Comments (3)
Hi -- so the issue is that PyTorch doesn't currently have a function that's a nice equivalent of TF's batch_jacobian function. This function basically just helps avoid a bunch of redundant calculations and is nicely parallelized. There's some discussion of this issue here.
By combining the sort of slow, python for-loop dependent implementation suggested by Gsunshine with some of the reshaping suggested here, a PyTorch version of the batch_jacobian could be done pretty easily, although potentially this function might be slow. Once you had this, you'd then replace the part of the PyTorch code in the "interactions" function that manually iterates through each input feature (and assumes a 2D batch x features tensor) with the batch_jacobian function you implemented.
Let me know if that sounds reasonable or like too much of a pain to be interesting! It's also possible that by searching around, you may find that someone else has made more progress on the batch_jacobian function in PyTorch recently (e.g. see discussion here).
from path_explain.
Also, I actually think that probably it would be easy to support higher dimensional input tensors if the user only wanted to calculate all interactions with a single feature, which we currently do with the "interaction_index" argument. In this case, the hessian is the same size as the input because we index into a particular gradient. So this might be an even easier starting place.
from path_explain.
Ok, as an update, we don't yet have a good way to support arbitrary dimension tensors in torch, but I did add an Embedding Explainer that should work with NLP models:
from path_explain.
Related Issues (12)
- output_indices not passed through for torch interactions HOT 2
- Can you help update a version with pytorch example? HOT 2
- Unable to use multiprocessing pool with PathExplainerTF HOT 1
- cannot import name 'EmbeddingExplainerTF' HOT 8
- Mismatch num_samples for Torch and Tensorflow implementation IG HOT 7
- No convergence for IH for larger input strings HOT 2
- Whether the feature attribution method can be applied on training data?
- can't find bert_explainer
- Using Longformer
- .npy files
- EmbeddingExplainerTorch not available in pip package
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 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. 📊📈🎉
-
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 ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from path_explain.