An Error while running Classical Pipeline Example given in docs/examples

[srinjoyganguly]

Hi @dimkart I hope you are doing well

I am trying to run the code given here on my Google Colab account - https://github.com/CQCL/lambeq/blob/main/docs/examples/classical_pipeline.ipynb

I am installing lambeq directly on Colab and it is picking up the latest version of DisCoPy

But I am continuously getting an error like this. I have pasted the full stack trace here -

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-11-84634b74856a> in <module>()
     39 dev_cost_fn, dev_costs, dev_accs = make_cost_fn(dev_pred_fn, dev_labels)
     40 
---> 41 result = train(train_cost_fn, x0, niter=20, callback=dev_cost_fn, optimizer_fn=torch.optim.AdamW, lr=0.1)

10 frames
<ipython-input-11-84634b74856a> in train(func, x0, niter, callback, optimizer_fn, lr)
      3     optimizer = optimizer_fn(x, lr=lr)
      4     for _ in range(niter):
----> 5         loss = func(x)
      6 
      7         optimizer.zero_grad()

<ipython-input-11-84634b74856a> in cost_fn(params, **kwargs)
     16 def make_cost_fn(pred_fn, labels):
     17     def cost_fn(params, **kwargs):
---> 18         predictions = pred_fn(params)
     19 
     20         logits = predictions[:, 1] - predictions[:, 0]

<ipython-input-10-dbb8534e3157> in predict(params)
      1 def make_pred_fn(circuits):
      2     def predict(params):
----> 3         return torch.stack([c.lambdify(*parameters)(*params).eval(contractor=tn.contractors.auto).array for c in circuits])
      4     return predict
      5 

<ipython-input-10-dbb8534e3157> in <listcomp>(.0)
      1 def make_pred_fn(circuits):
      2     def predict(params):
----> 3         return torch.stack([c.lambdify(*parameters)(*params).eval(contractor=tn.contractors.auto).array for c in circuits])
      4     return predict
      5 

/usr/local/lib/python3.7/dist-packages/discopy/tensor.py in eval(self, contractor)
    448         if contractor is None:
    449             return Functor(ob=lambda x: x, ar=lambda f: f.array)(self)
--> 450         array = contractor(*self.to_tn()).tensor
    451         return Tensor(self.dom, self.cod, array)
    452 

/usr/local/lib/python3.7/dist-packages/tensornetwork/contractors/opt_einsum_paths/path_contractors.py in auto(nodes, output_edge_order, memory_limit, ignore_edge_order)
    262         output_edge_order=output_edge_order,
    263         nbranch=1,
--> 264         ignore_edge_order=ignore_edge_order)
    265   return greedy(nodes, output_edge_order, memory_limit, ignore_edge_order)
    266 

/usr/local/lib/python3.7/dist-packages/tensornetwork/contractors/opt_einsum_paths/path_contractors.py in branch(nodes, output_edge_order, memory_limit, nbranch, ignore_edge_order)
    160   alg = functools.partial(
    161       opt_einsum.paths.branch, memory_limit=memory_limit, nbranch=nbranch)
--> 162   return base(nodes, alg, output_edge_order, ignore_edge_order)
    163 
    164 

/usr/local/lib/python3.7/dist-packages/tensornetwork/contractors/opt_einsum_paths/path_contractors.py in base(nodes, algorithm, output_edge_order, ignore_edge_order)
     86   path, nodes = utils.get_path(nodes_set, algorithm)
     87   for a, b in path:
---> 88     new_node = contract_between(nodes[a], nodes[b], allow_outer_product=True)
     89     nodes.append(new_node)
     90     nodes = utils.multi_remove(nodes, [a, b])

/usr/local/lib/python3.7/dist-packages/tensornetwork/network_components.py in contract_between(node1, node2, name, allow_outer_product, output_edge_order, axis_names)
   2083     axes1 = [axes1[i] for i in ind_sort]
   2084     axes2 = [axes2[i] for i in ind_sort]
-> 2085     new_tensor = backend.tensordot(node1.tensor, node2.tensor, [axes1, axes2])
   2086     new_node = Node(tensor=new_tensor, name=name, backend=backend)
   2087     # node1 and node2 get new edges in _remove_edges

/usr/local/lib/python3.7/dist-packages/tensornetwork/backends/pytorch/pytorch_backend.py in tensordot(self, a, b, axes)
     44   def tensordot(self, a: Tensor, b: Tensor,
     45                 axes: Union[int, Sequence[Sequence[int]]]) -> Tensor:
---> 46     return torchlib.tensordot(a, b, dims=axes)
     47 
     48   def reshape(self, tensor: Tensor, shape: Tensor) -> Tensor:

/usr/local/lib/python3.7/dist-packages/torch/functional.py in tensordot(a, b, dims, out)
   1032 
   1033     if out is None:
-> 1034         return _VF.tensordot(a, b, dims_a, dims_b)  # type: ignore[attr-defined]
   1035     else:
   1036         return _VF.tensordot(a, b, dims_a, dims_b, out=out)  # type: ignore[attr-defined]

RuntimeError: expected scalar type Float but found Double

I was able to successfully carry out experiments using the Quantum Pipeline code on Google Colab and did not faced any issues but for this one I am getting error. I have tried to fix the issue by converting variables or some function outputs to float() but I was unable to rectify this.

Can you please help me fix this issue?

Thank you so much!

[ianyfan]

Try setting torch.set_default_tensor_type(torch.FloatTensor) or torch.set_default_tensor_type(torch.cuda.FloatTensor) at the start of your script?

[dimkart]

Hello @srinjoyganguly and thanks for reporting this. @ianyfan has already responded what we think might be the problem, let us know if this works for you.

[srinjoyganguly]

Hi @ianyfan thank you so much for the help! I tried using the lines you gave at the start of my script but still getting the same issue. The same error persists. I realized that this line - torch.set_default_tensor_type(torch.cuda.FloatTensor) already existed in my code and I was getting the error. I tried with - torch.set_default_tensor_type(torch.FloatTensor) but the error is still the same. Are there any other ways?

@dimkart you are very welcome! Thank you a lot!

[dimkart]

Hi @srinjoyganguly, can you send us your notebook/code in lambeq-support@cambridgequantum.com (or post it here), so we can have a better look? It's really difficult to say what is going on just from the error message.

[ianyfan]

Changing the line to

torch.set_default_tensor_type(torch.cuda.DoubleTensor)

should allow the notebook to run, though we are working on a proper fix on the underlying code.

[srinjoyganguly]

Indeed @ianyfan the code works with DoubleTensor! Thank you so much for the help! I look forward to the proper fix in the underlying code! Thanks a lot @ianyfan and @dimkart !

[dimkart]

Thanks for pointing this out @srinjoyganguly. We have updated the notebook to use a double tensor until we fix this in DisCoPy. This issue will be now closed.

[srinjoyganguly]

You are very welcome @dimkart ! I am glad, thanks so much!