Ambiguous Datasets

Description

This repository contains ambiguous datasets generated using a conditional variational autoencoder (CVAE) approach. The datasets contain images that are ambiguous between a pair of classes. This class interpolation is done by conditional generation through the CVAE with a class-vector and mixing factor that gives ambiguous readout class-probability for the digit pair. Currently, only MNIST and EMNIST are supported, but the codebase is general enough that new datasets could be easily added.

How to run

First, install dependencies

# clone project   
git clone https://github.com/ABL-Lab/ambiguous-dataset

# install project   
cd ambiguous-dataset 
pip install -e .   
pip install -r requirements.txt

download ambiguous datasets (A-MNIST and A-EMNIST)

A-MNIST google drive link: https://drive.google.com/file/d/1JlJVoymk-3GAf9GHbgVeTTzQzPtvLUAt/view?usp=sharing

A-EMNIST google drive link: https://drive.google.com/file/d/19SspjkL24DngyIdLH5C0MTIXK0gQ2BhI/view?usp=sharing

Sequential MNIST (addition task): https://drive.google.com/file/d/1NMKNNo2lMp4Pg8UsF_cL5nzHRM5lMNRK/view?usp=sharing

Sequential A-MNIST (MNIST digits adding to ambiguous digit): https://drive.google.com/file/d/1vKvGwH_hvzQQrg8g7AxGxetvFEwfW_fQ/view?usp=sharing

Sequential A-EMNIST (EMNIST characters making a 3-letter word): https://drive.google.com/file/d/1TKN-B36AAvknmhS2L8Oknkr2HpXS62j-/view?usp=sharing

download trained model weights for experiments

gdrive link: https://drive.google.com/file/d/1a49OW61ShyAVZQos6iSLmBZTrTybvOtf/view?usp=sharing

unzip into a folder then:

from ambiguous.models.vae import MLPVAE
from ambiguous.models.readout import Readout
from ambiguous.models.cvae import Conv_CVAE

device='cuda' if torch.cuda.is_available() else 'cpu'

vae_path, readout_path, ccvae_path = ..., ..., ...
vae = MLPVAE(latent_dim=10, input_img_size=28).to(device)
vae.load_state_dict(torch.load(vae_path))

readout = Readout(latent_dim=10, h=512, n_classes=10).to(device)
readout.load_state_dict(torch.load(readout_path))

ccvae = Conv_CVAE(latent_dim=10,n_cls=10).to(device)
ccvae.load_state_dict(torch.load(ccvae_path))

if using cpu, pass map_location='cpu' when calling torch.load()

Importing Ambiguous Dataset to your own project

This project is setup as a package which means you can easily import any file into any other file like so.

from ambiguous.dataset.dataset import DatasetFromNPY, DatasetTriplet


root = 'path_to_ambiguous_dataset'
# older version
#trainset = DatasetFromNPY(root=root, download=False, train=True, transform=None)
#testset = DatasetFromNPY(root=root, download=False, train=False, transform=None)

# new version
trainset = DatasetTriplet(root=root, train=True, transform=None)
testset = DatasetTriplet(root=root, train=False, transform=None)

trainloader = DataLoader(trainset, batch_size=64, shuffle=True)
testloader = DataLoader(testset, batch_size=64, shuffle=False)
(clean1, amb, clean2), label = next(iter(amnist_loader))

Here are some examples of the generated triplets (clean, ambiguous, clean):

Citation

@article{YourName,
  title={Your Title},
  author={Your team},
  journal={Location},
  year={Year}
}

abl-lab / ambiguous-dataset Goto Github PK