Have you try (using your terminology) B50 50 steps?

It's quite challenging and models follow there a different regimes, I'd be interested to know your results.

Hello, Which subset of classes of the full ImageNet-1000 did you use for ImageNet-100? Is it the first 100 classes in alphabetical order of the folder names? Thanks!

Thank you for this nice work!

I am trying to reproduce some comparison methods using the code and configurations from arthurdouillard/incremental_learning.pytorch. But the results in b0 setting are very different from what are reported in your paper. I guess the main reason is the configuration for some hyperparameters. I believe you have carefully tuned these parameters in b0 setting.

Since you said that your code is also based on arthurdouillard/incremental_learning.pytorch, I am wondering if it is possible for you to release some configuration files first so that we reproduce some results using your setting and arthurdouillard's code?

Thank you so much!

After reading your paper, I have some questions 1.What is the accuracy of the model without the warm-up strategy? 2.What is the data enhancement used by the model? 3.What is the number of images stored?

I've been trying to reproduce your results of WA using my codebase (original inclearn) unsuccessfully.

And when I'm looking at your codebase I see that WA is not yet implemented? The distillation https://github.com/Rhyssiyan/DER-ClassIL.pytorch/blob/main/inclearn/models/incmodel.py#L73 is never used. Am I missing something or did you not yet uploaded the code & configs of your baselines?

Hey,

I see that your paper states that on CIFAR100-B0 you averaged three runs (cf caption of table 1). However in datasets.py, the dataset is defined with three runs (two news runs, and the three from PODNet).

Can I assume you only used the three from PODNet?

Thanks!

I want to compare based on the data you reported (reproduction methods and your method). Can you provide the accurate data of the graph in the paper and the appendix? Thank you!

Another question is the data in the figure is the average of the three orders given by PODNet?

In appendix D, you write that your method process 1.07ms/image vs an usual ResNet at 0.99ms/image.

With how many pruned resnet (and thus how many tasks) is the 1.07ms/image produced? I guess that DER after 50 tasks should be slower (at least a little) than DER after 2 tasks. (do you have numbers about speed w.r.t number of tasks?)

Thanks!

Hello,

Can you provide the final accuracy (not the avg) for all models (or at least for DER) on all settings of CIFAR and ImageNet please? Otherwise I'm relying on the graphs but it's not 100% accurate.

Thank you very much!

Why in Table 1, the capacity of 5 steps OURS (W/O P) model is 33.6 instead of 11.2x5=56? But in Table 2, the capacity of Ours (W/O P) model of 2 steps is 11.2x2=22.4?

Does the baseline use the same class order for all?

I ran into endless errors while run code sh scripts/run.sh, such as easydict, tensorboardx missing, and sacred version conflict with python 3.10, etc.

Could you please provide your requirements.txt or environment.yml for environment dependencies, even pip list result.

Thanks for your help.

Sorry to bother you, I wasn't able to reproduce the ImageNet-100 results. I'm using the config imagenet-100/b0_10s/configs/ without any change.

For the first task, I'm getting only 90.4% top1 accuracy vs. ~98% shown in Figure 6, and the final top1 accuracy 71.24% vs. 77.18% in Table 3.

Am I missing something?

I found in ur table 7 or original icarl paper. 10 task. Cifar100-B0. 32-layer ResNet.

I cannot achieve the results about 64.04.

Or give some insights?

Hello,

Followings our emails, can you provide at some point:

• the pruning implementation (which is core to your method)
• the final parameters count (in all settings and datasets), and not the average over all steps
• same for the latency, or at least the overhead in % induced after each task

Thank you very much!

Hello, I modified the ImageNet-100 config file to ImageNet-1000 following the supplementary implementation details and didn't get the results in the paper. Was there anything in addition that I need to do? Can you share the ImageNet-1000 configuration?

Thanks so much!