individualcalibration's People
individualcalibration's Issues
From our discussion at ICML
Hi! This is the code that I mentioned sharing during ICML. I think the line here https://github.com/ShengjiaZhao/IndividualCalibration/blob/master/utils.py#L69 could be improved by first taking an individual calibration error for each instance first.
Here is a small snippet of a testing loop, and how I implemented it after reading your paper...
SAMPLES = 2000
b = 100
for i, (x, y, _) in enumerate(args.test_set):
x, y = x.to(args.device), y.to(args.device)
mus, logvars = model.mc(x, SAMPLES)
mus = mus.squeeze()
logvars = logvars.squeeze()
cdf = 0.5 * (
1.0 + torch.erf((y - mus) / torch.exp(logvars / 2) / math.sqrt(2))
)
bins = torch.linspace(1.0 / b, 1.0, b).to(args.device) # expectation of U
cdf = cdf.repeat(b, 1, 1) # repeat it b times so we can test each sample at each bin
# the next line is weird, but [...] it is like a double unsqueeze
cal = cdf <= bins[(...,) + (None,) * 2] # how many values are below bin percentile
cal = cal.sum(dim=1)
cal = cal / float(SAMPLES) # empirical average number of samples below bin percentile
# at this point we have a (b, x_inst) tensor of the empirical average of the number of CDF samples which
# fall into each bin, we can take the calibration error of each sample to find the worst subset, and then
# also get the average calibration error of that worst subset since we already have the calibration error
# of each instance.
cal_err += (torch.abs(cal - bins.unsqueeze(1))).sum(dim=1) # sum calibration error over the sample dimension
for my reimplementation purposes, I was not concerned with getting the worst subset, but you could do that right after getting the empirical average of samples below the bin percentile step (second to last non-comment line).
Also, if you did it this way, you could avoid taking only the last sample in this line https://github.com/ShengjiaZhao/IndividualCalibration/blob/master/utils.py#L73 and instead measure the mean calibration error of the worst subset which were calculated over all samples.
Thanks for answering my questions! Please let me know what you think
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
๐ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. ๐๐๐
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google โค๏ธ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.