bbbdylan / unisiam Goto Github PK

I'm sorry to bother you again.I'm wondering how long does it probably take to run a test? Is it normal to have no reaction for a long time.Looking forward to your reply!

Hi there,

I am currently trying to utilise your approach. More specifically I am trying to use your pre-trained models as porvided here (https://drive.google.com/drive/folders/1N_5ZiI73TfFFFOudWDPuNXeYc2dHzKYU).

Currently, I am attempting to use your codebase as cloned with the pre-trained 'mini_r18.pth' model. Using this model for evaluation via

python ./train.py --data_path D:\Downloads\miniimagenet\miniimagenet --eval_path mini_r18.pth

results in significantly lower performance than what is stated in the original paper, see below:

Environment details: From what I can gather, a requirements or reproducible environment file is not supplied with the repo, however I have done my best to reproduce based on your other closed issues. I am using pytorch 1.8 with cuda 10.2 and python 3.7

Any help with rectifying or explaining this difference in obtained performance would be greatly appreciated!

Cheers!

FileNotFoundError: [Errno 2] No such file or directory: '.datasetminiminiimagenetimages/images/n0212916500000959.jpg'
您好，请问我感觉数据和路径都没问题，为什么会有这样的情况呢？

Hi there,

I ended up having some more luck with a previous issue I opened!

I am wondering now whether it would be possible to release the code used to train the original SimCLR model? It seems from the paper you used your own implementation.

Any help is greatly appreciated!

Cheers,
Calum

We followed the settings of config/unisiam/tiered/r18.sh and re-trained the model with backbone resnet18 on tieredImageNet, and its performance as follows,

classifier: LR, power_norm: True, 5-way 1-shot acc: 63.24+0.40
classifier: LR, power_norm: True, 5-way 5-shot acc: 81.14+0.31

It severe below the paper, how can I improve the performance? Can you release the guide of how to train the model on tieredImageNet?

Thank you very much.

Hi!
Greet job first! Thanks for your efforts.
I ran on my computer but didn’t get the desired results. Could you please share the configuration information for the pre-trained models?

I just reproduced your project on miniImageNet. Wonderful work!
But I've got a question, have you used ResNet-50 for knowledge distillation in your miniImageNet experiments, and how do the results compare to ResNet-18/ResNet-34. I didn't find it in the paper. Looking forward to your reply!