cbas

Quick setup

• Download/clone repo
• Move images folder (and it's contents) out from under the cbas root directory

cd cbas
cp -r images ../

• cd into images folder and run the shell script to unzip the images

cd ../images
bash setup_images.sh

Full setup

1.) Download/clone the COCO api from the COCO github page

• You can follow their readme or this one. For this one, you don't have to download the COCO dataset
• After unzipping, rename the root cocoapi-master/ to just coco/ or move the contents of cocoapi-master/ to an empty folder named coco/
• Create two additional folders:
  • coco/images/
  • coco/annotations/
• To install:
  • Run make under coco/PythonAPI/

2.) [OPTIONAL] Download the COCO dataset from the COCO download page

This is only necessary if you want to run the coco demos and/or build cbas from scratch. The training set is 18GB but downloads surprisingly fast.

On the COCO download page select:

• the "2017 Train images [118K/18GB]" link
  • The 2017 COCO eval set is so small compared to the train set (5K vs 118K) that I just split the 118K train set into train and val and didn't bother with their val set. It wouldn't hurt to add it to our val set.
• the "2017 Train/Val annotations [241MB]" link

Unzip, and place:

• the train2017/ image folder (containing 118k images) in: coco/images/
  • e.g. coco/images/train2017/
• the annotations in: coco/annotations/

3.) Download/clone this cbas repository

• Clone the cbas/ repository into the coco/ directory:
  • e.g.: coco/cbas/

4.)[OPTIONAL] Set up the pre-made CBAS-34 dataset.

• Unzip the and place into the coco/images/ folder:
  • e.g.coco/images/cbas34_train/
  • e.g.coco/images/cbas34_val/

5.) [OPTIONAL] Create CBAS-80 and CBAS-34 from scratch

• If you downloaded COCO, you should be able to run create_cbas80_and_cbas36.ipynb which will walk through creating these datasets

6.) Run the training demo on CBAS-34

• If you skipped steps 2 & 5, make sure to complete step 4.
• Now you can run the PyTorch demo cbas34_train_demo.ipynb to train and evaluate LeNet on CBAS-36