Jean Kossaifi

It's also specified in the DataLoader: https://github.com/face-analysis/emonet/blob/47006243857718250907eed4c9b3c39e1578c955/emonet/data/affecnet.py#L10

Yes, indeed, this one can be done with a sum, I just wrote this to have a minimal reproducible example. The issue I had was with general tensors, not necessarily...

Did you check that the computation is indeed done on GPU? It also depends on the size (and rank) of your tensors. It is possible that the NN version is...

Yes, I agree, it is better to prioritise consistency and not surprise end users with unexpected behaviours. We can either remove it completely for now and think of a better...

I'm +1 for moving to tenalg. An important note is that I actually already experimented with a few implementations (see the docstring): https://github.com/tensorly/tensorly/blob/1128fea84bd5a26129cb990987ca2a9d90a701e9/tensorly/cp_tensor.py#L650 The reason we ended up using this...

Thanks @mfproto , this is great suggestion. I don't have experience adding layers to ONNX, if you know how to would be awesome if you could start a PR!

Hi @ashim95 -- factorized tensors are just that: higher-order tensors in factorized form. Tensorized matrices are *matrices* that are tensorized. That tensorized form is expressed in factorized form. You can...

I think this broke the doc

Going to close, feel free to reopen if random init didn't solve your issue.

Hi @wanglu2014 -- in that case, I recommend you try the [randomized SVD](https://tensorly.org/dev/modules/generated/tensorly.decomposition.RandomizedCP.html#tensorly.decomposition.RandomizedCP) or even a gradient based decomposition, e.g. using [TensorLy-Torch](http://tensorly.org/torch/dev/).