Awesome

PRANC

This is the official code for paper PRANC: Pseudo RAndom Networks for Compacting deep models. PRANC is a method to compact the knowledge of a Deep Neural Net for over 50x. Here is the link to the paper.

Requirements:

PyYAML==6.0 torch==1.10.2 torchvision==0.11.3 tqdm==4.62.3

or you can simply use: pip3 install -r requirements.txt

Running:

For running the code, simply use: python3 launcher.py <config-file>

Config file content:

name: description of experiment
id: unique identifier, your choice but be careful, will be used to store the signature and model
gpus: list of available GPUs

pranc.seed: seed for initializing basis networks
num_alpha: number of basis networks

experiments.mode: [train, test] mode of the experiment
experiment.method: [normal, pranc] method of the experiment
experiment.loss: [cross-entropy, mse] loss function
experiment.lr: learning rate, can be ignored when testing
experiment.optimizer: [sgd, adam] training optimizer
experiment.momentum: momentum for sgd, can be ignored for other optimizers
experiment.weight_decay: weight decay for sgd, can be ignored for other optimizers
experiment.scheduler: [none, step, exponential], learning rate scheduler
experiment.gamma: gamma for exponential and step scheduler 
experiment.step: step for step scheduler
experiment.epoch: number of training epochs
experiment.batch_size: training batch size, optional for testing
experiment.resume: '<TASK_ID>/pranc'  for resuming pranc training. 
experiment.resume: '<TASK_ID>/best_model.pt'  for resuming normal training. 
experiment.load_model: '<TASK_ID>/pranc' for pranc testing
experiment.load_model: '<TASK_ID>/best_model.pt' for normal testing 
experiment.task: [mnist, cifar10, cifar100, tiny] the task that is going to be solved
experiment.model_arch: [lenet, resnet20, resnet56, alexnet, convnet] model architecture used in experiment

dataset.image_width: input image width, set 28 for mnist, 32 for cifar, 64 tiny, 256 for imagenet
dataset.dataset_path: path to the dataset

monitor.log_rate: training log rate (based on batches)
monitor.save_model: path to save the model. if touch, modify resume and load_model
monitor.save_path: path to save the pranc signature. if touch, modify resume and load_model

There is a sample config file in configs