Training Dataset on Amazon S3

I'm building an image / object classifier with TensorFlow. It works perfectly, however I want to build the database using Amazon S3.

I would like to know how to set image_dir to train my images using S3.

In this case, when I test with local images, I need to organize the images and folders as follows:

training_dataset/
-- dog/
---- img1.jpg
-- cat/
---- cat1.jpg
---- catt2.jpg
train.sh

My train.sh :

python retrain.py \
  --bottleneck_dir=tf_files/bottlenecks \
  --how_many_training_steps=500 \
  --model_dir=inception \
  --summaries_dir=tf_files/training_summaries/basic \
  --output_graph=tf_files/retrained_graph.pb \
  --output_labels=tf_files/retrained_labels.txt \
  --image_dir=training_dataset

The problem is, how to pass the image_dir parameter to the Amazon folder if the files are in CLOUD? Would I have to download the images to a folder? What if I have 30 million images? It does not roll.

Important: I'm using the TensorFlow repository, with the image sorter example, at the link below:

link