Issue with joint training on CIFAR100 with Resnet18.

[prateeky2806]

[Moving from Slack to Github, @ggraffieti]

Hi, I am trying to have joint training ResNet18 model on CIFAR100. I am getting accuracies which seem to be random. My working code is attached below. I feel like I am doing something wrong so any help would be appreciated.

import numpy as np
import avalanche as avl
from torch.nn import CrossEntropyLoss
from torch.optim import SGD
from avalanche.evaluation import metrics as metrics
import torchvision

def main():
    
    benchmark = avl.benchmarks.SplitCIFAR100(n_experiences=10, seed=1, return_task_id=True)

    model = torchvision.models.resnet18(num_classes=100)

    criterion = CrossEntropyLoss()
    optimizer = SGD(model.parameters(), lr=0.0001)

    interactive_logger = avl.logging.InteractiveLogger()
    evaluation_plugin = avl.training.plugins.EvaluationPlugin(
        metrics.accuracy_metrics(epoch=True, experience=True, stream=True),
        metrics.loss_metrics(epoch=True, experience=True, stream=True),
        loggers=[interactive_logger])
    
    cl_strategy = avl.training.JointTraining(
        model, optimizer, criterion,
        train_mb_size=128, train_epochs=2, eval_mb_size=128,
        evaluator=evaluation_plugin)
    
    cl_strategy.train(benchmark.train_stream)
    res = cl_strategy.eval(benchmark.test_stream)


if __name__ == '__main__':
    main()

[AntonioCarta]

You should increase your learning rate and number of epochs.