def save_state(self,
conf,
accuracy,
to_save_folder=False,
extra=None,
model_only=False):
if to_save_folder:
save_path = conf.save_path
else:
save_path = conf.model_path
lz.mkdir_p(save_path, delete=False)
torch.save(
self.model.state_dict(),
save_path / ('model_{}_accuracy:{}_step:{}_{}.pth'.format(
get_time(), accuracy, self.step, extra)))
if not model_only:
torch.save(
self.head.state_dict(),
save_path / ('head_{}_accuracy:{}_step:{}_{}.pth'.format(
get_time(), accuracy, self.step, extra)))
torch.save(
self.optimizer.state_dict(),
save_path / ('optimizer_{}_accuracy:{}_step:{}_{}.pth'.format(
get_time(), accuracy, self.step, extra)))
def crop_face(args):
assert osp.exists(args.data_dir), "The input dir not exist"
root_folder_name = args.data_dir.split('/')[-1]
dst_folder = args.data_dir.replace(root_folder_name, root_folder_name + '_OPPOFaces')
lz.mkdir_p(dst_folder, delete=False)
mtcnn = MTCNN()
ind = 0
all_img = []
for imgfn in itertools.chain(
glob.glob(args.data_dir + '/**/*.jpg', recursive=True),
glob.glob(args.data_dir + '/**/*.JPEG', recursive=True)):
ind += 1
if ind % 10 == 0:
print(f'proc {ind}, {imgfn}')
dstimgfn = imgfn.replace(root_folder_name, root_folder_name + '_OPPOFaces')
dst_folder = osp.dirname(dstimgfn)
lz.mkdir_p(dst_folder, delete=False)
img = cvb.read_img(imgfn) # bgr
img1 = Image.fromarray(img)
face = mtcnn.align_best(img1, limit=None, min_face_size=16, imgfn=imgfn)
face = np.asarray(face) # bgr
# face = cvb.bgr2rgb(face) # rgb
cvb.write_img(face, dstimgfn)
all_img.append( dstimgfn)
logging.info(f'finish crop all {ind} imgs')
lz.msgpack_dump(all_img, dst_folder + '/' + 'all_imgs.pk')
del mtcnn
torch.cuda.empty_cache()
def consumer(queue, lock):
while True:
imgfn, param, roi_box, dst_imgfn = queue.get()
pts68 = [predict_68pts(param[i], roi_box[i]) for i in range(param.shape[0])]
for img_fp, pts68_, dst in zip(imgfn, pts68, dst_imgfn):
try:
img_ori = cvb.read_img(img_fp)
pts5 = to_landmark5(pts68_[:2, :].transpose())
warped = preprocess(img_ori, landmark=pts5)
# plt_imshow(warped, inp_mode='bgr'); plt.show()
lz.mkdir_p(osp.dirname(dst), delete=False)
cvb.write_img(warped, dst)
except Exception as e:
logging.warning(f'error occur {e}, pls check!')
cvb.write_img(np.ones((112, 112, 3), np.uint8), dst)
def run(_):
cfgs = lz.load_cfg('./cfgs/single_ohnm.py')
procs = []
for args in cfgs.cfgs:
if args.loss != 'trivid':
print(f'skip {args.loss} {args.logs_dir}')
continue
if args.log_at is None:
args.log_at = np.concatenate([
range(0, 640, 31),
range(args.epochs - 8, args.epochs, 1)
])
args.logs_dir = lz.work_path + 'reid/work/' + args.logs_dir
if osp.exists(args.logs_dir) and osp.exists(args.logs_dir + '/checkpoint.64.pth'):
print(os.listdir(args.logs_dir))
continue
if not args.gpu_fix:
args.gpu = lz.get_dev(n=len(args.gpu),
ok=args.gpu_range,
mem_thresh=[0.09, 0.09], sleep=32.3)
lz.logging.info(f'use gpu {args.gpu}')
# args.batch_size = 16
# args.gpu = (3, )
# args.epochs = 1
# args.logs_dir+='.bak'
if isinstance(args.gpu, int):
args.gpu = [args.gpu]
if not args.evaluate and not args.vis:
assert args.logs_dir != args.resume
lz.mkdir_p(args.logs_dir, delete=True)
lz.pickle_dump(args, args.logs_dir + '/conf.pkl')
if cfgs.no_proc:
main(args)
else:
proc = mp.Process(target=main, args=(args,))
proc.start()
lz.logging.info('next')
time.sleep(random.randint(39, 90))
if not cfgs.parallel:
proc.join()
else:
procs.append(proc)
if cfgs.parallel:
