alexrame / fishr Goto Github PK
View Code? Open in Web Editor NEWOfficial PyTorch implementation of the Fishr regularization for out-of-distribution generalization
License: GNU General Public License v3.0
Official PyTorch implementation of the Fishr regularization for out-of-distribution generalization
License: GNU General Public License v3.0
Dear authors,
I have one question about code here:
fishr/coloredmnist/train_coloredmnist.py
Line 184 in 7b8fdf1
Thank you.
Hi,
Since you have used on some of the datasets available in DomainBed benchmark in your experiments, I was wondering whether your code works also on other datasets such as WILDS-FMoW which is available in DomainBed's code?
Thanks.
Sara A. Al-Emadi
Hi @alexrame,
I've been playing around with the Fishr implementation in DomainBed for a while now and going through the paper again today I noticed the following section:
For example, on PACS dataset (7 classes and |ω| = 14, 343) with a ResNet-50 and batch size 32, Fishr induces an overhead in memory of +0.2% and in training time of +2.7% (with a Tesla V100) compared to ERM; on the larger-scale DomainNet (345 classes and |ω| = 706, 905), the overhead is +7.0% in memory and +6.5% in training time.
Unfortunately, I'm not noticing these kind of percentages for the overhead. Rather, Fishr is about twice as slow for me as a similar ERM model. Now this could be due to a variety of reasons. For example, my model is quite small because of which the overhead might play a relatively lager role: I simply use an MLP with layer of size [166, 1024, 256, 64, 256, 64, 1]
. The backwards pass with Backpack is only for the last layer (so I believe |ω| = 64?). My batch size is 64.
I am writing to you to verify that the numbers from the paper were achieved with this implementation. Is that indeed the case? If so, do you think that in my case a larger overhead is to be expected? This would help me to narrow down a possible problem. I would greatly appreciate your response.
Hi,
When I tried downloading the DomainBed dataset through the code provided, it results in the following error:
Access denied with the following error: Cannot retrieve the public link of the file. You may need to change the permission to 'Anyone with the link', or have had many accesses. You may still be able to access the file from the browser: https://drive.google.com/uc?id=0B6x7gtvErXgfbF9CSk53UkRxVzg
I tried accessing the folder through the browser but I get an access denied error. Therefore, would appericate it if you could kindly provide the access to the download link.
Thanks.
Sara A. Al-Emadi
Hi Alex,
Thank you for your amazing work.
Can you provide the results of Fishr for leave-one-domain-out model selection for domainbed benchmark?
Thank you in advance
Dear authors,
I wanna run Fishr on ColoredMNIST in domainbed with mini-batch fashion. Must I conduct a hyperparameter sweep? I used the hyperparameter in coloredmnist/train_coloredmnist.py (i.e., \lambda=91257) and run domainbed.scripts.train.py. I found that fishr cannot achieve reasonable results, i.e., testing env with 0.1 accuracy. How to run fishr on ColoredMNIST with mini-batch? Thanks in advance.
Best,
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.