search

brjathu / iTAML

Official implementation of "iTAML : An Incremental Task-Agnostic Meta-learning Approach". CVPR 2020
96 stars 16 forks source link

Obviously catastrophic forgetting #18

Open ghost opened 3 years ago

ghost commented 3 years ago

Thanks for your work, i think your work is a milestone work in lifelong machine learning. But i run your code, the program does not seem to be running correctly. image And the log file is here: session_0 image session_1 image session_2 image session_3 image session_4 image

The only place I modify the program is start_sess = 0 So why is the program like this？

joeljosephjin commented 2 years ago

Im getting the same issue.. Is there a way to solve this?

ctm-hxds commented 2 years ago

Im getting the same issue.. Is there a way to solve this?

No, I tried a lot of methods and finally did not solve this problem

Im getting the same issue.. Is there a way to solve this?

Madghostek commented 4 months ago

The session log files contain accuracy of the base model, not the meta-model. There is meta_test method of learner object, which trains the meta-model using the base model and memory buffer, then the accuracy is as presented in the paper (the accuracy is printed to console after training meta tasks step). I'm not sure what's the point of the session files though, as they don't give information about the meta-model at all, at first it was confusing for me as well. Edit: I was running train_mnist.py, maybe on other datasets it's different