`contract_compressed` requires either `max_bond` or `cutoff` to be specified

[snsunx]

What happened?

I found that the TensorNetwork.contract_compressed method works on a TensorNetwork with two Tensors but does not work on a TensorNetwork with more than two Tensors.

What did you expect to happen?

I expect the method to return a Tensor object.

Minimal Complete Verifiable Example

import numpy as np
import quimb.tensor as qtn   

A = qtn.Tensor(np.random.rand(3, 4), inds=['a', 'b'], tags=['A'])
B = qtn.Tensor(np.random.rand(4, 5), inds=['b', 'c'], tags=['B'])
C = qtn.Tensor(np.random.rand(5, 4), inds=['c', 'd'], tags=['C'])
tn = A & B & C
tn.contract_compressed('greedy')

Relevant log output

Traceback (most recent call last):
  File "/central/home/ssun3/tests/test_tensor_networks.py", line 53, in <module>
    test_mps()
  File "/central/home/ssun3/tests/test_tensor_networks.py", line 49, in test_mps
    mps.contract_compressed('greedy')
  File "/home/ssun3/miniconda3/envs/tns/lib/python3.10/site-packages/quimb/tensor/tensor_core.py", line 6476, in contract_compressed
    t = tn._contract_compressed_tid_sequence(
  File "/home/ssun3/miniconda3/envs/tns/lib/python3.10/site-packages/quimb/tensor/tensor_core.py", line 6290, in _contract_compressed_tid_sequence
    _compress_neighbors(tid2, t2, d)
  File "/home/ssun3/miniconda3/envs/tns/lib/python3.10/site-packages/quimb/tensor/tensor_core.py", line 6236, in _compress_neighbors
    if bonds_size(t, t_neighb) > chi:
TypeError: '>' not supported between instances of 'int' and 'NoneType'

Anything else we need to know?

The chi variable is None, which resulted in the error.

Environment

I'm using Quimb version 1.4.0.

[snsunx]

If I set max_bond when calling the function, the error goes away. Is this the expected behavior? Is it possible that the user calls contract_compressed without setting max_bond, i.e. without intending to compress the tensor network?

[jcmgray]

Hi @snsunx, yes currently one needs to supply either max_bond or cutoff. Probably it would be useful for debugging purposes to support neither being supplied.

Cleaning up the interface and writing some docs is on my to-do list, but it is a tricky task, as this is quite an advanced function, where there is a trade-off between 'just working' with no settings, and explicitly forcing the user to engage in understanding the algorithm at some level and choosing the settings, since there is a large affect in terms of both complexity and accuracy.

I suppose I'm saying the interface is not quite finalized yet!

[snsunx]

That answers my question. Thank you for the clarification!