for proc in procs:
proc.join()
def run(_):
cfgs = lz.load_cfg('./cfgs/single_ohnm.py')
procs = []
for args in cfgs.cfgs:
if args.loss != 'tcx':
print(f'skip {args.loss} {args.logs_dir}')
continue
# args.log_at = np.concatenate([
# args.log_at,
# range(args.epochs - 8, args.epochs, 1)
# ])
args.logs_dir = 'work/' + args.logs_dir
if not args.gpu_fix:
args.gpu = lz.get_dev(n=len(args.gpu),
ok=args.gpu_range,
mem=[0.12, 0.07],
sleep=32.3)
lz.logging.info(f'use gpu {args.gpu}')
# args.batch_size = 16
# args.gpu = (3, )
# args.epochs = 1
# args.logs_dir+='.bak'
if isinstance(args.gpu, int):
args.gpu = [args.gpu]
if not args.evaluate:
assert args.logs_dir != args.resume
lz.mkdir_p(args.logs_dir, delete=True)
lz.pickle_dump(args, args.logs_dir + '/conf.pkl')
# main(args)
proc = mp.Process(target=main, args=(args, ))
proc.start()
lz.logging.info('next')
time.sleep(random.randint(39, 90))
procs.append(proc)
for proc in procs:
proc.join()
def do_align_by_list(inps):
from lz import mkdir_p, cvb
ind, tid, sid, fn, x, y, w, h, mtcnn = inps
dst_dir = f'{dst}/{sid}/{tid}'
dst_fn = f'{dst}/{sid}/{tid}/{ind}.png'
if osp.exists(dst_fn): return
# logging.info(f'{ind} start')
x, y, w, h = list(map(int, [x, y, w, h]))
imgp = img_path + fn
assert osp.exists(imgp), imgp
img = cvb.read_img(imgp)
face = img[y:y + h, x:x + w, :]
face_ali = alignface(face, mtcnn)
_ = mkdir_p(dst_dir, delete=False)
_ = cvb.write_img(face_ali, dst_fn)
def do_align_one(
ind,
val,
):
tid = val['TEMPLATE_ID']
sid = val['SUBJECT_ID']
fn = val['FILENAME']
dst_dir = f'{dst}/{sid}/{tid}'
dst_fn = f'{dst}/{sid}/{tid}/{ind}.png'
# if osp.exists(dst_fn): return
x, y, w, h = val.iloc[-4:]
x, y, w, h = list(map(int, [x, y, w, h]))
imgp = img_path + fn
assert osp.exists(imgp), imgp
img = cvb.read_img(imgp)
assert img is not None, 'impg'
face = img[y:y + h, x:x + w, :]
face_ali = alignface(face, mtcnn, img)
_ = mkdir_p(dst_dir, delete=False)
_ = cvb.write_img(face_ali, dst_fn)
def crop_face(args):
for k, v in default_args.items():
setattr(args, k, v)
assert osp.exists(args.data_dir), "The input dir not exist"
root_folder_name = args.data_dir.split('/')[-1]
src_folder = args.data_dir
dst_folder = args.data_dir.replace(root_folder_name, root_folder_name + '_OPPOFaces')
lz.mkdir_p(dst_folder, delete=False)
ds = TestData(src_folder)
loader = torch.utils.data.DataLoader(ds, batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=False,
pin_memory=True,
drop_last=False
)
# 1. load pre-tained model
checkpoint_fp = 'models/phase1_wpdc_vdc.pth.tar'
arch = 'mobilenet_1'
checkpoint = torch.load(checkpoint_fp, map_location=lambda storage, loc: storage)['state_dict']
model = getattr(mobilenet_v1, arch)(num_classes=62) # 62 = 12(pose) + 40(shape) +10(expression)
model_dict = model.state_dict()
# because the model is trained by multiple gpus, prefix module should be removed
for k in checkpoint.keys():
model_dict[k.replace('module.', '')] = checkpoint[k]
model.load_state_dict(model_dict)
if args.mode == 'gpu':
cudnn.benchmark = True
model = model.cuda()
model.eval()
# 2. load dlib model for face detection and landmark used for face cropping
queue = Queue()
lock = Lock()
consumers = []
for i in range(args.num_consumers):
p = Process(target=consumer, args=(queue, lock))
p.daemon = True
consumers.append(p)
for c in consumers:
c.start()
# 3. forward
ttl_nimgs = 0
ttl_imgs = []
data_meter = lz.AverageMeter()
model_meter = lz.AverageMeter()
