Skip to content

Latest commit

 

History

History
44 lines (40 loc) · 1.66 KB

README.md

File metadata and controls

44 lines (40 loc) · 1.66 KB

Setting Up Datasets

Required Datasets

  • ImageNet (ILSVRC2012)
  • COCO 2017

ImageNet

Download the ImageNet validation dataset and devkit from here (You will need to make an account). There should be no need to manually extract the files as Torch Vision should do it automatically.

COCO 2017

Download the COCO 2017 validation dataset and annotations from here and here. Extract the downloaded into the directory structure show below.

Example Directory Structure

📂datasets
 ┣ 📂coco
 ┃ ┣ 📂2017
 ┃ ┃ ┣ 📂annotations
 ┃ ┃ ┃ ┗ 📜instances_val2017.json
 ┃ ┃ ┣ 📂labels
 ┃ ┃ ┃ ┣ 📂train2017
 ┃ ┃ ┃ ┃ ┣ 📜000000000009.txt
 ┃ ┃ ┃ ┃ ┣ 📜000000000025.txt
 ┃ ┃ ┃ ┃ ┣ 📜000000000030.txt
 ┃ ┃ ┃ ┃ ┣ ... A lot more images
 ┃ ┃ ┣ 📂test2017
 ┃ ┃ ┃ ┣ 📜000000000001.jpg
 ┃ ┃ ┃ ┣ 📜000000000016.jpg
 ┃ ┃ ┃ ┣ 📜000000000019.jpg
 ┃ ┃ ┃ ┣ ... A lot more images
 ┗ 📂imagenet
 ┃ ┗ 📂2012
 ┃ ┃ ┣ 📂ILSVRC2012_devkit_t12
 ┃ ┃ ┣ 📂val
 ┃ ┃ ┃ ┣ 📂n01440764
 ┃ ┃ ┃ ┣ 📂n01443537
 ┃ ┃ ┃ ┣ 📂n01484850
 ┃ ┃ ┃ ┣ ...
 ┃ ┃ ┣ 📜ILSVRC2012_devkit_t12.tar.gz
 ┃ ┃ ┣ 📜ILSVRC2012_img_train.tar       - Torch Vision will use this directly without extracting
 ┃ ┃ ┣ 📜ILSVRC2012_img_val.tar
 ┃ ┃ ┣ 📜meta.bin

The training datasets are not required by this repo, but are included in example for completeness.