NotionNext BLOG https://tangly1024.com/ 这是一个由NotionNext生成的站点 Sun, 23 Jul 2023 05:00:01 GMT https://validator.w3.org/feed/docs/rss2.html https://github.com/jpmonette/feed zh-CN All rights reserved 2023, NotionNext <![CDATA[How to use EarlyStopping]]> https://tangly1024.com/article/early_stopping https://tangly1024.com/article/early_stopping Fri, 21 Jul 2023 00:00:00 GMT

Why

  1. Avoid over-fitting.
  1. Save time & gpu memory.

Implementation

github.com
https://github.com/Lightning-AI/lightning/blob/9a4e8a8c528d6475ab33c46a0a84e27273cc10bd/src/pytorch_lightning/callbacks/early_stopping.py#L38
github.com
https://github.com/Bjarten/early-stopping-pytorch/blob/master/MNIST_Early_Stopping_example.ipynb
EarlyStoppingHook in mmengine:
github.com
https://github.com/open-mmlab/mmengine/blob/main/mmengine/hooks/early_stopping_hook.py

Sample Code

EarlyStopping:

Use in yolo-v5

Effectiveness for object detection

 
]]> <![CDATA[Cluster and visualize connected graph by k-means and networkx]]> https://tangly1024.com/article/a5c6d495-4fa9-4aad-a501-959046b23146 https://tangly1024.com/article/a5c6d495-4fa9-4aad-a501-959046b23146 Fri, 21 Jul 2023 00:00:00 GMT

1. Get distance matrix from adjacent matrix M using floyd warshall algorithm.

 

2. Compute N clusters from distance matrix using k-means.

🤔
What if we need to specify number of nodes for each cluster?
 

3. Visualize graph using networkx.

💡
To make edges in the visualized graph scaled to their actual values in the adjacency matrix, we need to use nx.draw_kamada_kawai to draw the graph.
Visualized graph:
notion image
]]>