[DISCUSS] Adding a PyTorch Frontend

masahi · February 5, 2020, 3:57am

Hi all,

The PR by @alexwong at https://github.com/apache/incubator-tvm/pull/4497 is close to be ready for merge.

At this point I see no good reason not to add this to TVM, so when it becomes ready I’ll go ahead and merge it. Later I’ll send a control flow support to begin work on script models. Thanks.

adb · February 5, 2020, 5:38pm

This is really interesting!

I’m thinking this could enable a straight forward implementation of iterative retraining for networks that will be quantized.

stevezhang2529 · June 2, 2020, 6:21am

Hi All,

I am interested in this work, and I am wondering what is the status of Pytorch --> Relay IR?

Thanks, S.

xiandi · July 23, 2020, 3:13am

NotImplementedError: The following operators are not implemented: [‘aten::values’]

t-vi · July 26, 2020, 11:30am

I don’t think the PyTorch frontend supports sparse tensors at this time.

xiandi · July 26, 2020, 12:46pm

add Dict type in relay ?

t-vi · July 28, 2020, 8:37am

Ah, sorry. I had checked ATen’s values function. But so you already found the thread discussing data types in the runtime / IR, so that bit would seem to come first.

xiandi · August 9, 2020, 7:09am

[Frontend][PyTorch] NotImplementedError: The following operators are not implemented: [‘aten::is_floating_point’, ‘aten::true_divide’]