Which is the best way to test a new architecture


#1

Hi!

I followed all the tutorials and everything worked, but when I try to import a very specific NN architecture of mine I receive several errors. I’ve tried to import an implementation from Keras and from TensorFlow but both had unsupported characteristics. So I am thinking of reimplementing the architecture to work with TVM, but my question is: is it better to write it directly with the TVM framework? If so, where can I found some examples of it? If it is better to use another framework, which one has better support for import to TVM?

Thank you!


#2

It’s better to write in other frameworks because currently tvm does not support training.

I think MXNet has the best compatability. You can find some examples in this folder https://github.com/dmlc/tvm/blob/94acff30e82f9352e1652bd81be260b672419aea/nnvm/python/nnvm/testing/dqn.py#L29
Most of them are ported from mxnet.

If the “unsupported characteristics” are new operators, you can try to add these operators to tvm.