search

RenYurui / StructureFlow

Code for paper "StructureFlow: Image Inpainting via Structure-aware Appearance Flow"
Other
222 stars 42 forks source link

Question: inconsistency about norm_flow and norm_conv #7

Closed HuaZheLei closed 5 years ago

HuaZheLei commented 5 years ago

In the 'models.py', the parameters for FlowGen is self.flow_param = {'input_dim':3, 'dim':64, 'n_res':2, 'activ':'relu', 'norm_conv':'ln', 'norm_flow':'in', 'pad_type':'reflect', 'use_sn':False} However, in the 'network.py', the default parameters for FlowGen is class FlowGen(nn.Module): def __init__(self, input_dim=3, dim=64, n_res=2, activ='relu', norm_flow='ln', norm_conv='in', pad_type='reflect', use_sn=True): The parameters of norm_flow and norm_conv are inconsistent. I wonder if it is a bug.

RenYurui commented 5 years ago

It is not a bug. Actually, we find that using 'in' in flow_net will introduce artifacts to the final results, while 'ln' has better performance than 'in' in this task.

HuaZheLei commented 5 years ago

It is not a bug. Actually, we find that using 'in' in flow_net will introduce artifacts to the final results, while 'ln' has better performance than 'in' in this task.

Hi, thanks for your reply. As you mentioned, 'norm_flow' should be 'ln' rather than 'in'. But in 'model.py', you send 'in' to 'norm_flow', which makes the defalut parameter 'ln' invalid. I think in the traning process, 'norm_flow' should be 'In' and 'norm_conv' should be 'in'. Am I right?

RenYurui commented 5 years ago

Sorry for misleading. The parameters of the code are correct. Since the Texture Generator (named in the paper) outputs the final results directly. Therefore, we find its normalization need to be 'ln' for better performance. Therefore, the parameter 'norm_conv' in 'FlowGen' should be 'ln'. The parameter 'norm_flow' assigns the normalization of flow fields generator of Texture Generator. It can be 'ln' or 'in'.

In 'model.py' the correct parameters are sent to the network. However, the default parameters in 'network.py' are incorrect. We will correct them.

HuaZheLei commented 5 years ago

Sorry for misleading. The parameters of the code are correct. Since the Texture Generator (named in the paper) outputs the final results directly. Therefore, we find its normalization need to be 'ln' for better performance. Therefore, the parameter 'norm_conv' in 'FlowGen' should be 'ln'. The parameter 'norm_flow' assigns the normalization of flow fields generator of Texture Generator. It can be 'ln' or 'in'.

In 'model.py' the correct parameters are sent to the network. However, the default parameters in 'network.py' are incorrect. We will correct them.

Cool!