=EMBEDDING_SIZE, out_features=NUM_CLASS),
'CosFace':
def main(ARGS):
if ARGS.model_path == None:
raise AssertionError("Path should not be None")
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
####### Model setup
print('Model type: %s' % ARGS.model_type)
if ARGS.model_type == 'ResNet_50':
model = ResNet_50(ARGS.input_size)
elif ARGS.model_type == 'ResNet_101':
model = ResNet_101(ARGS.input_size)
elif ARGS.model_type == 'ResNet_152':
model = ResNet_152(ARGS.input_size)
elif ARGS.model_type == 'IR_50':
model = IR_50(ARGS.input_size)
elif ARGS.model_type == 'IR_101':
model = IR_101(ARGS.input_size)
elif ARGS.model_type == 'IR_152':
model = IR_152(ARGS.input_size)
elif ARGS.model_type == 'IR_SE_50':
model = IR_SE_50(ARGS.input_size)
elif ARGS.model_type == 'IR_SE_101':
model = IR_SE_101(ARGS.input_size)
elif ARGS.model_type == 'IR_SE_152':
model = IR_SE_152(ARGS.input_size)
else:
raise AssertionError(
'Unsuported model_type {}. We only support: [\'ResNet_50\', \'ResNet_101\', \'ResNet_152\', \'IR_50\', \'IR_101\', \'IR_152\', \'IR_SE_50\', \'IR_SE_101\', \'IR_SE_152\']'
.format(ARGS.model_type))
if use_cuda:
model.load_state_dict(torch.load(ARGS.model_path))
else:
model.load_state_dict(torch.load(ARGS.model_path, map_location='cpu'))
model.to(device)
# embedding_size = 512
model.eval()
# DATA_ROOT = './../evoLVe_data/data' # the parent root where your train/val/test data are stored
# INPUT_SIZE = [112, 112] # support: [112, 112] and [224, 224]
# BACKBONE_RESUME_ROOT = './../evoLVe_data/pth/backbone_ir50_ms1m_epoch120.pth' # the root to resume training from a saved checkpoint
# BACKBONE_RESUME_ROOT = './../pytorch-face/pth/IR_50_MODEL_arcface_casia_epoch56_lfw9925.pth'
MULTI_GPU = False # flag to use multiple GPUs; if you choose to train with single GPU, you should first run "export CUDA_VISILE_DEVICES=device_id" to specify the GPU card you want to use
# DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# EMBEDDING_SIZE = 512 # feature dimension
# BATCH_SIZE = 512
# BACKBONE = IR_50(INPUT_SIZE)
# if os.path.isfile(BACKBONE_RESUME_ROOT):
# print("Loading Backbone Checkpoint '{}'".format(BACKBONE_RESUME_ROOT))
# BACKBONE.load_state_dict(torch.load(BACKBONE_RESUME_ROOT, map_location='cpu'))
# else:
# print("No Checkpoint Found at '{}'.".format(BACKBONE_RESUME_ROOT))
# sys.exit()
print("=" * 60)
print(
"Performing Evaluation on LFW, CFP_FF, CFP_FP, AgeDB, CALFW, CPLFW and VGG2_FP, and Save Checkpoints..."
)
#### LFW
print("Performing Evaluation on LFW...")
lfw, lfw_issame = get_val_pair(ARGS.data_root, 'lfw')
accuracy_lfw, best_threshold_lfw, roc_curve_lfw = perform_val(
MULTI_GPU, device, ARGS.embedding_size, ARGS.batch_size, model, lfw,
lfw_issame)
print("Evaluation: LFW Acc: {}".format(accuracy_lfw))
#### CALFW WORKS
print("Performing Evaluation on CALFW...")
calfw, calfw_issame = get_val_pair(ARGS.data_root, 'calfw')
accuracy_calfw, best_threshold_calfw, roc_curve_calfw = perform_val(
MULTI_GPU, device, ARGS.embedding_size, ARGS.batch_size, model, calfw,
calfw_issame)
print("Evaluation: CALFW Acc: {}".format(accuracy_calfw))
#### CPLFW
print("Performing Evaluation on CPLFW...")
cplfw, cplfw_issame = get_val_pair(ARGS.data_root, 'cplfw')
accuracy_cplfw, best_threshold_calfw, roc_curve_calfw = perform_val(
MULTI_GPU, device, ARGS.embedding_size, ARGS.batch_size, model, cplfw,
cplfw_issame)
print("Evaluation: CPLFW Acc: {}".format(accuracy_cplfw))
#### CFP-FF
print("Performing Evaluation on CFP-FF...")
cfp_ff, cfp_ff_issame = get_val_pair(ARGS.data_root, 'cfp_ff')
accuracy_cfp_ff, best_threshold_cfp_ff, roc_curve_cfp_ff = perform_val(
MULTI_GPU, device, ARGS.embedding_size, ARGS.batch_size, model, cfp_ff,
cfp_ff_issame)
print("Evaluation: CFP-FF Acc: {}".format(accuracy_cfp_ff))
#### CFP-FP
print("Performing Evaluation on CFP-FP...")
