def prepare_image():
dataset_image_path="/home/qianqianjun/CODE/DataSets/DaeDatasets"
save_dir="/home/qianqianjun/CODE/实验结果/重建结果/改进编码器"
os.makedirs(save_dir,exist_ok=True)
length=5000
dataset=Datas(dataset_image_path,length)
loader=torch.utils.data.DataLoader(dataset,batch_size=25,shuffle=False,num_workers=4)
encoders,decoders=Encoders(parm),Decoders(parm)
device=torch.device("cpu")
nn.Module.load_state_dict(encoders,torch.load("../CelebA_out/2020-03-28_23:13:37/checkpoints/encoders.pth",map_location=device))
nn.Module.load_state_dict(decoders,torch.load("../CelebA_out/2020-03-28_23:13:37/checkpoints/decoders.pth",map_location=device))
if parm.useCuda:
encoders=encoders.cuda()
decoders=decoders.cuda()
for name,data in loader:
batch_data=batchDataTransform(data,3)
base=getBaseGrid(imgSize=parm.imgSize,Inbatch=True,batchSize=batch_data.shape[0])
if parm.useCuda:
batch_data = batch_data.cuda()
base = base.cuda()
z,zI,zW,intersI,intersW=encoders(batch_data)
texture,warp,out,_=decoders(zI,zW,intersI,intersW,base)
images=out.detach().cpu().numpy()
images=np.transpose(images,[0,2,3,1])
images=np.array(images*255,dtype=np.uint8)
for index,img in enumerate(images):
temp=cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
cv2.imwrite(os.path.join(save_dir,name[index]),temp)
def interpolation(attr_path, decoders,batch_data,
test_texture_code, test_warp_code,
test_texture_inters,test_warp_inters,**lambda_params):
# 属性向量加载
with open(attr_path, "rb") as f:
attribute = pk.load(f, encoding="utf-8")
texture_code = attribute["texture_code"]
warp_code = attribute["warp_code"]
texture_inters = attribute["texture_inter_outs_attribute"]
warp_inters = attribute["warp_inter_outs_attribute"]
# 插值系数
lambda_texture_inter_out = lambda_params["lambda_texture_inter_out"]
lambda_warp_inter_out = lambda_params["lambda_warp_inter_out"]
lambda_deepest_texture = lambda_params["lambda_deepest_texture"]
lambda_deepest_warp = lambda_params["lambda_deepest_warp"]
base = getBaseGrid(parm.imgSize, True, batch_data.size()[0])
if parm.useCuda:
decoders = torch.nn.Module.cuda(decoders)
texture_code = texture_code.cuda()
warp_code = warp_code.cuda()
for i in range(len(texture_inters)):
texture_inters[i] = texture_inters[i].cuda()
for i in range(len(warp_inters)):
warp_inters[i] = warp_inters[i].cuda()
base = base.cuda()
# 纹理信息插值
texture_code_interpolation_result = test_texture_code + lambda_deepest_texture * texture_code.unsqueeze(0).repeat(
test_texture_code.size()[0], 1)
texture_inter_interpolation_result = []
for i in range(len(test_texture_inters)):
texture_inter_interpolation_result.append(
test_texture_inters[i] + lambda_texture_inter_out[i] *
texture_inters[i].repeat(test_texture_inters[i].size()[0], 1, 1, 1)
)
### 变形信息插值
warp_code_interpolation_result = test_warp_code + lambda_deepest_warp * warp_code.unsqueeze(0).repeat(
test_warp_code.size()[0], 1)
warp_inter_interpolation_result = []
for i in range(len(test_warp_inters)):
warp_inter_interpolation_result.append(
test_warp_inters[i] + lambda_warp_inter_out[i] *
warp_inters[i].repeat(test_warp_inters[i].size()[0], 1, 1, 1)
)
# 结果图像重建
texture_result, warp_result, result, _ = decoders(texture_code_interpolation_result,
