Pretty print function for dense tensors

[tgkolda]

Would it be possible to add a pretty print function for dense tensors. Here is what I have in mind:

def pretty_print_tensor(X, fmt="10.4f", name="Slice"):
    if not isinstance(X, ttb.tensor):
        raise ValueError("Input must be a pyttb tensor")

    # Get the shape of the tensor
    shape = X.shape

    if name == "Slice":
        print("Tensor is of shape "+ " x ".join(map(str, shape)))
    else:
        print(f"{name} is a tensor of shape " + " x ".join(map(str, shape)))


    # Iterate over all possible slices (in Fortran order)
    for index in np.ndindex(shape[2:][::-1]): # Skip the first two dimensions and reverse the order
        index = index[::-1] # Reverse the order back to the original
        # Construct the slice indices
        slice_indices = (slice(None), slice(None)) + index
        slice_data = X[slice_indices]
        print(f"{name}(:, :, {', '.join(map(str, index))}) =")
        array = slice_data.data
        for row in array:
            print(" ".join(f"{val:{fmt}}" for val in row))
        

So pretty_print_tensor(X,fmt="2d",name="X") produces output like this:

X is a tensor of shape 3 x 3 x 2
X(:, :, 0) =
 3  9  1
 8  2  1
 4  3  9
X(:, :, 1) =
 6  9  5
 5  6  4
 1  4  1

Or pretty_print_tensor(X,fmt="5.1f") for a different tensor produces:

Tensor is of shape 3 x 4 x 3 x 2
Slice(:, :, 0, 0) =
  1.0   7.0   5.0   5.0
  8.0   9.0   1.0   7.0
  4.0   5.0   3.0   8.0
Slice(:, :, 1, 0) =
  4.0   9.0   9.0   9.0
  1.0   2.0   1.0   3.0
  3.0   5.0   6.0   5.0
Slice(:, :, 2, 0) =
  9.0   7.0   2.0   5.0
  2.0   7.0   5.0   4.0
  5.0   5.0   4.0   8.0
Slice(:, :, 0, 1) =
  7.0   3.0   6.0   7.0
  3.0   5.0   4.0   4.0
  6.0   4.0   5.0   9.0
Slice(:, :, 1, 1) =
  2.0   4.0   4.0   7.0
  7.0   6.0   1.0   5.0
  4.0   5.0   1.0   7.0
Slice(:, :, 2, 1) =
  9.0   3.0   9.0   5.0
  7.0   6.0   9.0   8.0
  1.0   3.0   8.0   2.0

[tgkolda]

Maybe it's a function in the tensor class, in which case the call might be X.pretty_print(...)

[ntjohnson1]

I have a PR up for printing the frontal slices #354. IIRC matlab has a global setting to control the printing precision so my bias would be to use something similar here as well. I haven't tried if our implementation is hooked up correctly where changing the numpy printing format populates correctly for us.

[tgkolda]

Thanks, Nick. I would still like this functionality as an option. The numpy-style printing is painful.

[ntjohnson1]

The primary painful piece after having control of the print format, and the frontal slices is just the brackets and a custom name display correct? Or is there some additional functionality that I'm overlooking?

[tgkolda]

So, the pain points are:

• Displaying the name
• Controlling the print format
• Displaying the size as 5 x 4 x 3 rather than (5,4,3)
• For 1D tensors, displaying the size as 5 rather than (5,).

FWIW, I don't think we necessarily want to change the universal print formatting. I can imagine scenarios where you just want different formats for different purposes in the same file.

Here's an updated version that handles 1D and 2D tensors more intelligently and changes the default name to Values.

def pretty_print_tensor(X, fmt="10.4f", name="Values"):
    if not isinstance(X, ttb.tensor):
        raise ValueError("Input must be a pyttb tensor")

    # Get the shape of the tensor
    shape = X.shape

    if name == "Values":
        print("Tensor is of shape "+ " x ".join(map(str, shape)))
    else:
        print(f"{name} is a tensor of shape " + " x ".join(map(str, shape)))

    # Check if the tensor is 1D
    if len(shape) == 1:
        print(f"{name}(:) =")
        for val in X.data:
            print(f"{val:{fmt}}")
        return

    # Check if the tensor is 2D
    if len(shape) == 2:
        print(f"{name}(:,:) =")
        array = X.data
        for row in array:
            print(" ".join(f"{val:{fmt}}" for val in row))
        return

    # Iterate over all possible slices (in Fortran order)
    for index in np.ndindex(shape[2:][::-1]): # Skip the first two dimensions and reverse the order
        index = index[::-1] # Reverse the order back to the original
        # Construct the slice indices
        slice_indices = (slice(None), slice(None)) + index
        slice_data = X[slice_indices]
        print(f"{name}(:, :, {', '.join(map(str, index))}) =")
        array = slice_data.data
        for row in array:
            print(" ".join(f"{val:{fmt}}" for val in row))        

[ntjohnson1]

I believe all of those pain points are decisions aligned with numpy conventions. Since that's the defacto scientific library for python it feels like deviating from that would make it harder for a python user to pick things up.

I believe pyttb is trying to capture the functionality of the MATLAB tensor toolbox but the aesthetics are not an objective. Here it feels like we are deviating from Python/Numpy standard approaches to better match MATLAB aesthetics.

I'll defer to @dmdunla since he is the one allowed to approve incoming PRs

[dmdunla]

For MATLAB users accustomed to the TTB for MATLAB, Tammy's suggestion makes sense. However, as Nick points out, this moves away from Python/NumPy conventions and thus would be distracting to people accustomed to pythonic standards.

Here are my suggestions:

1. Align Tammy's code to TTB for MATLAB as much as is possible (using the variable name in the output has already been determined as impossible at this point).
2. Name the methods that produce the TTB for MATLAB-like output as matlab_print. I believe this is both more descriptive and more objective than pretty_print, and including matlab inthe name of the function describes the intent (e.g., print the tensor like it's printed in MATLAB).
3. For the default printing of data classes, continue with our attempts to be faithful to TTB for MATLAB while incorporating numpy-like conventions (e.g., tensor of shape (2, 2, 2) vs. tensor of shape 2 x 2 x 2).
4. For the default printing, use the names of the data class data members when printing. For example, tensor has data, ktensor has weights and factor_matrices, and ttensor has core and factor_matrices. I'm not a fan of using the name U for factor matrices in general, and I'd like to avoid introduction of aliases to underlying data members in the __str__ and __repr__ methods.
5. Use frontal slices for now in all data classes. NumPy does not output based on layout/order at this point, so aligning with TTB for MATLAB, which outputs frontal slices, makes the most sense at this point. However, if we eventualy support other layout/order options (e.g., "C"), we should revisit this choice to make sure that frontal slices make sens for all layouts.

I do not think there is a single solution that will both aligh with both MATLAB and Python conventions, so the above are an attempt at some compromise.

This is my take ATM, and I'm open to more discussion.

[ntjohnson1]

SG. If we are accommodating this then I think it makes sense to create pyttb.matlab_support under which we can put matlab_print as a function. Each data class can have a private method _matlab_print this publicly facing function calls. That way we have some separation for the weird/non-standard/custom things required. If the support doesn't continue to grow then it's not a crazy amount of indirection.

[dmdunla]

Great idea. I like the separation of matlab_support into a distinct space in pyttb.