experiment result

[MrLinNing]

Hello peterliht， I ran through your code according to the instructions, did not modify any parameters, but found that the results vary greatly. What parameters did you modify before releasing the code? The following experimental results on resnet18： python train.py --model_dir experiments/resnet18_distill/resnext_teacher

My experimental environment is：

python 3.5.2
pytorch 0.4.0
GPU  TITAN Xp

[xht033]

me too. A huge gap between my experiment and the author's report.

[ChuangbinC]

@MrLinNing Do you get the experiment result closed to the author's result?

[michaelklachko]

I just ran an experiment on CIFAR-10, with the student being a simple LeNet-5 like network (64C - MP - 128C - MP - 400FC-10), and the teacher is a deeper version (128C-128C-MP-128C-128C-MP-128C-128C-512FC-10).

The teacher gets to ~93% accuracy, the student without KL is ~86.5%. With KL, the student gets to 87.5% consistently.

I didn't use this repo code, only copied the KL loss function to my code.

[xiaowenmasfather]

I found a 10% gap too, 84% nowhere near the expected 94.788% . Student net: Resnet-18, Teacher net: Resnext29. Parameters are the same with @peterliht 's original settings.

[wnma3mz]

I also got similar results. train_set: 84.914%, test_set: 83.89% The teacher model is derived from the author's pretrained_teacher_models.zip\pretrained_teacher_models\base_resnext29\. The accuracy of testing this model is: train_set: 100%, test_set: 96.23%. Other parameters are consistent with the author.

[haitongli]

Looking through another thread of issue discussions on the data loader, the accuracy inconsistency might be due to the way how student and teacher models got their data when we used shuffling.

[haitongli]

Has anyone used pytorch 0.3 to run and test?

[michaelklachko]

@peterliht why would you want to use Pytorch 0.3? The current stable version is 1.2.

@wnma3mz @xiaowenmasfather Resnet-18 should get to 94.0% without any teachers. If that's not the case, then you're doing something wrong.

[haitongli]

@peterliht why would you want to use Pytorch 0.3? The current stable version is 1.2.

@wnma3mz @xiaowenmasfather Resnet-18 should get to 94.0% without any teachers. If that's not the case, then you're doing something wrong.

I understand there is newer (and more stable) version of pytorch available. I just wanted to understand if people have seen different results across different pytorch versions. When first creating this repo 2 years ago, as specified in requirements.txt, v0.3 was used. I want to get a better understanding of issues that have prevented people from reproducing results and see if fixes can be done along with the most stable pytorch version.

[wnma3mz]

Hi @michaelklachko You‘re right. Resnet-18 with the author's hyperparameters can indeed reach 94%. So my doubt is, where is the problem? Has anyone encountered the same problem and helped me?

@peterliht Thanks for your suggestion, I will try it on version 0.3 later.

[haitongli]

@wnma3mz another thread might also be worth looking into @ #9 and also @ #4

[wnma3mz]

@peterliht Thank you for your prompt reply. I have already seen this issue, I have changed code according to this comment to ensure the correctness of the distillation.

[forjiuzhou]

@wnma3mz another thread might also be worth looking into @ #9 and also @ #4

I compare the max index of teacher's output with label, these two disagree with each other. I have commit a request to fix this issue.

[conditionWang]

I met the same problem of accuracy gap. I have tried adjusting the learning rate to a small one and observed an improvement, making my results close to those of Peterliht. You can try changing the learning rate and running the code again.

[tianli]

@wnma3mz another thread might also be worth looking into @ #9 and also @ #4

I compare the max index of teacher's output with label, these two disagree with each other. I have commit a request to fix this issue.

Your request (https://github.com/peterliht/knowledge-distillation-pytorch/pull/17) fix the problem and I am getting much improved result. I wonder why it is not merged into the master yet!

[tianli]

FYI, with the pull request #17, I was able to get accuracy 95.19% on reset18 with the resnext29 teacher.

[haitongli]

Thanks for all the discussions and the reminder from @tianli about the pull request. I haven't been able to keep track of this repo for a while. #17 has been merged.