def main(args):
print(args)
MULTI_GPU = False
DEVICE = torch.device("cuda:0")
DATA_ROOT = '/raid/Data/ms1m-retinaface-t1/'
with open(os.path.join(DATA_ROOT, 'property'), 'r') as f:
NUM_CLASS, h, w = [int(i) for i in f.read().split(',')]
if args.network == 'VIT':
model = ViT_face(image_size=112,
patch_size=8,
loss_type='CosFace',
GPU_ID=DEVICE,
num_class=NUM_CLASS,
dim=512,
depth=20,
heads=8,
mlp_dim=2048,
dropout=0.1,
emb_dropout=0.1)
elif args.network == 'VITs':
model = ViTs_face(loss_type='CosFace',
GPU_ID=DEVICE,
num_class=NUM_CLASS,
image_size=112,
patch_size=8,
ac_patch_size=12,
pad=4,
dim=512,
depth=20,
heads=8,
mlp_dim=2048,
dropout=0.1,
emb_dropout=0.1)
model_root = args.model
model.load_state_dict(torch.load(model_root))
#debug
w = torch.load(model_root)
for x in w.keys():
print(x, w[x].shape)
#embed()
TARGET = [i for i in args.target.split(',')]
vers = get_val_data('./eval/', TARGET)
acc = []
for ver in vers:
name, data_set, issame = ver
accuracy, std, xnorm, best_threshold, roc_curve = perform_val(
MULTI_GPU, DEVICE, 512, args.batch_size, model, data_set, issame)
print('[%s]XNorm: %1.5f' % (name, xnorm))
print('[%s]Accuracy-Flip: %1.5f+-%1.5f' % (name, accuracy, std))
print('[%s]Best-Threshold: %1.5f' % (name, best_threshold))
acc.append(accuracy)
print('Average-Accuracy: %1.5f' % (np.mean(acc)))
dataset_train = datasets.ImageFolder(os.path.join(DATA_ROOT, 'ms1mv2-asia'), train_transform)
# create a weighted random sampler to process imbalanced data
weights = make_weights_for_balanced_classes(dataset_train.imgs, len(dataset_train.classes))
weights = torch.DoubleTensor(weights)
sampler = torch.utils.data.sampler.WeightedRandomSampler(weights, len(weights))
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))
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)
with open(os.path.join(DATA_ROOT, 'property'), 'r') as f:
NUM_CLASS, h, w = [int(i) for i in f.read().split(',')]
assert h == INPUT_SIZE[0] and w == INPUT_SIZE[1]
dataset = FaceDataset(os.path.join(DATA_ROOT, 'train.rec'),
rand_mirror=True)
trainloader = torch.utils.data.DataLoader(dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=len(GPU_ID),
drop_last=True)
print("Number of Training Classes: {}".format(NUM_CLASS))
vers = get_val_data(EVAL_PATH, TARGET)
highest_acc = [0.0 for t in TARGET]
#embed()
#======= model & loss & optimizer =======#
BACKBONE_DICT = {
'VIT':
ViT_face(loss_type=HEAD_NAME,
GPU_ID=GPU_ID,
num_class=NUM_CLASS,
image_size=112,
patch_size=8,
dim=512,
depth=20,
heads=8,
mlp_dim=2048,
BACKBONE_NAME = cfg[
'BACKBONE_NAME'] # support: ['ResNet_50', 'ResNet_101', 'ResNet_152', 'IR_50', 'IR_101', 'IR_152', 'IR_SE_50', 'IR_SE_101', 'IR_SE_152']
HEAD_NAME = cfg[
'HEAD_NAME'] # support: ['Softmax', 'ArcFace', 'CosFace', 'SphereFace', 'Am_softmax']
INPUT_SIZE = cfg['INPUT_SIZE']
BATCH_SIZE = cfg['BATCH_SIZE']
RGB_MEAN = cfg['RGB_MEAN'] # for normalize inputs
RGB_STD = cfg['RGB_STD']
EMBEDDING_SIZE = cfg['EMBEDDING_SIZE'] # feature dimension
GPU_ID = cfg['TEST_GPU_ID'] # specify your GPU ids
print("Overall Configurations:")
print(cfg)
#val_data_dir = os.path.join(VAL_DATA_ROOT, 'val_data')
val_data_dir = VAL_DATA_ROOT
