Mind Your Outliers!
Mind Your Outliers! Investigating the Negative Impact of Outliers on Active Learning for Visual Question Answering
Siddharth Karamcheti, Ranjay Krishna, Li Fei-Fei, Christopher D. Manning
Annual Meeting for the Association of Computational Linguistics (ACL-IJCNLP) 2021.
Code & Experiments for training various models and performing active learning on a variety of different VQA datasets and splits. Additional code for creating and visualizing dataset maps, for qualitative analysis!
If there are any trained models you want access to that aren't easy for you to train, please let me know and I will do my best to get them to you. Unfortunately finding a hosting solution for 1.8TB of checkpoints hasn't been easy
Quickstart
Clones vqa-outliers
to the current working directory, then walks through dependency setup, mostly leveraging the environments/environment-
files. Assumes conda
is installed locally (and is on your path!). Follow the directions here to install conda
(Anaconda or Miniconda) if not.
We provide two installation directions -- one set of instructions for CUDA-equipped machines running Linux w/ GPUs (for training), and another for CPU-only machines (e.g., MacOS, Linux) geared towards local development and in case GPUs are not available.
The existing GPU YAML File is geared for CUDA 11.0 -- if you have older GPUs, file an issue, and I'll create an appropriate conda configuration!
Setup Instructions
# Clone `vqa-outliers` Repository and run Conda Setup
git clone https://github.com/siddk/vqa-outliers.git
cd vqa-outliers
# Ensure you're using the appropriate hardware config!
conda env create -f environments/environment-{cpu, gpu}.yaml
conda activate vqa-outliers
Usage
The following section walks through downloading all the necessary data (be warned -- it's a lot!) and running both the various active learning strategies on the given VQA datasets, as well as the code for generating Dataset Maps over the full dataset, and visualizing active learning acquisitions relative to those maps.
Note: This is going to require several hundred GB of disk space -- for targeted experiments, feel free to file an issue and I can point you to what you need!
Downloading Data
We have dependencies on a few datasets, some pretrained word vectors (GloVe), and a pretrained multimodal model (LXMERT), though not the one commonly released in HuggingFace Transformers. To download all dependencies, use the following commands from the root of this repository (in general, run everything from repository root!).
# Note: All the following will create/write to the directory data/ in the current repository -- feel free to change!
# GloVe Vectors
./scripts/download/glove.sh
# Download LXMERT Checkpoint (no-QA Pretraining)
./scripts/download/lxmert.sh
# Download VQA-2 Dataset (Entire Thing -- Questions, Raw Images, BottomUp Object Features)!
./scripts/download/vqa2.sh
# Download GQA Dataset (Entire Thing -- Questions, Raw Images, BottomUp Object Features)!
./scripts/download/gqa.sh
Additional Preprocessing
Many of the models we evaluate in this work use the object-based BottomUp-TopDown Attention Features -- however, our Grid Logistic Regression and LSTM-CNN Baseline both use dense ResNet-101 Features of the images. We extract these from the raw images ourselves as follows (again, this will take a ton of disk space):
# Note: GPU Recommended for Faster Extraction
# Extract VQA-2 Grid Features
python scripts/extract.py --dataset vqa2 --images data/VQA-Images --spatial data/VQA-Spatials
# Extract GQA Grid Features
python scripts/extract.py --dataset gqa --images data/GQA-Images --spatial data/GQA-Spatials
Running Active Learning
Running Active Learning is a simple matter of using the script active.py
in the root of this directory. This script is able to reproduce every experiment from the paper, and allows you to specify the following:
- Dataset in
< vqa2 | gqa >
- Split in
< all | sports | food >
(for VQA-2) andall
for GQA - Model (mode) in
< glreg | olreg | cnn | butd | lxmert >
(Both Logistic Regression Models, LSTM-CNN, BottomUp-TopDown, and LXMERT, respectively) - Active Learning Strategy in
< baseline | least-conf | entropy | mc-entropy | mc-bald | coreset-{fused, language, vision} >
following the paper. - Size of Seed Set (burn, for burn-in) in
< p05 | p10 | p25 | p50 >
where each denotes percentage of full-dataset to use as seed set.
