RecursionError

[DJHeisenberg01]

Hi Lambeq support, i'm sorry but i am here again,

I am trying to train a binary classification model using my own dataset, i am using the quantum pipeline example in lambeq documentation, i tried to run pipeline with 350 train sentences and 350 val sentences, i don't know the timing of quantum training but in 6 hours it doesn't return any epoch results. I tried to run it on Colab and the trainer.fit() returns this error:

RecursionError                            Traceback (most recent call last)
<ipython-input-15-9407b272b2c0> in <cell line: 1>()
----> 1 trainer.fit(train_dataset, val_dataset, evaluation_step=1, logging_step=100)

11 frames
... last 2 frames repeated, from the frame below ...

/usr/local/lib/python3.10/dist-packages/discopy/monoidal.py in then(self, *others)
    413     def then(self, *others):
    414         if len(others) != 1 or any(isinstance(other, Sum) for other in others):
--> 415             return super().then(*others)
    416         other, = others
    417         return self.upgrade(

RecursionError: maximum recursion depth exceeded

I don't know if I misconfigured the pipeline, but I didn't change the various configurations in the example.
Thanks again for help.

[dimkart]

Hi again, note that the logging step counts epochs; so, could you use logging_step=1 so we could get an idea of the actual time an epoch needs to run? It's true that quantum models need significantly longer times than classical ones, but until we have this number it's difficult to say if your code does something or not.

[DJHeisenberg01]

I also tried running it on VS Code with 10 circuits, but it seems to loop even though there are few circuits

[dimkart]

@DJHeisenberg01 It seems your problem is related with this issue in DisCoPy. We are currently investigating it with the DisCoPy team, we'll let you know as soon as possible.

[dimkart]

A fix has been added in DisCoPy repo (legacy branch), and a new 0.5.1.x version will be uploaded to PyPi soon. (Note that lambeq uses an older version of DisCoPy atm). For the moment, you can upgrade your environment using the following command:

pip install --upgrade git+https://github.com/discopy/discopy.git@legacy

When the PyPi release is out I'll also let you know. Also note that within the next two weeks we'll release a new version of lambeq which will support DisCoPy 1.0, plus will fix a bunch of other issues, so keep an eye for upgrades. Thanks again for letting us know.

[DJHeisenberg01]

Thanks a lot for the help, i'll try right now, i am doing a sperimental thesis on Lambeq and i'm testing pipelines with a big dataset (3500 sentences).

[dimkart]

DisCoPy PyPi version you can use as well: 0.6.0.

Good luck with your experiments, let us know how it goes.

[Thommy257]

Can you try to run the same data through our quantum_pipeline_jax.ipynb please?

[DJHeisenberg01]

I run the same data in quantum_pipeline_jax.ipynb, it works if i use stairs_reader, but if i use more sentences (i need it for a Thesis) like 1700 val and 1700 train it crashes because it uses all the Colab Pro RAM