post_meter = lz.AverageMeter()
lz.timer.since_last_check('start crop face')
for ind, data in enumerate(loader):
data_meter.update(lz.timer.since_last_check(verbose=False))
if (data['finish'] == 1).all().item():
logging.info('finish')
break
if ind % 10 == 0:
logging.info(
f'proc batch {ind}, data time: {data_meter.avg:.2f}, model: {model_meter.avg:.2f}, post: {post_meter.avg:.2f}')
mask = data['finish'] == 0
input = data['img'][mask]
input_np = input.numpy()
roi_box = data['roi_box'][mask].numpy()
imgfn = np.asarray(data['imgfn'])[mask.numpy().astype(bool)]
dst_imgfn = [img_fp.replace(root_folder_name, root_folder_name + '_OPPOFaces') for img_fp in imgfn]
ttl_imgs.extend(dst_imgfn)
ttl_nimgs += mask.sum().item()
with torch.no_grad():
if args.mode == 'gpu':
input = input.cuda()
param = model(input)
param = param.squeeze().cpu().numpy().astype(np.float32)
model_meter.update(lz.timer.since_last_check(verbose=False))
queue.put((imgfn, param, roi_box, dst_imgfn))
# pts68 = [predict_68pts(param[i], roi_box[i]) for i in range(param.shape[0])]
# pts68_proc = [predict_68pts(param[i], [0, 0, STD_SIZE, STD_SIZE]) for i in range(param.shape[0])]
# for img_fp, pts68_, pts68_proc_, img_, dst in zip(imgfn, pts68, pts68_proc, input_np, dst_imgfn):
# ## this may need opt to async read write
# img_ori = cvb.read_img(img_fp)
# pts5 = to_landmark5(pts68_[:2, :].transpose())
# warped = preprocess(img_ori, landmark=pts5)
# # plt_imshow(warped, inp_mode='bgr'); plt.show()
# lz.mkdir_p(osp.dirname(dst), delete=False)
# cvb.write_img(warped, dst)
#
# ## this may cause black margin
# # pts5 = to_landmark5(pts68_proc_[:2, :].transpose())
# # warped = preprocess(to_img(img_), landmark=pts5)
# # # plt_imshow(warped, inp_mode='bgr'); plt.show()
# # dst = img_fp.replace(root_folder_name, root_folder_name + '_OPPOFaces')
# # cvb.write_img(warped, dst)
# if args.dump_res:
# img_ori = cvb.read_img(img_fp)
# pts_res = [pts68_]
# dst = img_fp.replace(root_folder_name, root_folder_name + '_kpts.demo')
# lz.mkdir_p(osp.dirname(dst), delete=False)
# draw_landmarks(img_ori, pts_res,
# wfp=dst,
# show_flg=args.show_flg)
post_meter.update(lz.timer.since_last_check(verbose=False))
lz.msgpack_dump(ttl_imgs, dst_folder + '/' + 'all_imgs.pk')
del model, input
torch.cuda.empty_cache()
while not queue.empty():
time.sleep(1)
def __init__(self, conf, inference=False, need_loader=True):
print(conf)
if conf.use_mobilfacenet:
# self.model = MobileFaceNet(conf.embedding_size).to(conf.device)
self.model = torch.nn.DataParallel(
MobileFaceNet(conf.embedding_size)).cuda()
print('MobileFaceNet model generated')
else:
# self.model = Backbone(conf.net_depth, conf.drop_ratio, conf.net_mode).to(conf.device)
self.model = torch.nn.DataParallel(
Backbone(conf.net_depth, conf.drop_ratio,
conf.net_mode)).cuda()
print('{}_{} model generated'.format(conf.net_mode,
conf.net_depth))
if not inference:
self.milestones = conf.milestones
if need_loader:
# self.loader, self.class_num = get_train_loader(conf)
self.dataset = Dataset2()
self.loader = DataLoader(self.dataset,
batch_size=conf.batch_size,
num_workers=conf.num_workers,
shuffle=True,
pin_memory=True)
# self.loader = Loader2(conf)
self.class_num = 85164
print(self.class_num, 'classes, load ok ')
else:
import copy
conf_t = copy.deepcopy(conf)
conf_t.data_mode = 'emore'
self.loader, self.class_num = get_train_loader(conf_t)
print(self.class_num)
self.class_num = 85164
