Did you train the teacher net for each dataset, or just used the pre-trained weights of vgg?
Will there be a lot difference between?

They resize the image to 256x256

Have you achieved the PRO score in this paper? I am also trying to implementation this paper but this metric is always far from the paper's announcement.

Hello! I am trying to run your code in a gcp vm with all requirements are satisfied, but when I run the docker image for pytorch/tensorrt, i got the following, but docker ps doesnt return any active containers. Can you briefly explain how to run the code in docker from your docker image?

sudo docker run --runtime=nvidia taikiinoue45/pytorch:tensorrt

== NVIDIA TensorRT ==

NVIDIA Release 20.03 (build 10567415)
NVIDIA TensorRT 7.0.0 (c) 2016-2020, NVIDIA CORPORATION. All rights reserved.
Container image (c) 2020, NVIDIA CORPORATION. All rights reserved.
https://developer.nvidia.com/tensorrt
To install Python sample dependencies, run /opt/tensorrt/python/python_setup.sh
To install open source parsers, plugins, and samples, run /opt/tensorrt/install_opensource.sh. See https://github.c
om/NVIDIA/TensorRT/tree/20.03 for more information.

To obtain a high-resolution anomaly detection map, a lot of CNN forward computations are required on the whole image in a sliding window manner. This is a time-consuming process, and it takes one hour per image.

Fast Dense Feature Extraction with CNNs that have Pooling or Striding Layers presented a novel approach to nearly arbitrary CNN architectures for fast execution on the whole image in a sliding window manner. Therefore, I tried to implement it in STAD.