[ml_tools] A train config used more than once overwrites previous results.

[debrevitatevitae]

Short description

If we train a model more than once in a row and we want to use the same training settings (i.e. same TrainConfig), every training will overwrite the results of the previous ones.

What is the expected result?

When the option create_subfolder_per_run=True and training a second time with the same TrainConfig, a new subdirectory should be created, where the results of the second training are stored. In other words, the results of the second training should not overwrite the results of the first one.

What is the actual result?

In the situation above, the results of the first training are overwritten.

Steps/Code to reproduce

from itertools import count

import torch
from qadence import (
    feature_map,
    hamiltonian_factory,
    hea,
    QNN,
    QuantumCircuit,
    TrainConfig,
    Z,
)
from qadence.ml_tools import to_dataloader, train_with_grad
from torch.utils.data import DataLoader


def dataloader(data_size: int = 25, batch_size: int = 5, infinite: bool = True) -> DataLoader:
    x = torch.linspace(0, 1, data_size).reshape(-1, 1)
    y = torch.sin(x)
    return to_dataloader(x, y, batch_size=batch_size, infinite=infinite)


n_qubits = 2
fm = feature_map(n_qubits)
ansatz = hea(n_qubits=n_qubits, depth=2)
observable = hamiltonian_factory(n_qubits, detuning=Z)
circuit = QuantumCircuit(n_qubits, fm, ansatz)

model = QNN(circuit, observable, backend="pyqtorch", diff_mode="ad")
batch_size = 100
input_values = {"phi": torch.rand(batch_size, requires_grad=True)}

cnt = count()
criterion = torch.nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.1)


def loss_fn(model: torch.nn.Module, data: torch.Tensor) -> tuple[torch.Tensor, dict]:
    next(cnt)
    x, y = data[0], data[1]
    out = model(x)
    loss = criterion(out, y)
    return loss, {}


config = TrainConfig(max_iter=10, print_every=2, folder="dev", create_subfolder_per_run=True)

# train a first time
train_with_grad(model, dataloader(), optimizer, config, loss_fn)

# train a second time
train_with_grad(model, dataloader(), optimizer, config, loss_fn)

Tracebacks (optional)

No response

Environment details (optional)

No response

Would you like to work on this issue?

None