def swap_faces(Xs_raw, Xt_raw):
detector = MTCNN()
device = torch.device('cpu')
G = AEI_Net(c_id=512)
G.eval()
G.load_state_dict(torch.load('./saved_models/G_latest.pth', map_location=device))
G = G.cpu()
f_siz = 256
arcface = Backbone(50, 0.6, 'ir_se').to(device)
arcface.eval()
arcface.load_state_dict(torch.load('./saved_models/model_ir_se50.pth', map_location=device), strict=False)
test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
Xs_img = Image.fromarray(Xs_raw)
Xs = detector.align(Xs_img, crop_size=(f_siz, f_siz))
Xs = test_transform(Xs)
Xs = Xs.unsqueeze(0).cpu()
with torch.no_grad():
embeds, Xs_feats = arcface(F.interpolate(Xs, (112, 112), mode='bilinear', align_corners=True))
embeds = embeds.mean(dim=0, keepdim=True)
mask = np.zeros([f_siz, f_siz], dtype=np.float)
for i in range(f_siz):
for j in range(f_siz):
dist = np.sqrt((i-(f_siz/2))**2 + (j-(f_siz/2))**2)/(f_siz/2)
dist = np.minimum(dist, 1)
mask[i, j] = 1-dist
mask = cv2.dilate(mask, None, iterations=20)
Xt_img = Image.fromarray(Xt_raw)
Xt, trans_inv = detector.align(Xt_img, crop_size=(f_siz, f_siz), return_trans_inv=True)
Xt = test_transform(Xt)
Xt = Xt.unsqueeze(0).cpu()
with torch.no_grad():
st = time.time()
Yt, _ = G(Xt, embeds)
Yt = Yt.squeeze().detach().cpu().numpy()
st = time.time() - st
print(f'inference time: {st} sec')
Yt = Yt.transpose([1, 2, 0])*0.5 + 0.5
Yt_trans_inv = cv2.warpAffine(Yt, trans_inv, (np.size(Xt_raw, 1), np.size(Xt_raw, 0)), borderMode=cv2.BORDER_TRANSPARENT)
mask_ = cv2.warpAffine(mask, trans_inv, (np.size(Xt_raw, 1), np.size(Xt_raw, 0)), borderMode=cv2.BORDER_TRANSPARENT)
mask_ = np.expand_dims(mask_, 2)
Yt_trans_inv = mask_*Yt_trans_inv + (1-mask_)*(Xt_raw.astype(np.float)/255.)
merge = Yt_trans_inv
return merge
model_save_path = './saved_models/'
optim_level = 'O0'
# fine_tune_with_identity = False
device = torch.device('cuda')
# torch.set_num_threads(12)
G = AEI_Net(c_id=512).to(device)
D = MultiscaleDiscriminator(input_nc=3, n_layers=6, norm_layer=torch.nn.InstanceNorm2d).to(device)
G.train()
D.train()
arcface = Backbone(50, 0.6, 'ir_se').to(device)
arcface.eval()
arcface.load_state_dict(torch.load('./id_model/model_ir_se50.pth', map_location=device), strict=False)
opt_G = optim.Adam(G.parameters(), lr=lr_G, betas=(0, 0.999))
opt_D = optim.Adam(D.parameters(), lr=lr_D, betas=(0, 0.999))
G, opt_G = amp.initialize(G, opt_G, opt_level=optim_level)
D, opt_D = amp.initialize(D, opt_D, opt_level=optim_level)
try:
p_G = './saved_mask_models/G_latest.pth'
p_D = './saved_mask_models/D_latest.pth'
G.load_state_dict(torch.load(p_G, map_location=torch.device('cpu')), strict=False)
D.load_state_dict(torch.load(p_D, map_location=torch.device('cpu')), strict=False)
print('p_G : ',p_G)
print('p_D : ',p_D)
def serve():
data = {"success": False}
detector = MTCNN()
device = torch.device('cuda')
G = AEI_Net(c_id=512)
G.eval()
G.load_state_dict(
torch.load('./saved_models/G_latest.pth',
map_location=torch.device('cpu')))
G = G.cuda()
