search

blackprotoss / GSDM

Text Image Inpainting via Global Structure-Guided Diffusion Models (Accepted by AAAI-24)
MIT License
46 stars 4 forks source link

weird training results #7

Closed LonglongaaaGo closed 3 months ago

LonglongaaaGo commented 3 months ago

Hi @blackprotoss

Thanks for your awesome work. I tried to train the structural prediction model for 50 epochs on the ST dataset. But the testing results are not perfect, could you give some advice for improving?

image
blackprotoss commented 3 months ago

we have re-checked our released code and confirmed the reproducibility. Since we have not encountered such a problem before, one possible guess is that it might be related to the image channels or RGB issues.

LonglongaaaGo commented 3 months ago

Hi @blackprotoss , Thanks for the reply. One thing I am not sure of is that the original CNN model only returns the final RNN features. I also changed the code to run it successfully by adding final feature maps from the last layer conv. As shown below, I am still determining if i am correct, could you help with this? I appreciate!

image

blackprotoss commented 3 months ago

My mistake, I forgot to upload one folder crnn (some critical code files here). The files have been updated.

LonglongaaaGo commented 3 months ago

Hi @blackprotoss , Got! thank you so much for the help!