tanganke
commented
2 weeks ago - Use this repository:
To reproduce the results in the paper use this repository, you can refer to the README file at the results directory, which describes the shell command.
The checkpoints of pre-trained model and fine-tuned models can be downloaded from google drive, these checkpoints are from the work Editing Models with Task Arithmetic. These models are from open-clip library of version 2.0.2.
See https://github.com/mlfoundations/task_vectors/issues/1 for split definition of datasets.
- Use FusionBench:
We also fine-tuned some CLIP models by ourselves and publish the checkpoints public at HiggingFace:
We also reimplement this algorithm in FusionBench, by running these command, the models and datasets should be downloaded automatically from huggingface. The experimental results of multi-task model fusion are basically the same with the open-clip implementation here.
If you have any other difficulties, feel free to contact us.
Hello,I am very interested in your work. However, due to my limited knowledge in this area, I am not sure how to train the code or which file to start with. Could you please guide me on this? I would really appreciate it. Thank you very much!