Thanks for your great work! I have a couple of questions regarding the model configurations and checkpoint availability in the context of the UniTS_zeroshot_newdata.sh script.
1. Model Dimensionality in Zero-Shot Setting:
Why is the d_model parameter set to 64 for the zero-shot setting, while in the original multi-task pretraining setting, it is set to 128? Is there a specific reason behind this choice, or is it related to the performance or efficiency considerations?
2. Checkpoint Availability:
I greatly appreciate the series of checkpoints you've provided in Pretrained weights. However, is it possible to make the checkpoint in the zero-shot setting publicly available?
Hello,
Thanks for your great work! I have a couple of questions regarding the model configurations and checkpoint availability in the context of the UniTS_zeroshot_newdata.sh script.
1. Model Dimensionality in Zero-Shot Setting: Why is the
d_model
parameter set to 64 for the zero-shot setting, while in the original multi-task pretraining setting, it is set to 128? Is there a specific reason behind this choice, or is it related to the performance or efficiency considerations?2. Checkpoint Availability: I greatly appreciate the series of checkpoints you've provided in Pretrained weights. However, is it possible to make the checkpoint in the zero-shot setting publicly available?
Thank you for your time and assistance.