Discrepancy in Model File Size from Pre-trained Models Provided

[Peng-weil]

Hello, thank you for your work !

I've been using the pre-training configuration files (panoptic) and the code process provided in your GitHub repository to train a model using MaskFormer on ZSP MSCOCO dataset. After completing the pre-training, I noticed that the size of the generated .pth model file is significantly smaller (176MB) compared to the pre-trained model file size (approximately 500MB) provided in your Google Drive.

I followed the instructions and configurations exactly as stated, but the difference in model file sizes is quite substantial, and I am concerned about whether there might be an underlying issue with the training process or model configuration that I overlooked.

Could you possibly provide some insights or assistance on this matter? Any guidance on potential reasons for this discrepancy and how to address it would be greatly appreciated.

Thank you for your support and looking forward to your response.