Spaces:
Running
Conversion of Model with fp16 weights to ONNX
Hi there, amazing work on your GAN project. I have used your convert-stylegan2-model.ipynb to convert my stylegan2 model weights which is in currently in mixed precision to onnx but i have been unable to do so as the image look like this after the conversion to onnx.
I know this is due to some of the weights being in fp16 but i dont really know how to resolve this issue. Your expertise and knoweldge in this issue is kindly appreciated. Thanks alot.
You need to finetune the model on fp32.
Hi, how should I go about finetuning the model in fp32. What I think I should be doing is to copy the weights and convert them to fp32 and retrain the model based on the settings that I have included in the training process is that the right process? Also how long did training last for the finetuning process may I ask? Thank you so much for your prompt replies.
Yes, just simply resume training and set --fp32=1 .
The generated images will return to normal within 100kimg.