Anirudh257
Question about the episode training
Hello,
I inspired your work and currently following your work and studying. thank you for the great work.
I noticed that in your paper "Spatio-Temporal Relation Modeling for Few-Shot Action Recognition", you mention that your episodic training paradigm follows Laenen et al. (NeurIPS 2021) [16]:
"In this work, we follow an episodic training paradigm as in [16], where few-shot tasks are randomly sampled from the training set for learning the C-way K-shot classification task in each episode."
However, after reviewing Laenen et al. (NeurIPS 2021), I found that their primary contribution was in removing episodic training and introducing a non-episodic training strategy with NCA Loss. Their work argues that episodic training is inefficient and that a fully non-episodic approach leads to better performance.
(its link is here : https://github.com/fiveai/on-episodes-fsl, see the "src/train/loss.py")
Could you clarify in what way your episodic training paradigm follows their work? Specifically:
1. Are there aspects of Laenen et al.'s methodology that directly influenced your episodic training implementation? 2. Did your model incorporate any of their non-episodic training strategies or NCA loss, or was the reference more general?
I appreciate your time and look forward to your response!
Best regards,
