How to save the model of each stage

[cc19970821]

hello, i want to save the model to do the T-SNE, how to save the model of each stage?
i found you had set the ckpt but don't set the torch.save() to save the model.

[yaoyao-liu]

Hi,

I removed the command for the checkpoints to save space.
If you need to save the checkpoints, you may add the following commands:

torch.save(b1_model, ckp_name)
torch.save(b2_model, ckp_name_b2)  

after these lines:

class-incremental-learning/adaptive-aggregation-networks/trainer/trainer.py

Lines 165 to 166 in 28f0bf6

	else:
	raise ValueError('Please set the correct baseline.')

[ZHUANGHP]

@yaoyao-liu After saving the models at each stage, how could we measure the accuracy for such stage by loading the models from the checkpoints. For instance in training on cifar100 for 5 phases, could we load "iter_9_b1.pth" and "iter_9_b2.pth" (e.g., using --resume) without training to measure the accuracy of the last phase? I have tested the code and it does not seem to be trivial. If I directly load "iter_9_b1.pth", "iter_9_b2.pth" and test them using

b1_model.eval()
b2_model.eval()
test_loss = 0
correct = 0
total = 0
with torch.no_grad():
    for batch_idx, (inputs, targets) in enumerate(testloader):
        inputs, targets = inputs.to(device), targets.to(device)
        outputs, _ = process_inputs_fp(the_args, fusion_vars, b1_model, b2_model, inputs)
        loss = nn.CrossEntropyLoss(weight_per_class)(outputs, targets)
        test_loss += loss.item()
        _, predicted = outputs.max(1)
        total += targets.size(0)
        correct += predicted.eq(targets).sum().item()
print('Test set: {} test loss: {:.4f} accuracy: {:.4f}'.format(len(testloader), test_loss / (batch_idx + 1),
                                                               100. * correct / total))

then the accuracy would be roung 2%. Is this because fusion_vars should somehow be updated during the training?

[yaoyao-liu]

Hi @ZHUANGHP,

Thanks for your interest in our work.

If you need to run the test without training, we can create the test code following this:
https://github.com/yaoyao-liu/class-incremental-learning/tree/main/mnemonics-training/2_eval

If you have further questions, please feel free to email me. Have a nice day!

Best,
Yaoyao