search

LeelaChessZero / lczero-training

For code etc relating to the network training process.
148 stars 119 forks source link

404 at training data link #122

Open mhorton18 opened 4 years ago

mhorton18 commented 4 years ago

I am getting a 404 error at the link where the training data ".tar.gz" file exists. Any update on where to find the training data and how to use it?

mooskagh commented 4 years ago

Which link are you trying to follow?

mhorton18 commented 4 years ago

From the readme: http://lczero.org/training_data

mooskagh commented 4 years ago

The entire readme is outdated, I suggest joining the Discord chat at http://lc0.org/chat and asking there if you want to train your network.

Dboingue commented 4 years ago

what if the intent is to access all the data produced in past training instances? Training your own network, may not be the only use for such access. I am just checking that there will be no loss of data. As I do intend on understanding how to access it at some point. Hopefully not misunderstanding the issue and comments here.

mooskagh commented 4 years ago

The training data is located at http://data.lczero.org/files/training_data/, http://data2.lczero.org/files/training_data/, http://data4.lczero.org/files/training_data/ (also at data3, data5 and data6 but they currently don't have web server set up. They have training data from older runs).

The combined size of the training data that we keep is ~13TB. We do delete older training data because we don't have space and noone really needs them. I think we deleted at least 15TB of training data since the beginning of the project.

Dboingue commented 4 years ago

Thank you for this clear answer. I suspected as much, and it makes sense.  However, I would assume that all networks still underactive developments or testing do have their corresponding training data accessible somewhere.  It is important to me to be able to keep the relationship between the data and the corresponding networks, as I would like to view what happens at different layers (I am curious that way), if that has not been already studied in the context of chess. Also, I understand that funding may not be easy to get for such a non-enthusiastic concern (English lacking here), but what could people of similar perspective as mine do to help for the retention of such data and its access? Maybe I could make this its own issue? somewhere. chat, forum, or GH issue? thanks and regards.  great project.

Le 25/04/2020 à 15:34, Alexander Lyashuk a écrit :

The training data is located at http://data.lczero.org/files/training_data/, http://data2.lczero.org/files/training_data/, http://data4.lczero.org/files/training_data/ (also at data3, data5 and data6 but they currently don't have web server set up. They have training data from older runs).

The combined size of the training data that we keep is ~13TB. We do delete older training data because we don't have space and noone really needs them. I think we deleted at least 15TB of training data since the beginning of the project.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/LeelaChessZero/lczero-training/issues/122#issuecomment-619430032, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADNEZIHKJL4OUAZGR6PJ4TTROM3LBANCNFSM4MK7D6SQ.