Tomasz Sobczyk

- [ ] some results regarding net size (https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY/edit#heading=h.p7t1qfux64dc exps 84, 85, 86, 89) - [ ] about data used to train the nets (filtering, mirroring and stuff) - [x]...

[RFC] Initial network storage format spec.

2

Previous attempt here: https://github.com/glinscott/nnue-pytorch/pull/64 Since then I realized that we need something simpler. So this is about as simple as we can get while preserving the required functionality. See `docs/network_storage_format.md`.

Results of my recent benchmark (https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY/edit#heading=h.amq1v486w8yp) suggest that non-power-of-2 FT sizes are not ideal, and we could benefit from splitting the FT into two parts, the main part and the...

This patch adds a way to disable factorization of the L1 and the FT layer. Currently it's only possible from the code, with no option exposed to the user. A...

Use two streams, one per FT slice.

4

Since the computation of the two slices of the feature transformer output are independent we can try to run them on separate streams. On my GTX750 I can notice a...

This means that the sizes in .ckpt cannot be inferred and must match the original script. This makes it harder to explore different sizes. The sizes of the layers should...

serialize.py doesn't correctly read feature_set from ckpt

1

ckpt should serialize the feature_set used, but the net is always initialized with the default, which is not always possible.

get rid of some nodchip code

5

https://github.com/glinscott/nnue-pytorch/blob/master/lib/nnue_training_data_formats.h contains `PackedSfenValue`, `BitStream`, and `SfenPacker` which are taken from the nodchip trainer. Rewrite it so that there are no licensing issues. This might be a non-issue, but it's questionable....

Undocumented behaviour of bitshuffle when blocksize is not a multiple of 8*type_size.

2

Bitshuffling fallbacks to memcpy when `size % 8 !=0`. This is fixed in version 3, so I presume it's not fixed here for valid reasons. I would propose at least...

Passwords are collected, sent and stored in plaintext [RFC]

1

https://github.com/glinscott/fishtest/blob/e16909c4994ac8b1533bda5bd41588424c6d2ad7/server/fishtest/userdb.py#L86 maintainer's response: > This is hinting a password reuse policy. We have daily db backups dating several years, so I suggest to change the master password of your sensible...