g-u-n / eccv22-foster Goto Github PK

Mentioned in the paper:Through this scaling strategy, the absolute value of logits for old categories is
reduced, and the absolute value of logits for new ones is enlarged, thus forcing
the model Ft to produce larger logits for old categories and smaller logits for
new categories. Why can a reduction in the absolute value of logits of the old class force the model Ft to produce a larger logits of the old class❓

Hi,

thank you for the code and awesome paper !!

i would like to train using B0 Cifar-100 setting

amending init_cls from 10 to 0 in cifar100.json file isn't work

could you tell me how to train it??

I'd greatly appreciate any help you can offer.

Hi,

thank you for the repo and paper. It is a really great work!

Unfortunately, I tried to run your repo but I received a NameError: name 'cdist' is not defined in ECCV22-FOSTER/models/base.py", line 132, in _eval_nme.

I simply fixed the error adding the following import at the beginning the file models/base.py

from scipy.spatial.distance import cdist

I opened this issue to improve the repo. Probably you simply missed it 😉

Best,
Niccolò

Thank you for such a good work. May I ask if I want to run some experiments on imagenet100, is the path of the imagenet1K dataset entered here in the data.py file?

Thank you for such a good work, I find that possible problem about Logits Alignment. The following is a detailed description.

Equation (13) and (14) from the paper:

Implemented code from foster.py：


According to the code, I calculated beta1 and beta2 as follows:

I guess that this result is not equivalent to what is proposed in the paper.

I'm looking forward to your answer!

what is ”temperature scalar T to 2“ ？

thank you ！

Hi!
First of all, thank you very much for your repo. I am very interested in your paper.
I find the darknet eval is the top1 acc of the snet. However, its precision is not the same as that in the last list. Later, I found that the model used the original teacher network with two backbone networks when evaluating.
Because you update the self._network is in the begin of incremental_train().
This is the original plan to use the result of the teacher model, or is it a mistake.

想和您做对比实验，下面这两个有什么区别啊，哪个对应论文里的算法啊，谢谢
`Train CIFAR-100
python main.py --config=foster-cifar100.json

Train FOSTER-RMM
python main.py --config=foster-rmm.json`

Hello, thanks for your sharing. I ran the code and found that the result of "foster" on "cifar100" is worse than that in the paper.
I ran the B0-10steps, the top1 result using CNN is :[93.5, 80.6, 76.97, 71.8, 69.84, 67.5, 65.63, 62.32, 61.04, 59.47] NME:[92.7, 78.5, 73.67, 65.65, 62.32, 58.38, 56.76, 53.09, 51.79, 49.15]. Then I average the CNN result. The outcome is 70.867 which is lower than paper's result (72.90).
my json file is as follows. I would appreciate it if you could help me.
{
"prefix": "init10_steps10",
"dataset": "cifar100",
"memory_size": 2000,
"memory_per_class": 20,
"fixed_memory": true,
"shuffle": true,
"init_cls": 10,
"increment": 10,
"model_name": "foster",
"convnet_type": "resnet32",
"device": ["0"],
"seed": [1993],
"beta1":0.94,
"beta2":0.97,
"oofc":"ft",
"is_teacher_wa":false,
"is_student_wa":false,
"lambda_okd":1,
"wa_value":1,
"init_epochs": 200,
"init_lr" : 0.1,
"init_weight_decay" : 5e-4,
"boosting_epochs" : 170,
"compression_epochs" : 130,
"lr" : 0.1,
"batch_size" : 128,
"weight_decay" : 5e-4,
"num_workers" : 8,
"T" : 2
}

Hi!

Thank you for the repo and the paper,

I have run the origin FOSTER repo and get final top-1 CNN accuracy at about 60 in CIFAR100 while it's about 70 in the paper.

I was wondering about which part is wrong, could you tell me where the problem is?

It would be greatly appreciated for any help from you.

Here is the training log.
cil_1993_foster_resnet32_cifar100_10_10_0.94_0.97_False_False.log

Hello, I would like to use you as a comparative experiment. Could you please kindly improve the log.txt file of imagenet B50 10 steps? I would be very grateful if you could give me your help.

Is "Bound" joint-training 100 classes？Why Bound has different results ?

thank you！

Hi. Thank you for your great work!

While I am trying to reproduce your work for further research, I found following imagenet-100 configs yield inferior results for imagenet-1000.

Can you provide the config file for imagenet-1k training?

Best wishes!

Hi @G-U-N ,

Thanks for your work on FOSTER.

I have a question about the evaluation protocol in the paper. In table 1, B50-10steps, the DER is with 66.36. But in their paper, they claim to have a 72.45 results in table 2 (CIFAR100-B50, 10 steps). Is the margin caused by the difference in the evaluation protocol?

missing the config folder

Hello,

Can you please provide the config file of training ImageNet1k?

Thanks!

作者您好，打扰您了。
请问能否方便提供一下performance.png这个图中的每个方法每个阶段的具体数据，也就是您论文中的图3.谢谢！