qinguisun/autows-bench-101

Auto Weak Supervision Benchmark

AutoWS-Bench-101 🪑

Introduction

AutoWS-Bench-101 is a framework for evaluating automated WS (AutoWS) techniques in challenging WS settings---a set of diverse application domains on which it has been previously difficult or impossible to apply traditional WS techniques.

Installation

Install anaconda: Instructions here: https://www.anaconda.com/download/

Clone the repository:

git clone https://github.com/Kaylee0501/AutoWS-Bench-101.git
cd AutoWS-Bench-101

Create virtual environment:

conda env create -f environment.yml
source activate AutoWS-Bench-101

Install CLIP:

pip install git+https://github.com/openai/CLIP.git

Datasets

Our benchmark automatic download the dataset for you. Please run data_settings.py to download the specific dataset you need.

Name	# class	# train	# valid	# test
MNIST	10	57000	3000	10000
FashionMNIST	10	57000	3000	10000
KMNIST	10	57000	3000	10000
CIFAR10	10	47500	2500	10000
SphericalMNIST	10	57000	3000	10000
PermutedMNIST	10	57000	3000	10000
ECG	4	280269	14752	33494
EMBER	2	285000	15000	100000
NavierStokes	2	100	100	1900

Run the Experiment

To run the whole implementation, we provide a pipeline. This pipeline will walk you through a full example of our framework. It allows you to choose the datasets and automatic download for you, select the embeddings, and generate a bunch of labeling functions (LFs) with the LF selectors you preferred. It then trains a Snorkel label model and gives you the accuracy and coverage.

Please run (feel free to change the argument):

python fwrench/applications/pipeline.py --dataset='cifar10' --embedding='openai' --lf_selector='iws'
python fwrench/applications/pipeline.py --dataset='ember' --embedding='pca' --lf_selector='snuba'

Examples

Training MNIST with pca embedding and snuba selector

import logging
import random
import copy

import fire
import fwrench.embeddings as feats
import fwrench.utils.autows as autows
import fwrench.utils.data_settings as settings
import numpy as np
import torch
from sklearn.decomposition import PCA
from sklearn.metrics import accuracy_score
from wrench.logging import LoggingHandler

def main(
    dataset="mnist",
    dataset_home="./datasets",
    embedding="pca",  # raw | pca | resnet18 | vae

    lf_selector="snuba",  # snuba | interactive | goggles
    em_hard_labels=False,  # Use hard or soft labels for end model training
    n_labeled_points=100,  # Number of points used to train lf_selector
    #
    # Snuba options
    snuba_combo_samples=-1,  # -1 uses all feat. combos
    # TODO this needs to work for Snuba and IWS
    snuba_cardinality=1,  # Only used if lf_selector='snuba'
    snuba_iterations=23,
    lf_class_options="default",  # default | comma separated list of lf classes to use in the selection procedure. Example: 'DecisionTreeClassifier,LogisticRegression'
    seed=123,
    prompt=None,
):
    ################ HOUSEKEEPING/SELF-CARE 😊 ################################
    random.seed(seed)
    logging.basicConfig(
        format="%(asctime)s - %(message)s",
        datefmt="%Y-%m-%d %H:%M:%S",
        level=logging.INFO,
        handlers=[LoggingHandler()],
    )
    logger = logging.getLogger(__name__)
    device = torch.device("cuda")

    ################ LOAD DATASET 😊  ##########################################
    train_data, valid_data, test_data, k_cls, model = settings.get_mnist(
            n_labeled_points, dataset_home
    )
    emb = PCA(n_components=100)
    embedder = feats.SklearnEmbedding(emb)

    embedder.fit(train_data, valid_data, test_data)
    train_data_embed = embedder.transform(train_data)
    valid_data_embed = embedder.transform(valid_data)
    test_data_embed = embedder.transform(test_data)

    ################ AUTOMATED WEAK SUPERVISION ###############################
    test_covered, hard_labels, soft_labels = autows.run_snuba(
        valid_data, train_data, test_data, valid_data_embed,
        train_data_embed, test_data_embed, snuba_cardinality,
        snuba_combo_samples, snuba_iterations, lf_class_options,
        k_cls, logger,
    )
    acc = accuracy_score(test_covered.labels, hard_labels)
    cov = float(len(test_covered.labels)) / float(len(test_data.labels))
    logger.info(f"label model train acc:    {acc}")
    logger.info(f"label model coverage:     {cov}")
    return acc

if __name__ == "__main__":
    fire.Fire(main)

Training ECG with resnet18 embedding and goggles selector

import logging
import random
import copy

import fire
import fwrench.embeddings as feats
import fwrench.utils.autows as autows
import fwrench.utils.data_settings as settings
import numpy as np
import torch
from sklearn.decomposition import PCA
from sklearn.metrics import accuracy_score
from wrench.logging import LoggingHandler

def main(
    dataset="ecg",
    dataset_home="./datasets",
    embedding="resnet18",  # raw | pca | resnet18 | vae

    lf_selector="goggles",  # snuba | interactive | goggles
    em_hard_labels=False,  # Use hard or soft labels for end model training
    n_labeled_points=100,  # Number of points used to train lf_selector
    #
    lf_class_options="default",  # default | comma separated list of lf classes to use in the selection procedure. Example: 'DecisionTreeClassifier,LogisticRegression'
    seed=123,
    prompt=None,
):
    ################ HOUSEKEEPING/SELF-CARE 😊 ################################
    random.seed(seed)
    logging.basicConfig(
        format="%(asctime)s - %(message)s",
        datefmt="%Y-%m-%d %H:%M:%S",
        level=logging.INFO,
        handlers=[LoggingHandler()],
    )
    logger = logging.getLogger(__name__)
    device = torch.device("cuda")

    ################ LOAD DATASET #############################################
    train_data, valid_data, test_data, k_cls, model = settings.get_ecg(
            n_labeled_points, dataset_home
    )
    embedder = feats.ResNet18Embedding(dataset)

    embedder.fit(valid_data, test_data)
    valid_data_embed = embedder.transform(valid_data)
    test_data_embed = embedder.transform(test_data)
    train_data_embed = copy.deepcopy(valid_data_embed)
    train_data = copy.deepcopy(valid_data)

    ################ AUTOMATED WEAK SUPERVISION ###############################
    test_covered, hard_labels, soft_labels = autows.run_goggles(
        valid_data, train_data, test_data, valid_data_embed,
        train_data_embed, test_data_embed, logger,
    )
    acc = accuracy_score(test_covered.labels, hard_labels)
    cov = float(len(test_covered.labels)) / float(len(test_data.labels))
    logger.info(f"label model train acc:    {acc}")
    logger.info(f"label model coverage:     {cov}")
    return acc

if __name__ == "__main__":
    fire.Fire(main)

Credits

We extend the WRENCH codebase to build our framework. Thanks for their inspiration.