Looking for 9x9 model files in .minigo file format

[walmsley]

I'm looking to run a 9x9 model in MiniGo's C++ GTP program, and it seems as though the program only accepts the new .minigo file format.

I've found 9x9 model files at e.g. https://console.cloud.google.com/storage/browser/minigo-pub/v3-9x9/models/?prefix=000427 but the only file formats provided are .meta, .index, and .data-* files. Do .pb or .minigo files exist somewhere for these models? Otherwise, are there instructions somewhere on how to properly convert these files to a .minigo file format?

Thanks!

[sethtroisi]

It doesn't seem that we have converted the old 9x9 models.

@tommadams or I will try to do.

@tommadams is this the general jist

for MODEL_PATH in `gsutil ls ...` do
  python freeze_graph.py  --model_path=$MODEL_PATH
  python3 oneoffs/wrap_model.py \
    --src_path $MODEL_PATH.tflite \
    --dst_path $MODEL_PATH.minigo \
  --metadata engine=lite,input_features=agz,input_layout=nhwc,input_type=float,board_size=9
done

note there's a small typo in cc/README.md see s/minifo/minigo/

[tommadams]

I'll do this now

[tommadams]

Oh, it looks like we don't have frozen .pbs for those 9x9 models.

We'll need to freeze them. Does anyone know what the architecture flags used were?

[sethtroisi]

That's a question for @amj or @brilee

[brilee]

I don't recall that our old 9x9 model were even that good. Tom, didn't you spend some time MLPerf'ing some 9x9 runs? Those might be better if you leave it running for a day or two, and it'll spit out the right file format without any extra grungework.

On Thu, Apr 23, 2020 at 5:47 AM Seth Troisi notifications@github.com wrote:

That's a question for @amj https://github.com/amj or @brilee https://github.com/brilee

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/tensorflow/minigo/issues/989#issuecomment-618300808, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAKCKFSMTT7L76QKXX3G4BTROAFCHANCNFSM4MONXAKA .

[tommadams]

I'm digging through the github history to see if I can pull out the appropriate architecture flags.

The v0.6 MLPerf was 9x9 but the benchmark was set up to train a model that plays something "reasonable" in as short a time as possible. The trained models are pretty weak.

[tommadams]

Ok, I've uploaded .minigo versions of all the v3-9x9 models to the GCS bucket.

When you're building the C++ binaries, remember to pass --define=board_size=9 because the board size is a compile-time constant. For example:

gsutil cp gs://minigo-pub/v3-9x9/models/000496-polite-ray.minigo .
bazel build  -c opt  --define=board_size=9  cc:simple_example
./bazel-bin/cc/simple_example  --model=000496-polite-ray.minigo