lz.mkdir_p(conf.log_path, delete=True)
self.writer = SummaryWriter(conf.log_path)
self.step = 0
if conf.loss == 'arcface':
self.head = Arcface(embedding_size=conf.embedding_size,
classnum=self.class_num).to(conf.device)
elif conf.loss == 'softmax':
self.head = MySoftmax(embedding_size=conf.embedding_size,
classnum=self.class_num).to(conf.device)
else:
raise ValueError(f'{conf.loss}')
print('two model heads generated')
paras_only_bn, paras_wo_bn = separate_bn_paras(self.model)
if conf.use_mobilfacenet:
self.optimizer = optim.SGD(
[{
'params': paras_wo_bn[:-1],
'weight_decay': 4e-5
}, {
'params': [paras_wo_bn[-1]] + [self.head.kernel],
'weight_decay': 4e-4
}, {
'params': paras_only_bn
}],
lr=conf.lr,
momentum=conf.momentum)
else:
self.optimizer = optim.SGD(
[{
'params': paras_wo_bn + [self.head.kernel],
'weight_decay': 5e-4
}, {
'params': paras_only_bn
}],
lr=conf.lr,
momentum=conf.momentum)
print(self.optimizer)
# self.scheduler = optim.lr_scheduler.ReduceLROnPlateau(self.optimizer, patience=40, verbose=True)
print('optimizers generated')
self.board_loss_every = 100 # len(self.loader) // 100
self.evaluate_every = len(self.loader) // 10
self.save_every = len(self.loader) // 5
self.agedb_30, self.cfp_fp, self.lfw, self.agedb_30_issame, self.cfp_fp_issame, self.lfw_issame = get_val_data(
self.loader.dataset.root_path)
else:
self.threshold = conf.threshold
])
draw_ellipse = True
i, j = np.indices(pij2d.shape)
i = i.ravel()
j = j.ravel()
pij = pij2d.ravel().astype('float32')
# Remove self-indices
idx = i != j
i, j, pij = i[idx], j[idx], pij[idx]
n_topics = 2
n_dim = 2
print(n_points, n_dim, n_topics)
lz.mkdir_p(root_path + '/work/' + path, delete=True)
os.chdir(root_path + '/work/' + path)
model = VTSNE(n_points, n_topics, n_dim)
wrap = Wrapper(model, batchsize=4096, epochs=1)
for itr in range(235):
print(itr, end=' ')
wrap.fit(pij, i, j)
# Visualize the results
embed = model.logits.weight.cpu().data.numpy()
f = plt.figure()
if not draw_ellipse:
plt.scatter(embed[:, 0], embed[:, 1], c=y * 1.0 / y.max())
plt.axis('off')
plt.savefig('scatter_{:03d}.png'.format(itr), bbox_inches='tight')
plt.close(f)
parser.add_argument('--num_consumers', type=int, default=6)
parser.add_argument('--gpus', type=str, default="0") # todo allow multiple gpu
args = parser.parse_args()
assert osp.exists(args.data_dir), "The input dir not exist"
root_folder_name = args.data_dir.split('/')[-1]
src_folder = args.data_dir.replace(root_folder_name, root_folder_name + '_OPPOFaces')
if not osp.exists(src_folder):
logging.info('first crop face, an alternative way is run python crop_face_oppo.py --data_dir DATASET. ')
from crop_face_oppo_fast import crop_face
# from crop_face_oppo import crop_face
crop_face(args)
dst_folder = args.data_dir.replace(root_folder_name, root_folder_name + '_OPPOFeatures')
lz.mkdir_p(dst_folder, delete=False)
class TestData(torch.utils.data.Dataset):
def __init__(self, imgfn_iter):
self.imgfn_iter = imgfn_iter
try:
self.imgfns = lz.msgpack_load(src_folder + '/all_imgs.pk')
except:
logging.info(
"After crop_face_oppo.py runned, *_OPPOFaces/all_imgs.pk will be generetd, which logs img list. But, all_imgs.pk cannot be loaded, we are regenerating all img list now ...")
self.imgfns = list(self.imgfn_iter)
self.length = len(self.imgfns)
# self.imgfn_iter is not thread safe
# self.lock = torch.multiprocessing.Lock()
# self.length = int(10 * 10 ** 6) # assume ttl test img less than 10M