Weights & biases (I)

One 、 Brief introduction

Weights & Biases It is a machine learning platform for developers to build better models faster . Use W&B Lightweight of 、 Interoperable tools for fast tracking experiments 、 Version and iteration datasets 、 Evaluate model performance 、 Recurrence model 、 Visualization results and discovery regression , And share the findings with colleagues .

stay 5 Set in minutes W&B, Then quickly iterate over your machine learning pipeline , Make sure your data sets and models are tracked and versioned in a reliable recording system .

The main function

• Experimental tracking ： Real time visualization experiment

• Super parameter tuning ： Fast optimization model

• data + Model version control ： Version datasets and models

• Model management ： Manage the model lifecycle from training to production

• Data visualization ： Visual prediction across model versions

• Collaborative reporting ： Describe and share findings

• Integrate ：Py Torch、Keras、Hugging Face etc.

• private - trusteeship ：W&B Private cloud and local hosting of applications

Reference resources

1. App The user interface
2. Python library
3. Command line interface

Two 、 Quick start

Use weights and bias experimental tracking to build better models more effectively .

Try this short Google Colab To see the weights and deviations , No need to install code ！

2.1. Set up wandb

a) stay https://wandb.ai/site Sign up for a free account , Then log in to your wandb account .
b) stay Python 3 Use in the environment pip Install... On your machine wandb library
c) Log in to wandb library . Here you will find your API secret key ：https://wandb.ai/authorize. 

pip install wandb

wandb login

2.2. Start a new run

stay Python Initialize in script or notebook W&B New operation in . wandb.init() Will start tracking system metrics and console logs , Open the box . Run your code , Enter your API secret key , You will see the new run appear in W&B in .
More about wandb.init() →【 The following guidance page 】

import wandb
wandb.init(project="my-awesome-project")

2.3. Tracking indicators

Use wandb.log() Track metrics or framework integration for easy detection .

wandb.log({'accuracy': train_acc, 'loss': train_loss})

2.4. Trace superparameter

Save the super parameters , So that you can quickly compare experiments . 

wandb.config.dropout = 0.2

 2.5. Get alert

If your W&B Run Crashed or whether the custom trigger has been reached （ For example, your lost to Na N Or your ML The steps in the pipeline have been completed ）, Through Slack Or email for notification . For complete setup , Refer to the alert documentation .

1. stay W&B Open alarm in user settings
2. take wandb.alert() Add to your code

wandb.alert(
    title="Low accuracy", 
    text=f"Accuracy {acc} is below the acceptable threshold {thresh}"
)

And then in Slack（ Or your email ） View in W&B Alert message .

common problem

Where can I find my API secret key ？
Sign in www.wandb.ai after ,API The key will be on the authorization page .

How to use in an automated environment W&B？
If you are inconvenient to run shell Training models in an automated environment of commands , for example Google Of Cloud ML, You should check our environment variable configuration guide .

Do you provide local 、 Local installation ？
Yes , You can host locally and privately in your own machine or private cloud W&B, Try this quick tutorial notebook to see how . Be careful , To log in wandb Local server , You can set the host flag to the address of the local instance .

How to temporarily close wandb logging ？
If you are testing code and want to disable wandb Sync , Please set environment variable WANDB _ MODE=offline.