Feature Request : GPU-based dedicated environment to run the Lambeq models

[ACE07-Sev]

Motivation:
Given none of the currently existing platforms support the Lambeq packages, it would nice if there was a dedicated environment with GPU-enabled compiler to allow the user(s) to train models using TKET and have WSL interpreter for running WebParser. This environment would have the imports pre-installed by default and allow for .txt or .csv file imports for the dataset.

Downloadable Checkpoints:
This manner user(s) can train the models based on preference with high speed, and eliminate any variability given different IDEs and/or interpreters which may potentially disrupt the model. This would allow users to outsource the training process to a GPU-based environment and simply download the model.lt checkpoint to use locally. Given the necessity of a WSL interpreter for running WebParser, the user could also download .pkl format of Webparsed diagrams to be used on locally as an additional feature as well.

Error Suggestion :
There could also be a suggestion system based on errors, like TypeError : Ty('p') which essentially indicates the need for an AtomicType.PREPOSITIONAL_PHRASE, or the trainer operand issue which indicates an issue with some diagrams having different number of output wires and provide suggestions for changing the dataset accordingly.

Proposed Feature Summary:
Given how each online environment requires users to pip install the libraries and currently this is the issue with platforms such as Colab and Deepnote, the environment would have the Lambeq and DiscoPY packages pre-installed, and run on Linux OS with WSL being the default interpreter. Since TKET is currently very slow for intermediate scale models, GPU feature would be incredibly useful. And lastly having a function to generate downloadable files for the model.lt and Webparser diagrams.

[dimkart]

Thanks for your suggestions. Regarding having a "dedicated environment with GPU-enabled compiler", I'm not sure what do you mean. A python virtual environment? A docker image for an OS with everything pre-installed? And by "GPU-enabled" you mean wrt Pytorch, jax, both? Note that soon we are also adding PennyLane and Quantum TF support, so as you can see there are so many possibilities, so creating pre-made environments for everything is not that simple.

Regarding the downloadable checkpoints, I admit it's not clear to me what you are suggesting. Could you provide more information?

Regarding the "error suggestion" point, do you have any specific examples in mind that would help us understand better the issue?

[ACE07-Sev]

Greetings,

Thank you for considering the suggestion! So I'll go one by one.

For "dedicated environment with GPU-enabled compiler" I essentially mean an environment where you can simply run the lambeq codes and have the model run on a GPU for speed as some models like TketModel take a really long time (1hr approx per epoch) and would be great to have an environment where we can just run the code faster. Secondly this environment would essentially have all of the imports pre-installed (as Lambeq cannot be run on Google Colab nor Deepnote, so having an environment like Jupyter with the package installed would be great, and also some segments like the BobCatParser was not available to be imported and I essentially had to manually download the ghz.tar file and then use a file path for the model to use it).

So up to now we would have an environment which we can just run the code without installing any of the package components and run on a GPU for speed. So lastly is the downloadable checkpoints. So since when we run the model it generates a checkpoint for the trained model, we can use the environment to do the heavy work of running and training the model and we can just download the checkpoint generated and use it on a local device in for instance Pycharm.

Lastly some errors were quite easy to fix and had singular solutions so the environment can give a string output hinting how to resolve it. For instance TypeError : Ty('p') which essentially indicates the need for an AtomicType.PREPOSITIONAL_PHRASE, or the trainer operand issue which indicates an issue with some diagrams having different number of output wires and provide suggestions for changing the dataset accordingly. So some of these errors simply needed a hint to fix so we can have a suggestion module to give hints on certain errors.

[ACE07-Sev]

If any parts are unclear kindly inform and I'll go into more detail.

Thank you so much!

[Thommy257]

Hi and thanks for the clarifications. Regarding the dedicated GPU environment: Keep in mind that the TketModel is built on top of tket, which is a compiler for quantum computers. tket provides an interface with many toolboxes that allow computations on real quantum devices as well as quantum simulators. Therefore, there's no GPU speedup available per se. Some simulators support GPU computing, however the main purpose of the TketModel is not speed but the possibility to run your model on real quantum hardware (or at least to simulate the exact pipeline of a quantum experiment).

In contrast, we provide the NumpyModel to enable quick debugging and GPU support using jax. The NumpyModel is essentially the same model as the TketModel, however, performs an exact simulation of the quantum circuits using numpy. By leveraging the just-in-time (jit) compilation feature of jax and it's GPU support, it's an order of magnitude faster than the TketModel. Therefore, we encourage people to use the NumpyModel to develop and debug their models and then validate the results using the TketModel.

Regarding the downloadable checkpoints. The checkpoints are in theory downloadable. They are saved automatically in the projects directory and can be downloaded from any machine you are running your experiments on. We can't provide cloud computing resources unfortunately. The reason why lambeq doesn't work with Google Colaboratory is that it runs on python3.7, lambeq however requires python>=3.8.

Note that the SSL error with BobcatParser and WebParser was fixed in the release 0.2.4.

You are right, some error codes need improvement. We'll put that on our agenda.

[Thommy257]

Hi! you can simply do ```python model = PytorchModel.from_diagrams(diagrams) model.initialise_weights() model.to(torch.device(f'cuda:{device_id}')) ``` or specify the device id in the trainer, using the `device` argument.

[mspronesti]

Thanks a lot, I eventually figured it out and that's why I deleted the comment. Thanks again 😄

[Thommy257]

I see, the original question was how to use the PytorchModel with GPU support.