PyQit

Current Plan (sub to change any time >_<)

• ansatz -> QuantumPipeline wrapper that connects the ansatz and DL model (if any) -> Trainer (with/w/o torch)
• Data preprocessing techniques (no idea how to do that rn) - maybe using a lightning type of data module (thanks a lot lightning). But as lightning has a core dep of torch, I have to reinvent(?) it ig
• any ansatz could go with any DL backbone (ideally, not sure how much is feasible - tbd)
• torch is used with lightning as a imp soft-dep now!!

Will add better vignettes once i have my ideas consolidated in my mind

Have a look at a basic tutorial here (This is how the flow would look like in future, still working on making data module and Trainer etc more "user friendly" and expressive)

for some idea how it might look like:

# no torch
pyqit.set_backend("pennylane")
qml_model= QMLmodel(...) # may use their own ansatz?
dm = DataModule(...)
trainer = Trainer(...)
trainer.fit(qml_model, dm)
trainer.predict(qml_model, dm_new, return_format = "numpy") # or "torch" for torch tensors if torch is backend,
# should i add pennylane tensors as well? good question!

Using Pipeline

pyqit.set_backend("torch")
dm = DataModule(...)
model_a = QMLmodel(**params)
model_b = QMLmodel(**params) # or DLModel for that matter
trainer = Trainer( max_epochs=10, learning_rate=0.01)
pipeline = QuantumPipeline(
            [
                PipelineStage(model_a, name="stage_1", trainable=trainable_a),
                PipelineStage(model_b, name="stage_2", trainable=True),
            ],
            mode="sequential",
        )
pipeline.fit(datamodule=dm, trainers=trainer, fit_mode="sequential_greedy")
preds = pipeline.predict(X_new, batch_size=8)

You can also train just QMLmodel using Trainer here anyQMLmodel and DLmodel can be implemented by the user themselves or use the implemented ones from the package Then package would also have a complete model zoo.