For example, to run the BottomUp-TopDown Attention Model (butd
) with the VQA-2 Sports Dataset, with Bayesian Active Learning by Disagreement, with a seed set that's 10% the size of the original dataset, use the following:
# Note: If GPU available (recommended), pass --gpus 1 as well!
python active.py --dataset vqa2 --split sports --mode butd --burn p10 --strategy mc-bald
File an issue if you run into trouble!
Creating Dataset Maps
Creating a Dataset Map entails training a model on an entire dataset, while maintaining statistics on a per-example basis, over the course of training. To train models and dump these statistics, use the top-level file cartograph.py
as follows (again, for the BottomUp-TopDown Model, on VQA2-Sports):
python cartograph.py --dataset vqa2 --split sports --mode butd
Once you've trained a model and generated the necessary statistics, you can plot the corresponding map using the top-level file chart.py
as follows:
# Note: `map` mode only generates the dataset map... to generate acquisition plots, see below!
python chart.py --mode map --dataset vqa2 --split sports --model butd
Note that Dataset Maps are generated per-dataset, per-model!
Visualizing Acquisitions
To visualize the acquisitions of a given active learning strategy relative to a given dataset map (the bar graphs from our paper), you can run the following (again, with our running example, but works for any combination):
python chart.py --mode acquisitions --dataset vqa2 --split sports --model butd --burn p10 --strategies mc-bald
Note that the script chart.py
defaults to plotting acquisitions for all active learning strategies -- either make sure to run these out for the configuration you want, or provide the appropriate arguments!
Ablating Outliers
Finally, to run the Outlier Ablation experiments for a given model/active learning strategy, take the following steps:
- Identify the different "frontiers" of examples (different difficulty classes) by using
scripts/frontier.py
- Once this file has been generated, run
active.py
with the special flag--dataset vqa2-frontier
and the arbitrary strategies you care about. - Sit back, examine the results, and get excited!
Concretely, you can generate the frontier
files for a BottomUp-TopDown Attention Model as follows:
python scripts/frontier.py --model butd
Any other model would also work -- just make sure you've generated the map via cartograph.py
first!
Results
We present the full set of results from the paper (and the additional results from the supplement) in the visualizations/
directory. The sub-directory active-learning
shows performance vs. samples for various splits of strategies (visualizing all on the same plot is a bit taxing), while the sub-directory acquisitions
has both the dataset maps and corresponding acquisitions per strategy!
Start-Up (from Scratch)
Use these commands if you're starting a repository from scratch (this shouldn't be necessary to use/build off of this code, but I like to keep this in the README in case things break in the future). Generally, you should be fine with the "Usage" section above!
Linux w/ GPU & CUDA 11.0
# Create Python Environment (assumes Anaconda -- replace with package manager of choice!)
conda create --name vqa-outliers python=3.8
conda activate vqa-outliers
conda install pytorch torchvision torchaudio cudatoolkit=11.0 -c pytorch
conda install ipython jupyter
conda install pytorch-lightning -c conda-forge
pip install typed-argument-parser h5py opencv-python matplotlib annoy seaborn spacy scipy transformers scikit-learn
Mac OS & Linux (CPU)
# Create Python Environment (assumes Anaconda -- replace with package manager of choice!)
conda create --name vqa-outliers python=3.8
conda activate vqa-outliers
conda install pytorch torchvision torchaudio -c pytorch
conda install ipython jupyter
conda install pytorch-lightning -c conda-forge
pip install typed-argument-parser h5py opencv-python matplotlib annoy seaborn spacy scipy transformers scikit-learn
Note
We are committed to maintaining this repository for the community. We did port this code up to latest versions of PyTorch-Lightning and PyTorch, so there may be small incompatibilities we didn't catch when testing -- please feel free to open an issue if you run into problems, and I will respond within 24 hours. If urgent, please shoot me an email at [email protected]
with "VQA-Outliers Code" in the Subject line and I'll be happy to help!