Pass around saved/exported models everywhere

[amj]

instead of having our selfplay workers constantly look for the .data-00000-of-00001, .index, .meta files, we should properly export our graph. Not only would this make a lot of our scripts cleaner, it'd be a good way to run some optimizers etc. prior to farming it all out.

[killerducky]

Running an optimizer, is that like tensorrt? The leela-chess guys had some great results with that.

[amj]

I hadn't heard of that -- i was pointed towards toco . I'll go look up tensorrt

[killerducky]

leela-chess benchmarks:
glinscott/leela-chess#52 (comment)

Looks like they are working on integrating into TF:
tensorflow/tensorflow#16253

[amj]

Yes, that's it almost exactly. out_graph = tf.graph_util.convert_variables_to_constants(n.sess, n.sess.graph.as_graph_def(), ["policy_output", "value_output"])

(where n is a dual_net.DualNetworkTrainer object that's initialized)

[Error323]

Hi,

Indeed, TensorRT seems very promising, unfortunately as of now you won't be able to export your tensorflow network to uff format for importing as the Merge operator is unsupported. This leaves three possibilities:

• Create a network clone in TensorRT using their C++ API and use exported weights from tensorflow
• Wait for TensorRT to support tensorflow 1.4 and beyond
• Wait for the PR mentioned above to be merged

The PR probably makes the most sense. Nice project! (offtopic: @amj it looks like you're from the AlphaGo documentary, awesome!)

[amj]

@Error323 thanks for the great summary, much appreciated! We'll check it out. I think it's a pretty open space w.r.t. what sort of optimizations are possible so we might not lock in to TensorRT immediately. (I am the guy from doc, yep -- had to study up on NNs since then ;) I'm really watching the leela chess project with great interest!

[brilee]

I think we've done this in the course of implementing C++ selfplay?