def quant_val_reader_batch():
nl, nr, flods, flags = parse_filelist(args.test_data_dir)
test_dataset = LFW(nl, nr)
test_reader = fluid.io.batch(
test_dataset.reader, batch_size=1, drop_last=False)
shuffle_reader = fluid.io.shuffle(test_reader, 3)
def _reader():
while True:
for idx, data in enumerate(shuffle_reader()):
yield np.expand_dims(data[0][0], axis=0)
return _reader
def main():
global args
parser = argparse.ArgumentParser(description='PaddlePaddle SlimFaceNet')
parser.add_argument(
'--action', default='train', type=str, help='train/test/quant')
parser.add_argument(
'--model',
default='SlimFaceNet_B_x0_75',
type=str,
help='SlimFaceNet_B_x0_75/SlimFaceNet_C_x0_75/SlimFaceNet_A_x0_60')
parser.add_argument(
'--use_gpu', default=1, type=int, help='Use GPU or not, 0 is not used')
parser.add_argument(
'--lr_strategy',
default='piecewise_decay',
type=str,
help='lr_strategy')
parser.add_argument('--lr', default=0.1, type=float, help='learning rate')
parser.add_argument(
'--lr_list',
default='0.1,0.01,0.001,0.0001',
type=str,
help='learning rate list (piecewise_decay)')
parser.add_argument(
'--lr_steps',
default='36,52,58',
type=str,
help='learning rate decay at which epochs')
parser.add_argument(
'--l2_decay', default=4e-5, type=float, help='base l2_decay')
parser.add_argument(
'--train_data_dir', default='./CASIA', type=str, help='train_data_dir')
parser.add_argument(
'--test_data_dir', default='./lfw', type=str, help='lfw_data_dir')
parser.add_argument(
'--train_batchsize', default=512, type=int, help='train_batchsize')
parser.add_argument(
'--test_batchsize', default=500, type=int, help='test_batchsize')
parser.add_argument(
'--img_shape', default='3,112,96', type=str, help='img_shape')
parser.add_argument(
'--start_epoch', default=0, type=int, help='start_epoch')
parser.add_argument(
'--total_epoch', default=80, type=int, help='total_epoch')
parser.add_argument(
'--save_frequency', default=1, type=int, help='save_frequency')
parser.add_argument(
'--save_ckpt', default='output', type=str, help='save_ckpt')
parser.add_argument(
'--feature_save_dir',
default='result.mat',
type=str,
help='The path of the extract features save, must be .mat file')
args = parser.parse_args()
if args.use_gpu:
num_trainers = fluid.core.get_cuda_device_count()
else:
num_trainers = int(os.environ.get('CPU_NUM', 1))
print(args)
print('num_trainers: {}'.format(num_trainers))
if args.save_ckpt == None:
args.save_ckpt = 'output'
if not os.path.isdir(args.save_ckpt):
os.makedirs(args.save_ckpt)
with open(os.path.join(args.save_ckpt, 'log.txt'), 'w+') as f:
f.writelines(str(args) + '\n')
f.writelines('num_trainers: {}'.format(num_trainers) + '\n')
if args.action == 'train':
train_program = fluid.Program()
test_program = fluid.Program()
startup_program = fluid.Program()
if args.action == 'train':
train_out = build_program(train_program, startup_program, args, True)
test_out = build_program(test_program, startup_program, args, False)
test_program = test_program.clone(for_test=True)
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(startup_program)
if args.action == 'train':
train(exe, train_program, train_out, test_program, test_out, args)
elif args.action == 'quant':
quant_post_static(
executor=exe,
model_dir='./out_inference/',
quantize_model_path='./quant_model/',
sample_generator=quant_val_reader_batch(),
model_filename=None, #'model',
params_filename=None, #'params',
save_model_filename=None, #'model',
save_params_filename=None, #'params',
batch_size=np.random.randint(80, 160),
batch_nums=np.random.randint(4, 10))
elif args.action == 'test':
[inference_program, feed_target_names,
fetch_targets] = fluid.io.load_inference_model(
dirname='./quant_model/',
model_filename=None,
params_filename=None,
executor=exe)
nl, nr, flods, flags = parse_filelist(args.test_data_dir)
test_dataset = LFW(nl, nr)
test_reader = fluid.io.batch(
test_dataset.reader,
batch_size=args.test_batchsize,
drop_last=False)
image_test = fluid.data(
name='image_test', shape=[-1, 3, 112, 96], dtype='float32')
