In situ TensorView: In situ Visualization of Convolutional Neural Networks

Convolutional Neural Networks(CNNs) are complex systems trained to recognize images, texts and more. However, once trained, they are regarded as black-boxes that are not easy to analyze and understand. Visualizing the dynamics within such deep artificial neural networks can provide a better understanding of how they are learning and making predictions. In the field of scientific simulations, visualization tools like Paraview have long been utilized to provide insights. We present in situ TensorView to visualize the training and functioning of CNNs as if they are systems of scientific simulations. In situ TensorView is a loosely coupled in situ visualization open framework that provides multiple viewers with the ability to visualize and understand their networks. It leverages the capability of co-processing from Paraview to provide real-time visualization during training and predicting phases, and avoids heavy I/O overhead. Tensorview is easily coupled with Tensorflow, as it only requires the insertion of a few lines of code into a TensorFlow framework. In this work, we showcase visualizing LeNet-5 and VGG16 using in situ TensorView. With the insight provided by Tensorview, users can adjust network architectures, or compress pre-trained networks guided by visualization results.