This repository contains the code for the paper "On the Effects of Self-supervision and Contrastive Alignment in Deep Multi-view Clustering", CVPR 2023.
Abstract - Recent works in deep multi-view clustering (MVC) vary significantly in the motivation, implementation and evaluation of new methods. This makes it difficult to reproduce and compare the performance of models, and to identify directions for future work. Here we illustrate that many recent methods can be regarded as instances of a new unified framework, referred to as DeepMVC. This provides novel insight into the field, allowing us to develop new instances of DeepMVC in a principled manner. We conduct an extensive experimental evaluation of recent methods and our new instances, showing that these outperform previous methods on several datasets. Finally, we present key findings from our experiments, and suggest promising directions for future research. To enhance the openness of the field, we provide an open-source implementation of DeepMVC, including the baselines and the new instances. Our implementation also includes a consistent evaluation protocol, facilitating fair and accurate evaluation of methods.
Figure 1: Overview of the DeepMVC framework and its components.
Table 1: Main results from our experimental evaluation.
This repository contains the Caltech7 and Caltech20 datasets. NoisyMNIST, NoisyFashion, EdgeMNIST, EdgeFashion, and PatchedMNIST can be generated without additional downloads by running
python -m data.make_dataset noisymnist noisyfashionmnist edgemnist edgefashionmnist patchedmnist
from the src
directory.
The COIL-20 dataset is generated by first downloading the original images from
here,
and placing them in data/raw/coil20
.
The following command will then generate the training-ready dataset
python -m data.make_dataset coil20
Pre-training the Siamese network for MvSCN requires a version of the dataset where samples are grouped together in positive and negative pairs. The paired datasets are generated by running
python -m data.pair_dataset -d <dataset_name>
where <dataset_name>
is one of noisymnist
, noisyfashionmnist
, edgemnist
, edgefashionmnist
, patchedmnist
, caltech7
, caltech20
, coil20
.
Experiments are run with the command
python train.py -c <experiment_config>
where <experiment_config>
is the name of a config defined in a file in src/config/experiments
.
See Creating new experiments for further details on the Config objects.
Directories containing experiments configs are
- Main evaluation:
src/config/experiments/benchmark
- Ablation studies:
src/config/experiments/ablation
- Increasing number of views (to reproduce Figure 2 in the paper):
src/config/experiments/increasing_n_views
Note that some new instances have names in this implementation, compared to in the paper:
Name in paper | Name in implementation | Name in experiment config |
---|---|---|
AE-KM | SAEKM | saekm |
AE-DDC | SAE | sae |
AECoKM | CAEKM | caekm |
AECoDDC | CAE | cae |
InfoDDC | MIMVC | mimvc |
train.py
logs the experiment results to a Weights and Biases project.
The project name is specified by setting WANDB_PROJECT="<project_name>"
in src/lib/wandb_utils.py
(set to "mvc-framework"
by default.)
A docker image with all dependencies can be built by running
docker build \
--tag <image_tag> \
--rm \
--no-cache \
--pull \
-f docker/Dockerfile \
.
from the root directory of the repository.
Most config parameters can be overridden by specifying them at the command line. For instance, to run CoMVC on NoisyMNIST with batch size 256 instead of 100, run
python train.py -c noisymnist_comvc --batch_size 256
Note that the config hierarchy (described below) is "traversed" on the command line using /
, so running CoMVC on NoisyMNIST without the adaptive weight on the contrastive loss is done with
python train.py -c noisymnist --model_config/loss_config/contrast_adaptive_weight False
and with learning rate 1e-4
python train.py -c noisymnist --model_config/optimizer_config/learning_rate 0.0001
The Config
class
All experiments run with src/train.py
are instances of Experiment
, a subclass of the generic Config
class.
Configs are implemented as hierarchical dataclass-like objects (based on pydantic's BaseModel
), where Experiment
is the top-level class.
For instance, instances of Experiment
are required to specify the dataset_config
attribute as an instance of Dataset
, containing the name and parameters of the dataset to use for the experiment.
All Config
subclasses used in the experiments are defined in src/config/templates
.
This directory includes Config
subclasses for datasets, models, encoders, fusion, clustering modules, etc.
Defining a custom CoMVC experiment on NoisyMNIST from scratch is done with
from config.templates.experiment import Experiment
from config.templates.dataset import Dataset
from config.templates.models.simvc_comvc import CoMVC, CoMVCLoss
from config.templates.encoder import Encoder
from config.templates.optimizer import Optimizer
noisymnist_comvc_custom = Experiment(
dataset_config=Dataset(
name="noisymnist",
n_train_samples=20000, # Reduce the number of samples to train quicker.
),
model_config=CoMVC(
# Ene encoder per view
encoder_configs=[
Encoder(layers="cnn_small", input_size=[1, 28, 28]),
Encoder(layers="cnn_small", input_size=[1, 28, 28]),
],
loss_config=CoMVCLoss(
tau=0.07 # Change the softmax temperature in the contrastive loss.
),
optimizer_config=Optimizer(
learning_rate=1e-4 # Change the learning rate
),
),
n_clusters=10, # The number of clusters is specified in the Experiment config
batch_size=256, # so is the batch size
n_epochs=200, # and the number of training epochs
)
This file should be imported in src/config/experiments/__init__.py
.
The custom experiment can then be run with
python train.py -c noisymnist_comvc_custom