hendrikstrobelt/confusionflow

ConfusionFlow is a visualization tool that enables more nuanced monitoring of a neural network's training process.

ConfusionFlow is a visualization tool distributed as Python package that enables more nuanced monitoring of a neural network's training process.

• track and visualize the model performance of different timepoints (i.e., epochs) and dataset folds and compare different runs (e.g., different hyperparameter configurations).
• we provide wrappers for exporting performance logs in the required format.

ConfusionFlow is in an early-preview alpha. Expect some bugs and rough edges.

Additional Information

ConfusionFlow was developed as a visualization tool to provide users with more feedback while developing or tuning neural network based classifiers. Performance monitoring often only utilizes simple line charts (one would for example plot the model loss and accuracy) which might easily miss many details and changes in the model's error structure.

While the errors for a specific model state (e.g., at a certain epoch) can be represented as a confusion matrix, it is difficult to compare multiple confusion matrices or track changes over time. ConfusionFlow visualizes model confusion over multiple model states and let users compare different folds (e.g., train vs test set) or different hyperparameter configurations.

To the best of our knowledge there are still no other tools with similar functionality.

Limitations

Number of classes

Due to screenspace limitations the system is currently limited to around 10 classes. We are aware that we will not be able to handle datasets at ImageNet scale BUT those datesets are not very common, as they are usually very expensive to obtain. A large percentage of classification problems ranges around 10 or less classes where ConfusionFlow can provide additional feedback.

Runtime overhead

Logging the performance for multiple folds every epoch might severely slowdown the time to convergence (especially when creating logs for the complete train and test sets). While logging on a minibatch level might provide additional information, it also severely slows down the training even further. It is possible to alleviate this problem by persisting model checkpoints and creating the performance logs on different machines. However, this also involves a lot of engineering effort and will not be supported by the ConfusionFlow in the near future.

Installation

Confusionflow can be either downloaded directly from PyPI via pip install confusionflow or by cloning and installing the repository directly from source:

From Source

Make sure you have node installed (required for building the UI component).

Clone and install the repository:

git clone https://github.com/confusionflow/confusionflow
cd confusionflow
python setup.py install

Getting started

As a first step you must create some logs before you can start analyzing. Have a look at the examples folder and run one of the demos. The examples will create a new subdirectory logs where the performance logs will be stored.

You then can start the ConfusionFlow UI via:

confusionflow --logdir `<path_to_logdir>`

Usage Example

## tensorflow.keras
cd examples/tf.keras/fashion-mnist
python fashion-mnist_demo.py
confusionflow --logdir logs

## pytorch
cd examples/torch/mnist
python mnist_demo.py
confusionflow --logdir logs

If you are using your own datasets you must create a dataset-configuration first. We provide some example configurations for some popular datasets in examples/dataset-templates which should help you getting started.

Logging

We provide simple wrappers for Tensorflow + Keras and Pytorch for logging confusion matrices and exporting them in the required ConfusionFlow format.

Log Directory Layout

<logdir>
├── datasets  		            <--- dataset config files
│   ├── mnist.json
│   └── index.yml
├── foldlogdata                 <--- foldlog data
│   ├── example_log_mnist_train_data.json
│   └── example_log_mnist_test_data.json
├── foldlog                     <--- foldlog specifications
│   ├── example_log_mnist_train.json
│   └── example_log_mnist_test.json
├── runs                        <--- run specifications
│   ├── example_log.json
│   └── index.json
└── views                       <--- view specifications (currently unsused)

Documentation

The latest documentation can be viewed on docs.confusionflow.org

API

The Python package includes a simple Flask based server that implements the current API. The current API definition can be found in tools/swagger/api.yml and can be viewed using the Swagger Editor.

ConfusionFlow UI

The ConfusionFlow UI is currently developed at Caleydo/confusionflow-ui.

The Team

ConfusionFlow is a research project of the Institute of Computer Graphics at Johannes Kepler University Linz in collaboration with the IBM Visual AI Lab. ConfusionFlow is currently maintained by Peter Ruch and Holger Stitz.

Feedback

We would be really grateful for any feedback via the repository's issues section.