Is there any easy way to access vector of activations present in value / policy heads for given chess position

[int8]

I have PGN/fen (or any format of chess position) and I want to pass it forward through value/policy network of lc0 to get activation vector(s) - any hint on how to achieve that ?

[Naphthalin]

There are ways people have done this -- best to join the Lc0 Discord server and ask there. The GitHub issues are only for actual issues with the Lc0 engine binary.

[int8]

thanks - on my way :)

[int8]

Ok I will actually leave my comment here in case anyone is looking for solution in the future:

most likely steps in the right direction:

• convert chosen model to ONNX via lc0 leela2onnx command
• use onnx python package to modify converted model to include intermediate layer as output

import onnx
model = onnx.load("converted_model.onnx")
print([n.name for n in model.graph.node]) # this will list all intermediate nodes 
# now let's say I want to read values from `/block9/conv2/relu` (layer that goes to policy/value heads):

from onnx import ValueInfoProto
intermediate_layer_value_info = ValueInfoProto()
intermediate_layer_value_info.name = '/block9/conv2/relu'
model.graph.output.append(intermediate_layer_value_info)
model.graph.output.extend([intermediate_layer_value_info])
model.save("modified_model.onnx") 

• load modified model

import onnxruntime as ort
ort_sess = ort.InferenceSession('modified_model.onnx')
outputs = ort_sess.get_outputs()
print([o.name for o in outputs]) # now output will include intermediate layer 

• run inference (this is tricky because it requires preparation of input for the model, at this point I will need mapping of fen into that input w.r.t https://github.com/LeelaChessZero/lc0/blob/master/src/neural/encoder.cc#L134 (which is in progress)

I will try to update this issue

[bastiansg]

Hi! @int8!

I'm very interested in this topic; I wanted to know if you managed to encode a specific chess position using the Leela Chess Zero neural network.

[mooskagh]

One of the ways to encode the position is to use python bindings as described here: #1261.

It's also not that hard to encode it manually. The mentioned function https://github.com/LeelaChessZero/lc0/blob/master/src/neural/encoder.cc#L134 has some bits of complicated logic, but they all were used for experimental input plane encodings. All modern networks use INPUT_CLASSICAL_112_PLANE encoding which is fairly simple (and I believed it's exactly the same as in the original AlphaZero paper).

In any case, it will be faster to find solution in our Discord chat at http://lc0.org/chat.