def inf_train_gen(loader):
while True:
for batch in loader:
yield batch
ctx = [mx.gpu(int(i)) for i in opt.ctx.split(",")]
batch_size = opt.batch_size
num_iterations = opt.niters
margin_s = opt.margin_s
margin_m = opt.margin_m
train_set = get_recognition_dataset(opt.dataset, transform=transform_train)
train_data = DataLoader(train_set, batch_size, shuffle=True, num_workers=opt.num_workers, last_batch='discard')
batch_generator = inf_train_gen(train_data)
targets = opt.target
val_sets = [get_recognition_dataset(name, transform=transform_test) for name in targets.split(",")]
val_datas = [DataLoader(dataset, batch_size, last_batch='keep') for dataset in val_sets]
dtype = opt.dtype
train_net = get_model(opt.model, classes=train_set.num_classes, weight_norm=True, feature_norm=True)
train_net.initialize(init=mx.init.MSRAPrelu(), ctx=ctx)
lr_period = [int(iter) for iter in opt.lr_decay_iter.split(",")]
lr_scheduler = IterLRScheduler(mode=opt.lr_mode, baselr=opt.lr, step=lr_period,
step_factor=opt.lr_decay, power=2,
niters=num_iterations, warmup_iters=opt.lr_warmup_iters)
'num_iterations={}, lr_warmup_iters={}, loss_warmup_iters={}.'.format(
num_iterations, lr_warmup_iters, loss_warmup_iters))
return num_iterations, lr_warmup_iters, loss_warmup_iters
ctx = [mx.gpu(int(i)) for i in opt.ctx.split(",")]
batch_size = opt.batch_size
num_iterations = opt.niters
lr_warmup_iters = opt.lr_warmup_iters
loss_warmup_iters = opt.loss_warmup_iters
margin_s = opt.margin_s
margin_m = opt.margin_m
train_set = get_recognition_dataset(opt.dataset, transform=transform_train)
train_data = DataLoader(train_set,
batch_size,
shuffle=True,
num_workers=opt.num_workers,
last_batch='discard')
batch_generator = inf_train_gen(train_data)
if num_iterations == 0:
# Auto setting. You should have a large batch size to enable this(512 or larger is recommend).
# Epochs 25, loss warm up 35%, lr warm up 5% mixup iters 90%.
num_iterations, lr_warmup_iters, loss_warmup_iters = auto_train_setting(
train_data._dataset, epochs=opt.epochs)
targets = opt.target
val_sets = [
logger.info('use %d gpus' % num_gpu)
logger.info(ctx)
else:
ctx = [mx.cpu()]
logger.info('use cpu')
logger.info(ctx)
batch_size = (opt.batch_size * num_gpu) if num_gpu > 0 else opt.batch_size
lr_decay_epoch = [
int(i) + opt.warmup_epochs for i in opt.lr_decay_epoch.split(',')
]
logger.info('loading training and validation data')
image_src_root = opt.data_dir
train_set = get_recognition_dataset("faces_emore",
root=image_src_root,
transform=transform_train)
train_data = DataLoader(train_set,
batch_size,
shuffle=True,
num_workers=opt.num_workers)
# targets = ['lfw', 'cfp_fp', 'agedb_30']
targets = [i for i in opt.validation_targets.split(',')]
val_sets = [
get_recognition_dataset(name,
root=os.path.join(image_src_root, 'faces_emore'),
transform=transform_test) for name in targets
]
val_datas = [
DataLoader(dataset, batch_size, num_workers=opt.num_workers)
ctx = [mx.gpu(i) for i in range(num_gpu)]
batch_size = 128 * num_gpu
num_worker = 24
save_period = 500
iters = 200e3
lr_steps = [30e3, 60e3, 90e3, np.inf]
scale = 50
margin = 0.5
embedding_size = 256
lr = 0.001
momentum = 0.9
wd = 4e-5
train_set = get_recognition_dataset("emore", transform=transform_train)
train_data = DataLoader(train_set,
batch_size,
shuffle=True,
num_workers=num_worker)
targets = ['lfw']
val_sets = [
get_recognition_dataset(name, transform=transform_test) for name in targets
]
val_datas = [
DataLoader(dataset, batch_size, num_workers=num_worker)
for dataset in val_sets
]
net = get_mobile_facenet(train_set.num_classes,
epochs = 15
save_period = 3000
warmup_epochs = 1
scale = 60
margin = 0.5
lr = 0.1
momentum = 0.9
wd = 4e-5
use_float16 = False
trans = Transform(use_float16)
train_set = get_recognition_dataset("faces_emore",
transform=trans.transform_train)
train_data = DataLoader(train_set,
batch_size,
shuffle=True,
num_workers=num_worker,
last_batch="discard")
targets = ['lfw']
val_sets = [
get_recognition_dataset(name, transform=trans.transform_test)
for name in targets
]
val_datas = [
DataLoader(dataset, batch_size, num_workers=num_worker)
for dataset in val_sets
]
if isf:
data = nd.transpose(data, (2, 0, 1)).astype('float32')
flip_data = nd.transpose(flip_data, (2, 0, 1)).astype('float32')
return data, flip_data
return transform_test(data), transform_test(flip_data)
export_path = os.path.dirname(
opt.model_params) if opt.export_path == '' else opt.export_path
ctx = [mx.gpu(int(i)) for i in opt.ctx.split(",")]
batch_size = opt.batch_size
targets = opt.target
val_sets = [
get_recognition_dataset(name, transform=transform_test_flip)
for name in targets.split(",")
]
val_datas = [
DataLoader(dataset, batch_size, last_batch='keep') for dataset in val_sets
]
test_net = get_model(opt.model, need_cls_layer=False)
test_net.cast(opt.dtype)
test_net.load_parameters(opt.model_params, ctx=ctx, ignore_extra=True)
def validate(nfolds=10):
metric = FaceVerification(nfolds)
metric_flip = FaceVerification(nfolds)
for loader, name in zip(val_datas, targets.split(",")):