AnAppleCore / SVP-Neural-Network

Semi-bionic Visual Pathway Neural Network (SVPNN): a framework that simulates V1 and V2 at the front of neural networks

SVP-Neural-Network

Semi-bionic Visual Pathway Neural Network (SVPNN): A framework that simulates V1 and V2 at the front of neural networks

Introduction

SVPNN simulates the first half of human ventral visual pathway, by mimicking V1 and V2 area with fixed-weight Linear-Nonlinear models. Here's the overview of ventral pathway and SVPNN architecture. In this repository, we provide DenseNet-121, VGG-19 with batch normalization and AlexNet as candidates for back-end models.

---

Quick Start

NOTE: No Pretrained Model Available Now! Train before Validation please.

Requirements:

Dependencies:

tqdm
scipy
pandas
requests
torch >= 1.6.0
torchvision
imagecorruptions

Type the following command to install dependencies:

$ pip install -r requirements.txt

Train a model

The main script we use here is run.py, read it for more details.

For example if you want to train SVPNN with DenseNet-121 back-end:

$ python run.py --in_path /path/to/dataset/ImageNet --output_path /path/to/store/results --mode train --mode_arch densenet121

If you want to train DenseNet-121 without SVPNN framework, add -n or --no_svp at the tail of the above command.

$ python run.py --in_path /path/to/dataset/ImageNet --output_path /path/to/store/results --mode train --mode_arch densenet121 -n

Validation

Validation is a little bit, but the restore epoch number and model weight path is required (there two parameters will also be used when you want to restore the training process), for example:

$ python run.py -i /path/to/dataset/ImageNet -o ./results/densenet121 -repoch 30 -rpath ./results/densenet121 -m val -a densenet121

Robustness Evaluation is also provided, similar to validation but change --mode or -m to rval:

$ python run.py -i /path/to/dataset/ImageNet -o ./results/densenet121 -repoch 30 -rpath ./results/densenet121 -m val -a densenet121

Experiment Parameters setting:

Here we provide the parameters to adjust the environment settings for both training and validating. Here we show a context where 4 GPUs are used, epoch number 70, batch size 256, 10 dataloading workers and restore checkpoint from epoch 50 in path ./results/densenet121.

$ python run.py -i /path/to/dataset/ImageNet -o ./results/densenet121 -repoch 50 -rpath ./results/densenet121 -m train -g 4 -j 10 -e 70 -b 256 -n -a densenet121 -c speckle_noise -s 2