06 July 2021 - #Updated to windows usage
This is ImageNet dataset downloader. You can create new datasets from subsets of ImageNet by specifying how many classes you need and how many images per class you need. This is achieved by using image urls provided by ImageNet API.
In this blog post I wrote in a bit more detail how and why I wrote the tool. Also, I did a little analysis of the current state of the ImageNet URLs in the post.
This software is written in Python 3
The following command will randomly select 100 of ImageNet classes with at least 200 images in them and start downloading:
python ./downloader.py \
-data_root /data_root_folder/imagenet \
-number_of_classes 100 \
-images_per_class 200
The following command will download 500 images from each of selected class:
python ./downloader.py
-data_root /data_root_folder/imagenet \
-use_class_list True \
-class_list n09858165 n01539573 n03405111 \
-images_per_class 500
You can find class list in this csv where I list every class that appear in the ImageNet with number of total urls and total flickr urls it that class.
I've implementet parallel request processing and I've added multiprocessing_workers parameter which by default is 8. You can turn it higher, but I havent yet tested the limits of flickr allowed bandwith myself, so use it with care and you will have to find the limits yourself if you want to go for the maximum speed.
You can do something like this:
python ./downloader.py \
-data_root /data_root_folder/imagenet \
-number_of_classes 1000 \
-images_per_class 500 \
-multiprocessing_workers 24
The problem that prevents the script from working on Windows has been fixed. Minor changes in the script and with the recording of files.