Using add_online for Centroid model

[OMGAmici]

Is the correct way to use the add_online() method as simple as using it in place of the add() method?

So far, I have not been able to see any performance improvement / change by introducing it, which leads me to suspect I might be implementing it incorrectly or there might be a bug.

I've tried using this as in the example snippet below as well as inserting it after the first measure of the accuracy. But neither has any effect on the outcome.

with torch.no_grad():
    for samples, labels in tqdm(train_ld, desc="Training"):
        samples = samples.to(device)
        labels = labels.to(device)
        samples_hv = encode(samples).to(device)
        model.add_online(samples_hv, labels)

accuracy = torchmetrics.Accuracy("multiclass", num_classes=num_classes).to(device)

[mikeheddes]

Hi, thank you for opening this issue. Yes, it should be as simple as that. Some things to note: 1) the improvement in accuracy will dependent on the dataset you are using, and 2) online learning is typically used in multiple iterations (epochs) to further boost the accuracy.

I modified the examples/voicehd.py to use add_online and got the following results:

add: 85.119% add_online 1 epoch: 86.273% add_online 2 epochs: 83.515% add_online 3 epochs: 87.877%

I modified examples/voicehd.py as follows:

for _ in range(3):

    with torch.no_grad():
        for samples, labels in tqdm(train_ld, desc="Training"):
            samples = samples.to(device)
            labels = labels.to(device)

            samples_hv = encode(samples)
            model.add_online(samples_hv, labels)

    accuracy = torchmetrics.Accuracy("multiclass", num_classes=num_classes)

    with torch.no_grad():
        for samples, labels in tqdm(test_ld, desc="Testing"):
            samples = samples.to(device)

            samples_hv = encode(samples)
            outputs = model(samples_hv, dot=False)
            accuracy.update(outputs.cpu(), labels)

    print(f"Testing accuracy of {(accuracy.compute().item() * 100):.3f}%")

[OMGAmici]

Thank you! That makes sense.