NETRON — A model visualization tool

Model summary option in TensorFlow is quite handy for understanding the overall architecture and input transformation through the network. It also provides information like output feature map size and no of trainable parameters in each layer of the network.

Model summary of small Convolutional Neural Network

But modern day neural network architectures like Resnet and Inception Net are quite complex (Even more complex models for Object Detection and Image segmentation tasks), model.summary() is not best suited for analyzing such complex model architecture.

NETRON provides a very easy to use graphical user interface for visualizing machine learning models by loading the complete machine model(Ex: .h5 file)or just its configuration(Ex: .json file).

GitHub — lutzroeder/netron: Visualizer for neural network, deep learning, and machine learning models

example visualization of Inception V3 model (only a part)

The following code snippet can be used for generating a json file(faster than saving complete model .h5 format) with the model configuration information of pretrained InceptionV3 models available tf.keras.applications(same can be done for a user defined model also)

Similarly the entire model can be saved to .h5 with following code snippet

Just to get a feel of NETRON, have a look of web API based representation of YOLO architecture using following link without having a local installation(which I recommend to have) in your PC

yolo.cfg (netron.app)