Some questions about Track1

[Frozen...@outlook.com]

Dear organizers:

We are team Meta_Learners. Thanks for the great competition! We have some questions:

1. When will the competition ends? The competition platform shows that currently there are 3 days and 21 hours to go, which means it ends at Oct. 1. midnight. However, the website says it ends at Oct. 2. midnight. Which one is correct?
2. Can we use the `dataset_spec.json` at the final phase (to get the meta-train class number)?

Many thanks!

Sincerely,

Meta_Learners

2021.09.28

[Adrian EL BAZ]

Hello team Meta_learners! 

Indeed the competition ends Oct 2nd midnight UTC. As of now, it ends in 3 days and 5 hours which match: I've changed the date limit. Thank you for pointing this out!

2- Regarding your second question, I've done a few minor changes in the metadl code (last commit on neurips2021 branch) to get the number of meta-train classes automatically in case you are using batch mode. If this is the case (batch mode) then you can simply change your __init__() method in MyMetaLearner to __init__(**kwargs) and then initialize a dedicated variable like: self.nbr_metatrain_classes = kwargs["nbr_classes"]
You can find an example in the updated transfer baseline. 

In any case, anything that is working in the codalab platform should be working during the final phase (even accessing meta_train classes).

By the way, I noticed some improvements since I reset the workers the other day, did you as well ?

Best,

Adrian

--
You received this message because you are subscribed to the Google Groups "metalearningchallenge" group.
To unsubscribe from this group and stop receiving emails from it, send an email to metalearningchal...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/metalearningchallenge/SY6P282MB3718A28FCA6DACC6DBD08E00B8A89%40SY6P282MB3718.AUSP282.PROD.OUTLOOK.COM.

[超宇 关]

Nice to hear from you Adrian!

Thanks for the convenient interface! By saying "In any case, anything that is working in the codalab platform should be working during the final phase (even accessing meta_train classes).", is that means that we can access the dataset_spec.json​ of meta-train dataset in the final phase? We want to first train in batch mode for a while, and then switch to episode mode to make finer parameter update, thus may need to know the class number in episode mode (We can easily create a batch from an episode, but is hard in reverse).

As for the workers, we haven't made many successful submissions recently and cannot tell whether there are any improvements. We'll make more submissions recently and will let you know if they become better!

Thanks again!

Meta_Learners

2021.09.30

---

From: Adrian EL BAZ <eb.ad...@gmail.com>
Sent: Thursday, September 30, 2021 03:23
To: 关 超宇 <Frozen...@outlook.com>
Cc: metalearni...@googlegroups.com <metalearni...@googlegroups.com>
Subject: Re: Some questions about Track1

 

[Adrian EL BAZ]

Hello!

1- Yes you will be able to access the dataset_spec.json file as in the current phase. Since your solution seems to be a particular case, it may be better to not use the changes that we've made. 

2- Alright let me know about any changes.

Best,

Adrian