kuc2477 / pytorch-ewc Goto Github PK

Great implementations, many thanks.

I'm new to the continuous learning and a little bit confused about its problem setting. I see that you generated different permuted MNIST data for every task (or session), but it looks labels of all 8 tasks are shared (0~10). My understanding is that continuous learning is about learning new task without forgetting old tasks. Should we give new labels (e.g., 11) as a new task? I appreciate your help.

For example, check Figure 2(e) of this paper: https://arxiv.org/pdf/1606.09282.pdf

I am trying to run EWC on my dataset with resnet50 model. While updating the fisher matrix using your function, My code says Cuda out of memory due to "log_liklihoods.append(output[:, target])" in the code. I read this "https://stackoverflow.com/questions/59805901/unable-to-allocate-gpu-memory-when-there-is-enough-of-cached-memory" and figured out the problem using 'detach()'. After doing detach etc, I get an error: RuntimeError: One of the differentiated Tensors appears to not have been used in the graph. Set allow_unused=True if this is the desired behavior.
To further solve this, I set "allow_unused=True" in autograd. As a result, all my gradients go to 0. Why is this happening?

I found that performing forward-backward in a loop is much faster than using autograd.grad with retain_graph=True. Your current code is:

        loglikelihood_grads = zip(*[autograd.grad(
            l, self.parameters(),
            retain_graph=(i < len(loglikelihoods))
        ) for i, l in enumerate(loglikelihoods, 1)])

https://github.com/kuc2477/pytorch-ewc/blob/master/model.py#L75

after I change it to:

        for batch, label in zip(buffer_data, buffer_label):
            self.zero_grad()
            loglikelihood = F.cross_entropy(self(batch), label)
            loglikelihood.backward()
            for n, p in self.named_parameters():
                n = n.replace('.', '__')
                grads[n] = grads.get(n, 0) + p.grad ** 2

it runs much faster. Can you please investigate this?

Thank you.
Thanh Tung

As the code, we can get negative log-likelihood by F.log_softmax. How can I get the log-likelihood of regression problem??

The computation for fisher matrix should be the average of the gradient^2. But in your implementation. you did the average of the loss, then only use that averaged loss to calculate the gradient and estimate the fisher matrix.
Is this a bug or the math is equivalent.
check the tensorflow version: https://github.com/ariseff/overcoming-catastrophic

Code in line 80 of model.py :
fisher_diagonals = [(g ** 2).mean() for g in loglikelihood_grads]
should be:
fisher_diagonals = [(g ** 2).mean(dim=0) for g in loglikelihood_grads]

When we train the current task, will we use the data of the previous task？ ewc need task A data to compute fisher info, when we train task B how can we use ewc to constraint model's parameters change?