Hello, I hope this message finds you well. I have a question regarding your paper. Could you kindly clarify whether the online encoder or the target encoder is used for feature extraction in downstream tasks after the model has been trained? I noticed that the article does not specify this. When I looked at the notebook/feature_visualization.ipynb file, I saw that the feature extraction uses encoder_k, which is the target encoder. Does this mean that the target encoder is also used for other evaluation experiments? In the BYOL paper, they mentioned retaining the online encoder. I was wondering if you could explain the differences in the experiments if we keep the online encoder versus the target encoder. Your clarification would be very helpful for my understanding. Thank you!
Hello, I hope this message finds you well. I have a question regarding your paper. Could you kindly clarify whether the online encoder or the target encoder is used for feature extraction in downstream tasks after the model has been trained? I noticed that the article does not specify this. When I looked at the notebook/feature_visualization.ipynb file, I saw that the feature extraction uses encoder_k, which is the target encoder. Does this mean that the target encoder is also used for other evaluation experiments? In the BYOL paper, they mentioned retaining the online encoder. I was wondering if you could explain the differences in the experiments if we keep the online encoder versus the target encoder. Your clarification would be very helpful for my understanding. Thank you!