search

wtliao / text2image

Text to Image Generation with Semantic-Spatial Aware GAN
177 stars 34 forks source link

fix text_encoder #26

Open lcxsnow opened 1 year ago

lcxsnow commented 1 year ago

In the paper, it is mentioned that the text_encoder will be fixed during the image synthesis training. However, in the provided code, the text_encoder is frozen, meaning it won’t be trained or optimized. Did I misunderstand something?

wtliao commented 1 year ago

In the paper, it is mentioned that the text_encoder will be fixed during the image synthesis training. However, in the provided code, the text_encoder is frozen, meaning it won’t be trained or optimized. Did I misunderstand something?

You can set it as trainable in the code. We did ablation studies both for freezing and training the text_encoder.

lcxsnow commented 1 year ago

What do you mean by the text_encoder is frozen ? I thought you have to pre-train the DAMSM model, which place in the code said that ? requires_grad = False means to freeze the model. Yes, pre-train the DAMSM model, and freeze it when you are training GAN.

lcxsnow commented 1 year ago

In the paper, it is mentioned that the text_encoder will be fixed during the image synthesis training. However, in the provided code, the text_encoder is frozen, meaning it won’t be trained or optimized. Did I misunderstand something?

You can set it as trainable in the code. We did ablation studies both for freezing and training the text_encoder.

Another question, I thought FID is more important than IS, because IS can't see the intra. Why do you pay more attention to IS in the paper.