How about the support for RNN?

[justdark]

Not all thnets supported layers have been implemented

is the LSTM layer supported? I use thexport on my model,It seems the LSTM with embedding layer have been decomposed to a extreme complicated structure with more than 7000+ functions, even exceed the default maxium recursive depth of Python. (the length of input series is 400).
some of them

<torch.autograd._functions.basic_ops.Add object at 0x10fe1c960>
<torch.autograd._functions.basic_ops.Mul object at 0x10fe1e5c0>
<torch.autograd._functions.pointwise.Sigmoid object at 0x10fe1e138>
<torch.autograd._functions.tensor.Chunk object at 0x10fe1e050>
<torch.autograd._functions.pointwise.Tanh object at 0x10fe1e308>
<torch.autograd._functions.tensor.Chunk object at 0x10fe1e050>
<torch.autograd._functions.tensor.Concat object at 0x10ee39308>

is that affordable for thnets? Can I use the model file directly for production?

[mvitez]

No, LSTM is not supported. thexport does not detect loops, so it failed to detect it and give an error.

[justdark]

I use

import sys
sys.setrecursionlimit(10000)`

can forcibly export the graph,is that not usable? I think it automaticly act like loops.

[mvitez]

No, it's not usable.

[justdark]

ok~~ I'll figure out how to handle it. tks~~