picture
Source: Deep Learning Enthusiasts
This article is about 700 words , it is recommended to read 5 minutes
This article introduces a number of tools that can visualize the deep learning training process to help you better understand deep learning , which is very practical.


The deep learning training process has always been in a black box state. Many students asked me how to explain it? In fact, many of them are still unexplainable, but through visualization, we can know exactly what features are learned by deep learning during the training process? What characteristics of the target are you interested in? We already have many channels to know about these, so let me introduce you to a few better tools!


picture


1. Deep learning network structure drawing tool


Address: https://cbovar.github.io/ConvNetDraw/


picture


2.caffe visualization tool


Input: caffe configuration file Output: network structure


Address: http://ethereon.github.io/netscope/#/editor

picture

3. Visual DL, a deep learning visualization tool


Visual DL is developed by Baidu, based on echar and PaddlePaddle, and supports mainstream frameworks such as PaddlePaddle, PyTorch and MXNet. ps: This is my favorite. After all, the rendering ability of echar is good hahaha, but unfortunately it does not support caffe and tensorflow.


Address: https://github.com/PaddlePaddle/VisualDL

4. Structure visualization tool PlotNeuralNet


Developed by a student in Computer Science at Saarland University.

Address: https://github.com/HarisIqbal88/PlotNeuralNet

In fact, there are many visualization tools, but what I want to say today is that the visualization of the training process is similar to the visualization of TF, but this operation is easier!

picture

How detailed does this tool really show the training process? In short, the author of the project has prepared an interactive interface for you, you just need to open the browser and load the interface. CNN Explainer uses TensorFlow.js to load pre-trained models for visualization, and Svelte as the framework for interaction and D3.js for visualization. The final product has no threshold for use even for beginners who are completely ignorant. Let's take a look at the specific effect.

picture


convolution


picture

picture

picture


Hyperparameters


picture

softmax

picture

picture

ReLU


picture

MaxPool


picture

picture

Through the whole process, everyone must have a detailed understanding of the process. If you are skilled, you can directly visualize the training process through the deep learning platform. The process must be more detailed than this.


Editor: Huang Jiyan

Proofreading: Lin Yilin


picture

picture