Jean Kossaifi

CP memory issue with high rank

3

We should incorporate the l2 regularization more efficiently without actually creating a diagonal matrix (Id*reg_l2) but instead directly augment the pseudo-inverse: https://github.com/tensorly/tensorly/blob/1128fea84bd5a26129cb990987ca2a9d90a701e9/tensorly/decomposition/_cp.py#L398

We reverted to the current version of MXNet. We want to update to MXNet 2.x for full numpy interface but it currently doesn't seem to be working https://github.com/apache/incubator-mxnet/issues/18695

Tests to fix

2

Some of the tests are randomly failing and either they, or the underlying function being tested, needs to be fixed, this issue tracks the progress and remaining work to be...

Incremental/online SVD

1

We already have partial-SVD for computing only k-first eigenvals and randomized SVD thanks to @merajhashemi for fast approximation. Would be good to add incremental SVD when we want better approximation...

The test sometimes fails. Detailed traceback: ``` __________________________ test_validate_tucker_rank ___________________________ def test_validate_tucker_rank(): """Test validate_tucker_rank with random sizes""" tol = 0.01 tensor_shape = tuple(np.random.randint(1, 100, size=5)) n_param_tensor = np.prod(tensor_shape) # Rounding...

Dot product between ones and COO tensor

2

Currently, `sparse.ones` returns a sparse tensor with a fill value of 1 which seems incompatible with the default value (0), at least when taking a dot product. Minimal code to...

Procrustes alignment: accepting different number of points.

3

Accepting different number of parameters by padding the smallest shape with zeroes (or the mean of the shape if said mean is not zero to not modify the overall behaviour).

CP function

8

CP via ALS is probably the most used function in TensorLy and comes with lots of options. One issue is that due to these successive additions, bugs (see e.g. this...

Better random init of factorized tensors

1

I wrote a better init in [TensorLy-Torch](https://github.com/tensorly/torch/blob/main/tltorch/factorized_tensors/init.py)) we should probably bring it in at some point, it allows the *reconstruction* to have zero mean and given standard deviation.

Doc and examples

4

We probably need to add some more examples and doc for new features in TensorLy. Would be great for instance to have an example for @earmingol @hmbaghdassarian as a user...