Fix loss

Current loss only compute loss but it should also compute the output gradient of the last layer.

To change this I will make so that loss function take a graph and the ground truth labels as argument and return the loss.

changed milestone to %aidge - April 2024

assigned to @cmoineau

changed the description

Note: this may require to use compile_gradient in the loss function ... not a big fan but I don't have a better idea yet.

Also this remove the gradient from the backward signature.

mentioned in merge request aidge_core!118 (merged)

mentioned in merge request aidge_backend_cpu!60 (merged)

mentioned in merge request !5 (merged)

mentioned in merge request aidge_interop_torch!1 (merged)

I may have overdone things, after reflexion, maybe the best idea would have been to let MSE take a tensor as argument as we can set the grad of this tensor and not give the whole graph ...

I have made a refactor in this way, now learning looks like:

# Set object for learning
scheduler = aidge_core.SequentialScheduler(model)

# setup optimier
opt = aidge_learning.SGD(momentum=0.9)
learning_rates = aidge_learning.constant_lr(0.001)
opt.set_learning_rate_scheduler(learning_rates)
opt.set_parameters(list(aidge_core.producers(model)))

tot_acc = 0
for i, (input, label) in enumerate(aidge_dataprovider):
    input.init_gradient()
    scheduler.forward(data=[input])
    # Really long line should be a faster way ...
    pred = list(model.get_output_nodes())[0].get_operator().get_output(0)
    opt.reset_grad()
    loss = aidge_learning.loss.MSE(pred, label)
    acc = np.sum(np.argmax(pred, axis=1) == np.argmax(label, axis=1))
    tot_acc += acc
    scheduler.backward()
    opt.update()
    print(f"Nb samples {(i+1)*BATCH_SIZE}, loss: {loss[0]}, acc:{(acc/BATCH_SIZE)*100}%, tot_acc:{(tot_acc/((i+1)*BATCH_SIZE))*100}%")

All branches have been merged

@pineapple @olivierbichler Do not hesitate to take the time to post review and re-open the issue to point me modification to make if necessary.

closed