image_test1 = fluid.data(
name='image_test1', shape=[-1, 3, 112, 96], dtype='float32')
image_test2 = fluid.data(
name='image_test2', shape=[-1, 3, 112, 96], dtype='float32')
image_test3 = fluid.data(
name='image_test3', shape=[-1, 3, 112, 96], dtype='float32')
image_test4 = fluid.data(
name='image_test4', shape=[-1, 3, 112, 96], dtype='float32')
reader = fluid.io.DataLoader.from_generator(
feed_list=[image_test1, image_test2, image_test3, image_test4],
capacity=64,
iterable=True,
return_list=False)
reader.set_sample_list_generator(
test_reader,
places=fluid.cuda_places() if args.use_gpu else fluid.CPUPlace())
test_out = (fetch_targets, reader, flods, flags)
print('fetch_targets[0]: ', fetch_targets[0])
print('feed_target_names: ', feed_target_names)
test(exe, inference_program, test_out, args)
else:
print('WRONG ACTION')
def build_program(program, startup, args, is_train=True):
if args.use_gpu:
num_trainers = fluid.core.get_cuda_device_count()
else:
num_trainers = int(os.environ.get('CPU_NUM', 1))
places = fluid.cuda_places() if args.use_gpu else fluid.CPUPlace()
train_dataset = CASIA_Face(root=args.train_data_dir)
trainset_scale = len(train_dataset)
with fluid.program_guard(main_program=program, startup_program=startup):
with fluid.unique_name.guard():
# Model construction
model = models.__dict__[args.model](
class_dim=train_dataset.class_nums)
if is_train:
image = fluid.data(
name='image', shape=[-1, 3, 112, 96], dtype='float32')
label = fluid.data(name='label', shape=[-1, 1], dtype='int64')
train_reader = fluid.io.batch(
train_dataset.reader,
batch_size=args.train_batchsize // num_trainers,
drop_last=False)
reader = fluid.io.DataLoader.from_generator(
feed_list=[image, label],
capacity=64,
iterable=True,
return_list=False)
reader.set_sample_list_generator(train_reader, places=places)
model.extract_feature = False
loss, acc = model.net(image, label)
optimizer = creat_optimizer(args, trainset_scale)
optimizer.minimize(loss)
global_lr = optimizer._global_learning_rate()
out = (loss, acc, global_lr, reader)
else:
nl, nr, flods, flags = parse_filelist(args.test_data_dir)
test_dataset = LFW(nl, nr)
test_reader = fluid.io.batch(
test_dataset.reader,
batch_size=args.test_batchsize,
drop_last=False)
image_test = fluid.data(
name='image_test', shape=[-1, 3, 112, 96], dtype='float32')
image_test1 = fluid.data(
name='image_test1',
shape=[-1, 3, 112, 96],
dtype='float32')
image_test2 = fluid.data(
name='image_test2',
shape=[-1, 3, 112, 96],
dtype='float32')
image_test3 = fluid.data(
name='image_test3',
shape=[-1, 3, 112, 96],
dtype='float32')
image_test4 = fluid.data(
name='image_test4',
shape=[-1, 3, 112, 96],
dtype='float32')
reader = fluid.io.DataLoader.from_generator(
feed_list=[
image_test1, image_test2, image_test3, image_test4
],
capacity=64,
iterable=True,
return_list=False)
reader.set_sample_list_generator(
test_reader,
places=fluid.cuda_places()
if args.use_gpu else fluid.CPUPlace())
model.extract_feature = True
feature = model.net(image_test)
out = (feature, reader, flods, flags)
return out
parser.add_argument(
'--test_data_dir', default='./lfw', type=str, help='lfw_data_dir')
parser.add_argument(
'--resume', default='output/0', type=str, help='resume')
parser.add_argument(
'--feature_save_dir',
default='result.mat',
type=str,
help='The path of the extract features save, must be .mat file')
args = parser.parse_args()
place = fluid.CPUPlace() if args.use_gpu == 0 else fluid.CUDAPlace(0)
with fluid.dygraph.guard(place):
train_dataset = CASIA_Face(root=args.train_data_dir)
nl, nr, flods, flags = parse_filelist(args.test_data_dir)
test_dataset = LFW(nl, nr)
test_reader = paddle.batch(
test_dataset.reader,
batch_size=args.test_batchsize,
drop_last=False)
net = models.__dict__[args.model](class_dim=train_dataset.class_nums)
if args.resume:
assert os.path.exists(args.resume + ".pdparams"
), "Given dir {}.pdparams not exist.".format(
args.resume)
para_dict, opti_dict = fluid.dygraph.load_dygraph(args.resume)
net.set_dict(para_dict)
test(test_reader, flods, flags, net, args)