search

USTC-JialunPeng / Diverse-Structure-Inpainting

CVPR 2021: "Generating Diverse Structure for Image Inpainting With Hierarchical VQ-VAE"
MIT License
175 stars 19 forks source link

the condition h is NOT used in the train_structure_generator.py #3

Closed John-niu-07 closed 3 years ago

John-niu-07 commented 3 years ago

How do you train the structure_generator? I see the code in file train_structure_generator.py, where the condition h is totally ignored.

USTC-JialunPeng commented 3 years ago

The condition 'h' is injected into each residual gated convolution layer of the autoregressive network. The corresponding code is with arg_scope([nn.gated_resnet], num_res_filters=nr_res_channel, nonlinearity=resnet_nonlinearity, dropout_p=dropout_p, num_head=nr_head, h=h) in the '/net/structure_generator.py' (Line 75)

John-niu-07 commented 3 years ago

The condition 'h' is injected into each residual gated convolution layer of the autoregressive network. The corresponding code is with arg_scope([nn.gated_resnet], num_res_filters=nr_res_channel, nonlinearity=resnet_nonlinearity, dropout_p=dropout_p, num_head=nr_head, h=h) in the '/net/structure_generator.py' (Line 75)

many thanks :-)