black-eyes-sgc / deepviz

Visualization tools for deep convolutional networks

DeepViz

DeepViz is an interactive tool for visualizing deep convolutional neural networks that perform image classification.

Installation

DeepViz's Python library dependencies are listed in ./webui/requirements.txt; you can install them using pip by running pip install -r requirements.txt. Some of these libraries depend on scipy and numpy, which can take a while to build from their sources. If you want to quickly get up and running on a machine that doesn't have these dependencies installed, it might be easiest to install a Python distribution like Anaconda that includes these libraries, then install the rest of the libraries using pip install -r requirements.txt.

Usage

Sample Datasets

DeepViz requires several input datasets: the image corpus used to train the model, a series of checkpoints of the model during its training, and a set of precomputed statistics about the model's predictions. To get started quickly, you can use these precomputed datasets:

• CIFAR-10 Image Corpus
• Model Checkpoints
• Model Stats DB

Running the Web UI

To start the web UI, run

./webui/runserver.py --model MODEL --cifar CIFAR --model-stats MODEL_STATS

where MODEL is the path to a path to a directory of snapshots of models trained by cuda-convnet, CIFAR is a path to a copy of the CIFAR-10 image corpus, and MODEL_STATS is a database of statistics computed offline on the trained models (see section below).

If everything worked correctly, you should be able to browse to http://localhost:5000 to view the UI.

Training a Model

Computing Model Statistics

The model statistics database contains pre-computed statistics about the model's predictions at each checkpoint. These databases can be generated from a trained model using the build_model_stats_db script:

./webui/deepviz_webui/build_model_stats_db.py --model MODEL --cifar CIFAR --num-classes 10 --output-dir MODEL_STATS