Nice work and the code. I find some equation errors in the main text.

1. The upper bound of I(X,T_j) should be E_{p(x)}D_{KL} not the E_{p(t_j)}D_{KL} in Equ.(7) and Equ.(19) in supplementary materials.
2. The D_{KL} in Equ.(11) misses a minus before 1/2. But your released code is correct.

Hi,
Great work! When I run the tools/train.py file with the config file I get this following error.
Also, where should I put the checkpoints after downloading them?
Thanks

Traceback (most recent call last):
File "tools/train.py", line 15, in
from mmcls.apis import set_random_seed, train_model
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcls/apis/init.py", line 2, in
from .inference import inference_model, init_model, show_result_pyplot
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcls/apis/inference.py", line 10, in
from mmcls.datasets.pipelines import Compose
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcls/datasets/init.py", line 2, in
from .base_dataset import BaseDataset
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcls/datasets/base_dataset.py", line 13, in
from mmcls.models.losses import accuracy
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcls/models/init.py", line 2, in
from .backbones import * # noqa: F401,F403
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcls/models/backbones/init.py", line 3, in
from .conformer import Conformer
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcls/models/backbones/conformer.py", line 9, in
from mmcv.cnn.bricks.transformer import AdaptivePadding
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcv/cnn/bricks/transformer.py", line 22, in
from mmcv.ops.multi_scale_deform_attn import MultiScaleDeformableAttention # noqa F401
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcv/ops/init.py", line 10, in
from .corner_pool import CornerPool
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcv/ops/corner_pool.py", line 8, in
ext_module = ext_loader.load_ext('_ext', [
File "/home/amir-e/.conda/envs/kf/lib/python3.8/site-packages/mmcv/utils/ext_loader.py", line 15, in load_ext
assert hasattr(ext, fun), f'{fun} miss in module {name}'
AssertionError: top_pool_forward miss in module _ext

Excuse me, I wanted to follow this wonderful work. I have successfully conducted KF experiments on cifar10 and
imagenet; however, a problem occurs when running KD experiments.
on cifar10, I run with: python tools/dist_train.py configs/cifar10-kd/wideresnet28-10_resnet18_b128x1
_cifar10.py 3
I get an error:

so I changed line 159 of the /KnowledgeFactor-main/cls/mmcls/models/classifiers/kd.py with:
loss_cls = self.criterionCls(student_logit, gt_label.squeeze(dim=1))
However, training with the modified code achieved more than 96% acc,which is higher than the acc reported in the paper

So I think my modification is wrong, would you please tell me the right way? thank you very much!

Excuse me, are these ckpts pretrained by ImageNet? I trained the wideresnet's teacher model with a pretrained ckpt solely on cifar10 initially and found it started at a accuracy of 50%, so I guess the ckpts aren't pretrained on the datasets of cifar10/100. Is that right?

Excuse, a question about the derivations of the Variational Bound for Mutual Information is not in the supplementary materials .

Nice work! But a problem occur when i test running :
Pretrained ResNet18 load successfully in the fold configs/Imagenet-kf ,while it shows "Teacher model not loaded “ when I test running with configs/imagenet-kf/wideresnet28-2*.py. Then I catch the exception , following is some of the error.

RuntimeError ('Error(s) in loading state_dict for ModuleDict:\n\tsize mismatch for backbone.layer1.0.conv1.weight: copying a param with shape torch.Size([160, 16, 3, 3]) from checkpoint, the shape in current model is torch.Size([32, 16, 3, 3]).\n\tsize mismatch for backbone.layer1.0.conv2.weight: copying a param with shape torch.Size([160, 160, 3, 3]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).\n\tsize mismatch for backbone.layer1.0.bn2.weight: copying a param with shape torch.Size([160]) from checkpoint, the shape in current model is torch.Size([32]).\n\tsize mismatch for backbone.layer1.0.bn2.bias: copying a param with shape torch.Size([160]) from checkpoint, the shape in current model is torch.Size([32]).\n\tsize mismatch for backbone.layer1.0.bn2.running_mean: copying a param with shape torch.Size([160]) from checkpoint, the shape in current model is torch.Size([32]).\n\tsize mismatch for backbone.layer1.0.bn2.running_var: copying a param with shape torch.Size([160]) from checkpoint, the shape in current model is torch.Size([32]).\n\tsize mismatch for backbone.layer1.0.downsample.0.weight: copying a param with shape torch.Size([160, 16, 1, 1]) from checkpoint, the shape in current model is torch.Size([32, 16, 1, 1]).

What does it mean?

What's the formula for it?