> Can we use the valid dataset(valid.zip contains 3,951 SD files.) together with training dataset(dev.zip contains 99,776 SD files) updated on 7/30 in phase 1 for trainning to predict the dataset(the test.zip contains 15,760 SD files) in phase 2?

If we cannot use both valid and training dataset for training, how can the organizor distinguish the of competitor's models of not using the both from those who just used the training dataset only ?

Posted by: Sebs @ Sept. 23, 2019, 11:40 p.m.

one question, how can you remove yourself from the leaderboard?

Posted by: EversChen @ Sept. 24, 2019, 8:25 a.m.

because we didn't make any progress recentlty and feel deseparete to see the final results.
cry out loud T.T

Posted by: Sebs @ Sept. 24, 2019, 9:33 a.m.

Personally, if you have other ideas for splitting validation sets, splitting them after merging should be possible. In addition, after determining the model parameters and estimating the number of training cycles, it is feasible to use all the data for training.

Posted by: Masker_L @ Sept. 24, 2019, 10:13 a.m.

Personally I think the button of "Remove From the Leaderboard" is a bug. Some top ranked results are hidden so nobody knows their real rank until the last seconds of this competition.

Posted by: zoro @ Sept. 24, 2019, noon

If we can use the valid dataset for training for not, a bug too. T.T

Posted by: Sebs @ Sept. 24, 2019, 12:10 p.m.

I felt like training with valid data should help a lot. Most of molecules on both valid and test data set are big molecules.

Posted by: zoro @ Sept. 24, 2019, 12:35 p.m.

Do you mean use about 3900 valid dataset only for training?

Some days back qlab force all the test result to be public, and i believe it's good. that means you can't hide any more, it will be fair to all of us.
So you are right, it will be another way to hide if the remove works, you can report this bug to qlab if you are the possible of top 3 of the candidates.

for me, i didn't get any progress for several weeks, anyway it's also good to see how you guys works at the end of holiday, let's see.

Posted by: EversChen @ Sept. 25, 2019, 5:21 p.m.

Top 3 are not for me anymore maybe because everyone is hiding the MAE.
I updated my latest MAE score immediately on the leaderboard to be the Top 1 for about 1 month, and only to find everyone is hiding their progress.
It's very sad for all of us.
It's a bug on my ground ,too.
I think every result should be public to the competitors.
With few GPU and 10 days left, I have no choice but to ask a question to merge dev and valid like this, and for sure some competitors are doing this to be the top rankers.
Waiting for the reply.
Love u all.

Posted by: Sebs @ Sept. 25, 2019, 5:39 p.m.

agree, latest MAE score should be shown in public

Posted by: qlab @ Sept. 25, 2019, 6:59 p.m.

By the way, where can I have free GPUs?
I am a student in chemistry and personaly I only have one GTX 1050ti on my windows.
There are few days left and I still got so many submitting chances remain.
Can anybody help me and gave me some advice.
I don't have too much money to rent any computing sources too. T.T

Posted by: Sebs @ Sept. 25, 2019, 7:23 p.m.

Congratulation! @Sebs. Thank you for your advice

Posted by: Josephxu @ Oct. 7, 2019, 11:57 p.m.

@Josephxu, thank u, champion of Merck Cup, I know you for a long time already.
hope to see you and your teammates again in nanjing university.
There are lots of interesting chemistry things I would like to seek advice from you.

Posted by: Sebs @ Oct. 8, 2019, 12:15 a.m.

Cry~~~, I tried to add some quantum properties like ACSF, CM calculated by Dscribe toolkits. However it is hard to train the models well. And my self-built evaluation system comes out some bugs at the end. If I have an opportunity to go NJU, I would be very glad to talk with you about those interesting chemistry. And consult you about the solutions. ^_^

Posted by: Josephxu @ Oct. 8, 2019, 1:05 a.m.
Post in this thread