examples

Training Integral neural networks.

Edsr

python edsr.py --integral --batch-size 16

Specify super resolution scale factor with argument --scale (default=4).

Imagenet

python imagenet.py --integral <IMAGENET FOLDER PATH>

Add --data-parallel to use DataParallel training.

Evaluation of trained INNs.

To resample (prune) and evaluate the integral model run commands below:

Edsr

python edsr.py --integral --resample --checkpoint=<INTEGRAL MODEL CHECKPOINT> --evaluate --batch-size 16

Imagenet

python imagenet.py --integral --resample --evaluate --checkpoint <INTEGRAL MODEL CHECKPOINT> <IMAGENET FOLDER PATH> 

Checkpoints of trained integral models are available at link.

Fast pruning of DNNs.

Here we do not train integral neural networks. Instead we convert pre-trained DNN to integral and tune integration partition with desired size only.

DNN -> INN -> Integration grid tuning.

python edsr.py --integral --resample --grid-tuning --batch-size 16