This repo contains the frontend code for my graduation project, Design and Implementation of a Visual Analysis System for Loss Function Sampling with Multiple Branches in Optimization Trajectories.

Make sure backend of this repo running.

npm i
npm start

The main layout look like this:

First, you should specify a project to be analyzed in the Project Panel. Then the training trajectories will be shown in the Main view. Every circle in the graph represents a checkpoint which generated on the training process.

Use scroll to zoom the canvas. Hold Shift and mouse drag to pan the canvas.

To select a subset of checkpoint, press and hold mouse left and pass through the checkpoints. The selected checkpoints will be marked red. To use lasso, hold Ctrl then draw the area to be select. To deselect some of selected checkpoints, hold Alt key and do the thing explain above.

You can also use Select All and Unselect All on the toolbar.

The projection button on toolbar will project the selected checkpoints on 2-d plane for main view showing. Hover these button for instruction.

Click the checkpoint to open training menu.

Change the hyperparameter if you want or left them default to the last training. Click Train button to start training.

Disturb magnitude is for disturbing only. When Disturb clicked, a random vector multiply by this constant will be added to the checkpoint you selected.

Navigate to http://localhost:40000/docs#/default/__newproj_post and fill the initial parameter.

We investigated a wide range of different architecture for image classification on CIFAR10 dataset. They are:

• LeNet
• CNN-12, CNN-24, CNN-36, CNN-48, CNN-96, CNN-48x2, CNN-48x3
• VGG-9, VGG-16
• ResNet-20, ResNet-56, ResNet-110, and the no-short-connection version of these
• DenseNet-121
• EfficientNet-s

We generated 3 models for every architecture and trained them over the same preset:

• batch size: 128
• momentum: 0.9
• weight decay: 0.0005
• optimizer: sgd
• learning rate: 0.1 for first 150 epoches, then 0.01 for 151225, 0.001 for 226275, 0.0001 for 276~300
• data augmentation: RandomCrop and RandomHorizontalFlip

The file size of each arch is shown below:

We plot these trajectories using PCA:

Train loss:

We plot some of these trajectories using t-SNE (we cannot plot others due to memory limit):

t-SNE arguments:

• perplexity: 10
• iteration: 3000

arch	plot
CNN-12
CNN-24
CNN-36
CNN-48x2
CNN-48x3
CNN-48x3

We projected these trajectories to 50d using PCA and then use t-SNE:

An interesting picture of EfficientNet-s (the label above pink line indicates the euclidean distance between the points):

The t-sne projected plot indicated the training trajectory may looks like a triangular frustum.

靠，图好多，哥们摆了，自己去img文件夹看