lfw, cfp_fp, agedb_30, calfw, cplfw, vgg2_fp, lfw_issame, cfp_fp_issame, agedb_30_issame, calfw_issame, cplfw_issame, vgg2_fp_issame = get_val_data(
VAL_DATA_ROOT)
#======= model =======#
BACKBONE_DICT = {
'ResNet_50': ResNet_50,
'ResNet_101': ResNet_101,
'ResNet_152': ResNet_152,
'IR_50': IR_50,
'IR_100': IR_100,
'IR_101': IR_101,
'IR_152': IR_152,
'IR_SE_50': IR_SE_50,
'IR_SE_101': IR_SE_101,
'IR_SE_152': IR_SE_152,
'MobileFaceNet': MobileFaceNet
}
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))
# 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(VAL_ROOT)
calfw, cplfw, calfw_issame, cplfw_issame = get_val_data(VAL_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)
len(dataset_train.classes))
weights = torch.DoubleTensor(weights)
sampler = torch.utils.data.sampler.WeightedRandomSampler(
weights, len(weights))
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))
agedb_30, cfp_fp, lfw, agedb_30_issame, cfp_fp_issame, lfw_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)
weights = torch.DoubleTensor(weights)
sampler = torch.utils.data.sampler.WeightedRandomSampler(
weights, len(weights))
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))
#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)
lfw, lfw_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)
dataset_train = datasets.ImageFolder(os.path.join(DATA_ROOT, 'imgs'), train_transform)
# create a weighted random sampler to process imbalanced data
weights = make_weights_for_balanced_classes(dataset_train.imgs, len(dataset_train.classes))
weights = torch.DoubleTensor(weights)
sampler = torch.utils.data.sampler.WeightedRandomSampler(weights, len(weights))
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))
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))
DATA_ROOT = cfg['DATA_ROOT'] # the parent root where your train/val/test data are stored
BACKBONE_RESUME_ROOT = cfg['BACKBONE_RESUME_ROOT'] # the root to resume training from a saved checkpoint
HEAD_RESUME_ROOT = cfg['HEAD_RESUME_ROOT'] # the root to resume training from a saved checkpoint
BACKBONE_NAME = cfg['BACKBONE_NAME'] # support: ['ResNet_50', 'ResNet_101', 'ResNet_152', 'IR_50', 'IR_101', 'IR_152', 'IR_SE_50', 'IR_SE_101', 'IR_SE_152']
HEAD_NAME = cfg['HEAD_NAME'] # support: ['Softmax', 'ArcFace', 'CosFace', 'SphereFace', 'Am_softmax']
INPUT_SIZE = cfg['INPUT_SIZE']
BATCH_SIZE = cfg['BATCH_SIZE']
RGB_MEAN = cfg['RGB_MEAN'] # for normalize inputs
RGB_STD = cfg['RGB_STD']
EMBEDDING_SIZE = cfg['EMBEDDING_SIZE'] # feature dimension
GPU_ID = cfg['TEST_GPU_ID'] # specify your GPU ids
print("Overall Configurations:")
print(cfg)
val_data_dir = os.path.join(DATA_ROOT, 'val_data')
lfw, cfp_fp, agedb, cplfw, calfw, lfw_issame, cfp_fp_issame, agedb_issame, cplfw_issame, calfw_issame = get_val_data(val_data_dir)
#======= model =======#
BACKBONE_DICT = {'ResNet_50': ResNet_50,
'ResNet_101': ResNet_101,
'ResNet_152': ResNet_152,
'IR_50': IR_50,
'IR_101': IR_101,
'IR_152': IR_152,
'IR_SE_50': IR_SE_50,
'IR_SE_101': IR_SE_101,
'IR_SE_152': IR_SE_152,
'MobileFaceNet': MobileFaceNet}
BACKBONE = BACKBONE_DICT[BACKBONE_NAME](INPUT_SIZE)
print("=" * 60)
