An implementation of adding biologically inspired shape bias to popular data sets, exploring the effects on common CNN architectures.
The code is intended to run in Google Colab (Colab), although can be configured if you have your own GPU.
Open a Colab notebook (or alternatively see the link for a set up - open in playground to run, may need to connect to GPU [step 1]: https://colab.research.google.com/drive/1UWgWIGu0RCuGbANeQPRXZ_Ta4wYKlmZq)
There are a few key steps to get up and running in the Colab environment.
- Connect to hosted GPU. (runtime -> change runtime type -> Hardware -> GPU)
- Download the repository/source code (using git clone preferably)
- Install requirements in the cloned repository
- Choose dataset, architecture, transformations and other hyperparameters you wish to run & run command
Example of set up in Colab:
# Note: '!' in Colab is to run command line
# Clone Repository
!git clone https://github.com/Mikcl/ShapeBias.git
# Change Directory
%cd ./ShapeBias/
# Install Requirements
!pip install -r requirements.txt
# Run Experimenent - train a VGG-16 model on dataset 0 (CIFAR-10)
!python main.py -a vgg-16 --dataset 0
Ensure all requirements are installed before running experiments: pip install -r requirements.txt
usage: main.py [-h] [-a ARCH] [-j N] [--epochs N] [--start-epoch N] [-b N]
[--lr LR] [--momentum M] [--wd W] [-p N] [-e] [--pretrained]
[--gpu GPU] [--own PATH] [-f] [-l LF] [-c CHANNELS] [--decay D]
[-d DATASET] [--data DATA] [--concat] [--same] [--DOG]
[-o O [O ...]] [--gabor] [-s S [S ...]] [-u U [U ...]]
[--savecsv] [--savemodel]
Shape Bias Training
optional arguments:
-h, --help show this help message and exit
-a ARCH, --arch ARCH model architecture: alexnet | densenet121 |
densenet161 | densenet169 | densenet201 | googlenet |
inception_v3 | mnasnet0_5 | mnasnet0_75 | mnasnet1_0 |
mnasnet1_3 | mobilenet_v2 | resnet101 | resnet152 |
resnet18 | resnet34 | resnet50 | resnext101_32x8d |
resnext50_32x4d | shufflenet_v2_x0_5 |
shufflenet_v2_x1_0 | shufflenet_v2_x1_5 |
shufflenet_v2_x2_0 | squeezenet1_0 | squeezenet1_1 |
vgg11 | vgg11_bn | vgg13 | vgg13_bn | vgg16 | vgg16_bn
| vgg19 | vgg19_bn | wide_resnet101_2 |
wide_resnet50_2 (default: resnet18)
-j N, --workers N number of data loading workers (default: 4)
--epochs N number of total epochs to run
--start-epoch N manual epoch number (useful on restarts)
-b N, --batch-size N mini-batch size (default: 256), this is the total
batch size of all GPUs on the current node when using
Data Parallel or Distributed Data Parallel
--lr LR, --learning-rate LR
initial learning rate
--momentum M momentum
--wd W, --weight-decay W
weight decay (default: 1e-4)
-p N, --print-freq N print frequency (default: 10)
-e, --evaluate evaluate model on validation set
--pretrained use pre-trained model
--gpu GPU GPU id to use.
--own PATH path to your own given model state dict (note arch of
this model must match -a)
-f, --finetune fine-tune model fc layer on training set
-l LF, --loadfinetuned LF
use fintuned own model with defined number of output
classes
-c CHANNELS, --channels CHANNELS
number of channels (default: 3)
--decay D decay learning rate every D epochs by factor of 10
-d DATASET, --dataset DATASET
which dataset to train on, default- None (define
custom path to directory), 0 - CIFAR10, 1 - CIFAR100
--data DATA path to custom dataset and where output folder is
--concat concat transformed data with original data
--same train and validated on same (type of transformation)
dataset, primarily used from custom transformations
--DOG DOG transformation, use --options for non default
hyper parameters, [sigma k]
-o O [O ...], --options O [O ...]
options (list) for the transformation, pass as: -o s k
--gabor gabor 2D CWT
-s S [S ...], --scales S [S ...]
scales (list) for the 2D Gabor Wavelet: -s 2 2.5
-u U [U ...], --orientations U [U ...]
orientations (list) for the 2D Gabor Wavelet: -u 1 2 3
--savecsv save the csv to google drive [only in collab]
--savemodel save the model to google drive [only in collab]
Can perform experiments on various datasets.
To utilise a custom dataset, use --data ./path/to/dataset, where .../dataset/ contains the directories /dataset/train and /dataset/val where images are grouped in directories (representing their class name).
Example:
...dataset/
...dataset/train
...dataset/train/dog/
...dataset/train/dog/bulldog.jpeg
...dataset/train/dog/poodle.jpeg
...dataset/train/car/
...dataset/train/car/hatchback.jpeg
...dataset/train/car/jeep.jpeg
...
...dataset/val
...dataset/val/dog/
...dataset/val/dog/boxer.jpeg
...dataset/val/car/
...dataset/val/car/limousine.jpeg
...
To download and format the Tiny ImageNet dataset, run ./scripts/prep.sh
Support for CIFAR-10 and CIFAR-100 have been added thus far, which can be utilised by --dataset N
| Data set | N |
|---|---|
| CIFAR-10 | 0 |
| CIFAR-100 | 1 |
-
Train CNN vgg-16 (learning rate=0.01, weight_decay=0.0005) on CIFAR-10 and validate on CIFAR-10
python main.py -a vgg-16 --lr 0.01 --wd 0.0005 --dataset 0 -
Train for 75 epochs and reduce the learning rate by a factor of 10 every 30 epochs
python main.py --dataset 0 --decay 30 --epochs 75 -
Train on CIFAR-10 and CIFAR-10 transformed by a Difference Of Gaussian and validate on CIFAR-10, and save the model into Google Drive with
--savemodelpython main.py --dataset 0 --concat --DOG --savemodel -
Train on only CIFAR-10 transformed by a 2D Gabor continuous wavelet transform (2D-CWT) and validate only on the same transform applied to CIFAR-10 validation, hyperparameter of orientation
-u 2is used for the wavelet.(note:
-uand-scan take multiple orientations and scales respectively for the--gabor2D Wavelet)python main.py --dataset 0 --same --gabor -u 2 -
Evaluate a model
./path/to/model.pth.taron a CIFAR-100 dataset.(note: should be trained to have 100 classes as output)
python main.py --own ./path/to/model.pth.tar --dataset 1 -e -
Finetune a model
./path/to/model.pth.taron a CIFAR-100 datasetpython main.py --own ./path/to/model.pth.tar --dataset 1 -f -
Evaluate a finetuned model
./path/finetuned/model.pth.taron a CIFAR-100 dataset. Pass 100 because 100 classes.python main.py --own ./path/finetuned/model.pth.tar --dataset 1 -l 100 -e -
Train on a custom dataset
./path/to/dataset/python main.py --data ./path/to/dataset/
Please see Image Net Example for details on Multi-processing Distributed Data Parallel Training if you wish to extend functionality
https://github.com/pytorch/examples/
When attempting to fine tune models, you may find that you will need import the model saved into the colab environment. With large model files this will take considerable time to 'Upload' the file directly to the Colab environment. A more efficient way to do this is (in Colab) to Mount your google drive, from the files section. Then run the script below inside ./ShapeBias/ to retrieve model.pth.tar from your Drive.
!cp ../drive/My\ Drive/model.pth.tar ./
Which will copy the model into your working directory very quickly. Files inside your drive can be found at drive.google.com