arcface = Backbone(50, 0.6, 'ir_se').to(device)
arcface.eval()
arcface.load_state_dict(torch.load('./face_modules/model_ir_se50.pth',
map_location=device),
strict=False)
test_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
if flask.request.method == 'POST':
st = time.time()
metadata = flask.request.form
source_image = metadata['source_image']
target_image = metadata['target_image']
source_image = base64.b64decode(source_image.encode('utf-8'))
Xs_raw = cv2.imdecode(np.frombuffer(source_image, np.uint8),
cv2.IMREAD_COLOR)
target_image = base64.b64decode(target_image.encode('utf-8'))
Xt_raw = cv2.imdecode(np.frombuffer(target_image, np.uint8),
cv2.IMREAD_COLOR)
Xs = detector.align(Image.fromarray(Xs_raw[:, :, ::-1]),
crop_size=(256, 256))
Xs_raw = np.array(Xs)[:, :, ::-1]
Xs = test_transform(Xs)
Xs = Xs.unsqueeze(0).cuda()
with torch.no_grad():
embeds = arcface(
F.interpolate(Xs[:, :, 19:237, 19:237], (112, 112),
mode='bilinear',
align_corners=True))
Xt, trans_inv = detector.align(Image.fromarray(Xt_raw[:, :, ::-1]),
crop_size=(256, 256),
return_trans_inv=True)
Xt_raw = Xt_raw.astype(np.float) / 255.0
Xt = test_transform(Xt)
Xt = Xt.unsqueeze(0).cuda()
mask = np.zeros([256, 256], dtype=np.float)
for i in range(256):
for j in range(256):
dist = np.sqrt((i - 128)**2 + (j - 128)**2) / 128
dist = np.minimum(dist, 1)
mask[i, j] = 1 - dist
mask = cv2.dilate(mask, None, iterations=20)
with torch.no_grad():
Yt, _ = G(Xt, embeds)
Yt = Yt.squeeze().detach().cpu().numpy().transpose([1, 2, 0
]) * 0.5 + 0.5
Yt = Yt[:, :, ::-1]
Yt_trans_inv = cv2.warpAffine(
Yt,
trans_inv, (np.size(Xt_raw, 1), np.size(Xt_raw, 0)),
borderValue=(0, 0, 0))
mask_ = cv2.warpAffine(mask,
trans_inv,
(np.size(Xt_raw, 1), np.size(Xt_raw, 0)),
borderValue=(0, 0, 0))
mask_ = np.expand_dims(mask_, 2)
Yt_trans_inv = mask_ * Yt_trans_inv + (1 - mask_) * Xt_raw
img_data = Yt_trans_inv * 255
retval, buffer = cv2.imencode('.jpg', img_data)
pic_str = base64.b64encode(buffer)
pic_str = pic_str.decode()
data['success'] = True
data['image'] = pic_str
st = time.time() - st
print(f'process time: {st} sec')
return flask.jsonify(data)
os.makedirs(inner_dir)
gpu_str = parse_gpu_id(opts.gpu_ids)
os.environ['CUDA_VISIBLE_DEVICES'] = gpu_str
G = AETNet().cuda()
D = MultiscaleDiscriminator(input_nc=3,
n_layers=6,
norm_layer=torch.nn.InstanceNorm2d).cuda()
G.train()
D.train()
arcface = Backbone(50, 0.6, 'ir_se').cuda()
arcface.eval()
arcface.load_state_dict(torch.load('./saved_models/model_ir_se50.pth'),
strict=False)
opt_G = optim.Adam(G.parameters(), lr=opts.lr_G, betas=(0, 0.999))
opt_D = optim.Adam(D.parameters(), lr=opts.lr_D, betas=(0, 0.999))
resume_epoch = 0
resume_iter = 0
l_adv = 1
l_att = 10
l_id = 5
l_rec = 10
inner_count = 0 # save mid results idx. -> idx.jpg
if opts.resume:
G_checkpoint = torch.load(os.path.join(checkponits_dir, "G_latest.pth"))
D_checkpoint = torch.load(os.path.join(checkponits_dir, "D_latest.pth"))
param_checkpoint = torch.load(os.path.join(checkponits_dir, "params.pth"))