cfp_fp, cfp_fp_issame = get_val_pair(ARGS.data_root, 'cfp_fp')
accuracy_cfp_fp, best_threshold_cfp_fp, roc_curve_cfp_fp = perform_val(
MULTI_GPU, device, ARGS.embedding_size, ARGS.batch_size, model, cfp_fp,
cfp_fp_issame)
print("Evaluation: CFP-FP Acc: {}".format(accuracy_cfp_fp))
#### AgeDB_30
print("Performing Evaluation on AgeDB_30...")
agedb_30, agedb_30_issame = get_val_pair(ARGS.data_root, 'agedb_30')
accuracy_agedb, best_threshold_agedb, roc_curve_agedb = perform_val(
MULTI_GPU, device, ARGS.embedding_size, ARGS.batch_size, model,
agedb_30, agedb_30_issame)
print("Evaluation: AgeDB_30 Acc: {}".format(accuracy_agedb))
#### VggFace2_FP
print("Performing Evaluation on VggFace2_FP...")
vgg2_fp, vgg2_fp_issame = get_val_pair(ARGS.data_root, 'vgg2_fp')
accuracy_vgg2_fp, best_threshold_vgg2_fp, roc_curve_vgg2_fp = perform_val(
MULTI_GPU, device, ARGS.embedding_size, ARGS.batch_size, model,
vgg2_fp, vgg2_fp_issame)
print("Evaluation: VggFace2_FP Acc: {}".format(accuracy_vgg2_fp))
print("=" * 60)
print("FINAL RESULTS:")
print(
"Evaluation: LFW Acc: {}, CFP_FF Acc: {}, CFP_FP Acc: {}, AgeDB Acc: {}, CALFW Acc: {}, CPLFW Acc: {}, VGG2_FP Acc: {}"
.format(accuracy_lfw, accuracy_cfp_ff, accuracy_cfp_fp, accuracy_agedb,
accuracy_calfw, accuracy_cplfw, accuracy_vgg2_fp))
print("=" * 60)
lfw, cfp_ff, cfp_fp, agedb, calfw, cplfw, vgg2_fp, lfw_issame, cfp_ff_issame, cfp_fp_issame, agedb_issame, calfw_issame, cplfw_issame, vgg2_fp_issame = get_val_data(
DATA_ROOT)
#======= model & loss & optimizer =======#
BACKBONE_DICT = {
'ResNet_50':
ResNet_50(INPUT_SIZE),
'ResNet_101':
ResNet_101(INPUT_SIZE),
'ResNet_152':
ResNet_152(INPUT_SIZE),
'IR_50':
IR_50(INPUT_SIZE),
'IR_101':
IR_101(INPUT_SIZE),
'IR_152':
IR_152(INPUT_SIZE),
'IR_SE_50':
IR_SE_50(INPUT_SIZE),
'IR_SE_101':
IR_SE_101(INPUT_SIZE),
'IR_SE_152':
IR_SE_152(INPUT_SIZE),
'ShuffleNetV2_0.5':
shufflenet_v2_x0_5(pretrained=True, only_features=True),
'ShuffleNetV2_1.0':
shufflenet_v2_x1_0(pretrained=False, only_features=True),
'ShuffleNetV2_1.5':
shufflenet_v2_x1_5(pretrained=False, only_features=True),
'ShuffleNetV2_2.0':
train_loader = torch.utils.data.DataLoader(
dataset_train, batch_size = BATCH_SIZE, sampler = sampler, pin_memory = PIN_MEMORY,
num_workers = NUM_WORKERS, drop_last = DROP_LAST
)
NUM_CLASS = len(train_loader.dataset.classes)
print("Number of Training Classes: {}".format(NUM_CLASS))
# validate on LFW, CFP_FF, CFP_FP, AgeDB, CALFW, CPLFW and VGGFace2_FP
lfw, cfp_ff, cfp_fp, agedb, calfw, cplfw, vgg2_fp, lfw_issame, cfp_ff_issame, cfp_fp_issame, agedb_issame, calfw_issame, cplfw_issame, vgg2_fp_issame = get_val_data(DATA_ROOT)
#======= model & loss & optimizer =======#
BACKBONE_DICT = {'ResNet_50': ResNet_50(INPUT_SIZE), 'ResNet_101': ResNet_101(INPUT_SIZE), 'ResNet_152': ResNet_152(INPUT_SIZE),
'IR_50': IR_50(INPUT_SIZE), 'IR_101': IR_101(INPUT_SIZE), 'IR_152': IR_152(INPUT_SIZE),
'IR_SE_50': IR_SE_50(INPUT_SIZE), 'IR_SE_101': IR_SE_101(INPUT_SIZE), 'IR_SE_152': IR_SE_152(INPUT_SIZE)}
BACKBONE = BACKBONE_DICT[BACKBONE_NAME]
print("=" * 60)
print(BACKBONE)
print("{} Backbone Generated".format(BACKBONE_NAME))
print("=" * 60)
HEAD_DICT = {'ArcFace': ArcFace(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS),
'CosFace': CosFace(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS),
'SphereFace': SphereFace(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS),
'Am_softmax': Am_softmax(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS)}
HEAD = HEAD_DICT[HEAD_NAME]
print("=" * 60)
print(HEAD)
print("{} Head Generated".format(HEAD_NAME))