W&B Smoothing Features

Customize line plot smoothing to better visualize noisy data
Created on March 16|Last edited on March 16
Comment
Here's a quick overview of how smoothing works on line plots. See details in the docs.
﻿Unsmoothed line﻿
﻿Exponential moving average (default)
﻿Gaussian smoothing﻿
﻿Running average﻿
﻿Try it yourself﻿﻿﻿
﻿﻿﻿﻿﻿
1. Unsmoothed lineHere's our starting point: an unsmoothed timeseries logged with wandb.log() over a series of steps.
﻿
Run: magic-morning-171
﻿
﻿
2. Exponential moving averageHere's what that chart looks like with exponential moving average smoothing. This is the default type of smoothing on line plots. I duplicated the chart above, clicked edit, and set the default smoothing to 0.5. 
﻿
Run: magic-morning-171
﻿
Original line in the backgroundYou can see a very faint line in the background. This shows the original data, before smoothing. Toggle Show Original off in settings to hide this ghosted line.
﻿
﻿
﻿
DetailsExponential weighted average is calculated by taking the weighted average of the terms coming before a value where the weights correspond to an exponential decay function. 
The basic formula for this is St=αYt+(1−α)Yt−1S_t=\alpha Y_t +(1-\alpha)Y_{t-1}St​=αYt​+(1−α)Yt−1​﻿﻿
We also add a debiasing term to make our calculation identical to tensorboard without which points close to the beginning would be closer to zero, so the actual formula we use is:
﻿St=(αYt+(1−α)Yt)(1−αt)S_t=\frac{(\alpha Y_t + (1-\alpha) Y_t)}{(1-\alpha^t)}St​=(1−αt)(αYt​+(1−α)Yt​)​﻿﻿
One downside of an exponential weighted average is that it only considers the points to the left of a given point, so a smooth line when smoothed will have an artificially decreased slope.
3. Gaussian smoothingHere's the same line plot with gaussian smoothing. I duplicated the panel above, clicked edit and set the smoothing type to Gaussian, and the smoothing factor to 1.
﻿
Run: magic-morning-171
﻿
﻿
4. Running averageHere's the same line plot with running average smoothing. I duplicated the panel, clicked edit and set the smoothing type to Running Average, and smoothing factor to 5.
﻿
Run: magic-morning-171
﻿
﻿
﻿
5. Try it yourself!Open a chart in your own project and try tweaking the smoothing options.
﻿
﻿
Need sample data to play with? Here's what I ran in Google Colab to make this example:
!pip install wandb -qqq
Then run this cell to log some noisy sample data, which will create a line plot in W&B:
import wandb
import math
import random
# Start a run that logs a sample metric at 100 steps
run = wandb.init(project="smoothing-example")
for step in range(100):
  wandb.log({
      "sample_metric": 0.1 * (math.log(1 + step + random.random()) + random.random())
  })
# Mark that run finished - useful in a Jupyter context
wandb.finish()
﻿
Add a comment