warp_code_interpolation_result,
texture_inter_interpolation_result,
warp_inter_interpolation_result, base)
return result
decoders,
torch.load(os.path.join(model_path, "decoders.pth"), map_location=device))
img_dir = "/home/qianqianjun/桌面/杨洋男"
names = os.listdir(img_dir)
imgs_path = [os.path.join(img_dir, name) for name in names]
batch_data = batchDataTransform(torch.tensor([
cv2.cvtColor(
cv2.resize(cv2.imread(img_path),
(parm.imgSize, parm.imgSize)), cv2.COLOR_BGR2RGB)
for img_path in imgs_path[:9]
],
dtype=torch.float32),
channel=3)
base = getBaseGrid(imgSize=parm.imgSize,
Inbatch=True,
batchSize=batch_data.shape[0])
if parm.useCuda:
batch_data = batch_data.cuda()
encoders = encoders.cuda()
decoders = decoders.cuda()
base = base.cuda()
print("开始")
import time
start = time.time()
z, zI, zW, inters_t, inters_w = encoders(batch_data)
texture, warp, out, _ = decoders(zI, zW, inters_t, inters_w, base)
end = time.time()
print(end - start)
import torch.nn as nn
model_path = "../CelebA_out/2020-03-28_23:13:37/checkpoints"
attr_file_path = "/home/qianqianjun/CODE/DataSets/CelebA/Anno/list_attr_celeba.txt"
image_data_path = "/home/qianqianjun/CODE/DataSets/DaeDatasets"
encoders = Encoders(parm)
decoders = Decoders(parm)
device = torch.device("cpu")
nn.Module.load_state_dict(
encoders,
torch.load(os.path.join(model_path, "encoders.pth"), map_location=device))
nn.Module.load_state_dict(
decoders,
torch.load(os.path.join(model_path, "decoders.pth"), map_location=device))
selecter = ImgSeleter(attr_file_path, image_data_path)
imgs = selecter.getImagesWithoutAttribute("Mustache", ["Male"],
amount=207,
shuffle=False)[200:]
tensor = batchDataTransform(torch.tensor(imgs, dtype=torch.float32), channel=3)
z, zI, zW, inters_I, inters_W = encoders(tensor)
base = getBaseGrid(parm.imgSize, True, tensor.shape[0])
texture, res_warp, out, warp = decoders(zI, zW, inters_I, inters_W, base)
result = torch.cat([tensor, texture, out], dim=0)
saveIntermediateImage(result.data.clone(),
n_sample=21,
nrow=7,
filename="重建结果",
output_dir="out")
texture_inter_interpolation_result=[]
for i in range(len(test_texture_inters)):
texture_inter_interpolation_result.append(
test_texture_inters[i] + lambda_texture_inter_out[i] *
texture_inters[i].repeat(test_texture_inters[i].size()[0], 1, 1, 1)
)
### 变形信息插值
warp_code_interpolation_result= test_warp_code + lambda_deepest_warp * warp_code.unsqueeze(0).repeat(test_warp_code.size()[0], 1)
warp_inter_interpolation_result=[]
for i in range(len(test_warp_inters)):
warp_inter_interpolation_result.append(
test_warp_inters[i] + lambda_warp_inter_out[i] *
warp_inters[i].repeat(test_warp_inters[i].size()[0], 1, 1, 1)
)
# 结果图像重建
base=getBaseGrid(parm.imgSize,True,batch_data.size()[0])
if parm.useCuda:
base=base.cuda()
texture_result,warp_result,result,_=decoders(texture_code_interpolation_result,
warp_code_interpolation_result,
texture_inter_interpolation_result,
warp_inter_interpolation_result,base)
result1=torch.cat([batch_data[:7],result[:7]],dim=0)
result2=torch.cat([batch_data[7:],result[7:]],dim=0)
res=torch.cat([result1,result2],dim=0)
saveIntermediateImage(
img_list=res.data.clone(),
output_dir="out",