def load_verts_faces_uv(self, json_path):
with open(json_path, 'r') as f:
uv_para = json.load(f)
verts_uvs = np.array(uv_para['textureMapping']['pointData'])
faces_uvs = np.array(uv_para['textureMapping']['triangles'])
verts_uvs = to_tensor(verts_uvs, self.device).unsqueeze(0).float()
faces_uvs = to_tensor(faces_uvs, self.device).unsqueeze(0).long()
return verts_uvs, faces_uvs
def __init__(self, json_path, *args, **kwargs):
# json_path = ".../face12.json"
super().__init__(*args, **kwargs)
with open(json_path, 'r') as f:
uv_para = json.load(f)
verts_uvs = np.array(uv_para['textureMapping']['pointData'])
faces_uvs = np.array(uv_para['textureMapping']['triangles'])
verts_uvs = to_tensor(verts_uvs).unsqueeze(0).float()
faces_uvs = to_tensor(faces_uvs).unsqueeze(0).long()
self.texture = ImagelessTexturesUV(verts_uvs=verts_uvs,
faces_uvs=faces_uvs).to(self.device)
def generate_vertices(self, sp, ep):
# sp, ep need to have a batch dimension
batch_size = ep.shape[0]
if self.use_bfm_gpu:
vertices = [self.bfm.generate_vertices(sp[i], ep[i]) for i in range(batch_size)]
else:
vertices = [self.bfm.generate_vertices(sp[i].detach().cpu().numpy(), ep[i].detach().cpu().numpy()) for i in
range(batch_size)]
vertices = [to_tensor(v, self.device) for v in vertices]
vertices = torch.stack(vertices)
return vertices
def __init__(self, opt, use_bfm_gpu=False):
super().__init__(opt)
self.to(opt.device)
# This factory uses the Basel Face Model
self.use_bfm_gpu = use_bfm_gpu # Loads the BFM into GPU for faster mesh generation
if self.use_bfm_gpu:
self.bfm = BFM2017Tensor(self.opt.path_bfm, device=opt.device)
else:
self.bfm = BFM2017(self.opt.path_bfm, self.opt.path_uv)
self.faces_tensor = to_tensor(self.bfm.faces, self.device).unsqueeze(0)
verts_uv, faces_uv = self.load_verts_faces_uv(opt.path_uv)
self.uv_renderer = UVRenderer(verts_uv, faces_uv, image_size=self.opt.image_h, cameras=Camera.get_camera())
def run(self, pipeline, batch, latent_from, latent_to, n):
result = list()
# linear interpolation
all_latents = interpolation(n, latent_from=latent_from, latent_to=latent_to)
all_latents = to_tensor(all_latents, batch[DIK.STYLE_LATENT].device)
for latent in all_latents:
batch2 = batch.copy()
batch2[DIK.STYLE_LATENT] = latent.reshape(batch[DIK.STYLE_LATENT].shape)
batch2 = pipeline.forward_latent2image(batch2, 0)
img_out = batch2[DIK.IMAGE_OUT][0]
img_out = self.tensor2image(img_out)
if self.mask_key is not None:
img_out = img_out * batch[self.mask_key][0].expand_as(img_out).cpu()
result.append(img_out)
return result