Decouple the conv data format from the input feature layout

[tommadams]

On the one hand, it's more efficient for the CPU to generate NCHW input features.
On the other, TensorFlow supports NHWC convolutions on a wider variety of platforms.

When I added support for NHWC, I coupled the conv data format to the input feature layout.

We should add another option to support different tensor layouts for input features and convolutions, inserting transpose operations as necessary.

[brilee]

Layout also has a performance impact on TPU, so you probably want to consider that as well.

…

On Thu, Mar 12, 2020 at 5:09 PM Tom Madams ***@***.***> wrote: On the one hand, it's more efficient for the CPU to generate NCHW input features. On the other, TensorFlow supports NHWC convolutions on a wider variety of platforms. When I added support for NHWC, I coupled the conv data format to the input feature layout. We should add another option to support different tensor layouts for input features and convolutions, inserting transpose operations as necessary. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#974>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAKCKFRYJLSCTEI4ZPFL533RHFFQ3ANCNFSM4LGVNNGQ> .

[tommadams]

TPU apparently doesn't support NCHW conv at all...

[SHKD13]

Is it purely theoretical topic or the part of MiniGo v18 building process?

[tommadams]

This is more for MLPerf, which uses smaller models and has a much higher CPU:GPU compute load than a regular Minigo run.

[SHKD13]

MLPerf is a kind of a "lab", where you can try some algorithm's changes and improvements on a smaller Networks before implement it for a regular full sized run? Sorry, if I get something wrong :)

[tommadams]

MLPerf is a suite of machine learning benchmarks that many major tech companies collaborate on. Minigo is the reinforcement learning benchmark for MLPerf.

[tommadams]

See https://github.com/tensorflow/minigo/tree/master/ml_perf for the implementation. It's basically a smaller & simplified version of the full pipeline

[SHKD13]

Thanks for clarification and links! Now I can see that I didn't know about the MLPerf background of MiniGo.
I thought, MG is enthusiastic and independent Go project like LZ or KataGo. But missed the point about something bigger behind, which drives the MiniGo team