hendrycks / pre-training Goto Github PK

Hi hendrycks,

Thanks for providing the finetuned models for the uncertainty task on Tiny Imagenet.

When I run the test.py using the snapshots/tune/wrn_baseline_epoch_19.pt under uncertainty/TinyImageNet. I found the test error is around 99.47% but it should be expected to be around 35% according to the csv log. So I wonder if there needs some preprocessing for the Tiny ImageNet dataset? Currently, I directly use the val part of Tiny Imagenet from https://image-net.org/download-images.php. The whole size of the Tiny Imagenet is 236 MB.

Many Thanks.

Hi,

In the paper, it writes: "Without assuming such knowledge, we use the maximum softmax probabilities to score anomalies and show that models which are pre-trained then tuned provide superior anomaly scores". So I want to confirm whether you fine-tune the whole model or just the classifier (the last layer). Thank you.

Best Regards,
Hongxin

Hi Hendrycks,

I want to evaluate the adversarial robustness of your proposed method about the datasets CIFAR10 and CIFAR100, can you provide the pretrained models of the proposed method about these two datasets?
Thank you!

Best wishes,
Gavin

Hi, I assume this is the input normalization of mean=0.5 and std=0.5.

• Is this a common way for transfer learning?
• If you have tried, what happens if there's no normalization?

Hi,
Could you share the dataset or the script for removing 153 CIFAR-10-related classes from the Downsampled ImageNet dataset?

Hi, Thank you for the great work shared. Could you please help to identify correct files to train a model using ImageNet where CIFAR10 related classes (listed in the paper) removed. e.g. to remove n03345487 class how can I update the imagenet_downsampled.py file?

Hi, where can i get the pretrained models?