Captum v0.3.1 #591
Merged
Captum v0.3.1 #591
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Summary: Pull Request resolved: meta-pytorch#528 Allows output to be on a different device than input by moving output difference to input device. Reviewed By: miguelmartin75 Differential Revision: D25001273 fbshipit-source-id: a9b6d8e8bb585d5360c53272a5502f4e8f257459
Summary: I pasted the README into a Google Doc to see if the spelling and grammar check would find anything. It was able to find a few mistakes that I've now corrected. Pull Request resolved: meta-pytorch#542 Reviewed By: vivekmig Differential Revision: D25221538 Pulled By: NarineK fbshipit-source-id: c53358b507c5edd6e2f2cd8e7a147aab4d7dbdaf
Summary: Addding multiple layer support in LayerIntegratedGradients Added two test cases. Both compare output to regular IG. One patches multiple embedding layers, the other patches layers of a new model. Added a new model due to existing models not having layers/sub-modules accepting multiple arguments. Updated documentation. Reviewed By: vivekmig Differential Revision: D25042339 fbshipit-source-id: cdfb6c03040fa9e049697d6b54765d27c4d0287b
…ch#551) Summary: This PR adds the ability to compare multiple models in Captum Insights.  In order to test this, I went through two scenarios. First, I made sure there are no regressions to single model workflows like this: 1. Start the Insights example with `python3 -m captum.insights.example` 2. Ensure that the original functionality is still working and there are no changes, other than the visual changes for column headers Then, I tested comparing multiple models by duplicating the existing example one: 1. Go to `example.py` 2. Duplicate the example model, by changing `models=[model]` to `models=[model, model, model]` 3. Check to make sure that it renders properly, and that selecting different target classes works to properly update the data for each visualization Pull Request resolved: meta-pytorch#551 Reviewed By: edward-io Differential Revision: D25379744 Pulled By: Reubend fbshipit-source-id: 4999c1ef0f18b8f735cd47a890cef413a7c6548e
Summary: Pull Request resolved: meta-pytorch#554 Reviewed By: NarineK Differential Revision: D25418107 Pulled By: edward-io fbshipit-source-id: 960ed22c5f6845ac9fedff8793196660d6fd5529
) Summary: Related issue: meta-pytorch#544 Adds flag `return_html`to `visualize_text()` that allows for the IPython HTML to be returned if the flag is set to `True`. This is useful for cases where users may want to save the output from outside of a notebook etc. Usage looks like: ```python from captum.attr import visualization as viz .... # get attributions and data record html_obj = viz.visualize_text([score_viz]) ``` Pull Request resolved: meta-pytorch#548 Reviewed By: vivekmig Differential Revision: D25424019 Pulled By: bilalsal fbshipit-source-id: 27a90f2775d90cbc848858fe1b439ddc2855cba4
Summary: As also mentioned in meta-pytorch#549, we had conflicting argument names in the API. In this PR we deprecate and rename those argument names. More specifically: 1. `n_samples` in NoiseTunnel is being renamed to `nt_samples` 2. `n_perturbed_samples` in Lime and KernelSHAP are being renamed to `n_samples` in order to remain consistent with Shapely Values (Sampling) Pull Request resolved: meta-pytorch#558 Reviewed By: edward-io Differential Revision: D25514136 Pulled By: NarineK fbshipit-source-id: 142c974da2a8430be234fe4ffc79e36faf2bf8d9
Summary: Pull Request resolved: meta-pytorch#534 Introduces a utility class called `ModelInputWrapper` to wrap over a model in order to treat inputs as separate layers. This does so by mapping each input fed to `forward` using an `Identity` operation. This way if attribute_to_inputs=True or False it should work. Add two tests: - Test whether _foward_layer_eval retrieves the appropriate input values - Compare regular IG with layer IG and layer wrapped inputs Updated tutorial and documentation Reviewed By: NarineK Differential Revision: D25110896 fbshipit-source-id: bb8dd4947ae88e183af94c09cf906f9687fbe8ff
Summary: Fixes GPU test failures on master by upgrading pip version. Pull Request resolved: meta-pytorch#568 Reviewed By: bilalsal Differential Revision: D25686863 Pulled By: vivekmig fbshipit-source-id: c1c860fb0666fb529e1d0b4462acaca5e5cea6b6
Summary: Pull Request resolved: meta-pytorch#570 This makes Lime work appropriately with int / long features; currently input only worked appropriately with float features. Reviewed By: bilalsal Differential Revision: D25693888 fbshipit-source-id: b96477f8c6805f554b324ffadbb00e971c12051f
Summary: Pull Request resolved: meta-pytorch#569 Update assertion to warning when using sklearn below 0.23.0 Reviewed By: bilalsal Differential Revision: D25686004 fbshipit-source-id: c5ae1aec5361716ed866cb8d0b25090b64e83926
Summary: Adding support for batch_size in NoiseTunnel as proposed in: meta-pytorch#497 Pull Request resolved: meta-pytorch#555 Reviewed By: vivekmig Differential Revision: D25700056 Pulled By: NarineK fbshipit-source-id: ea34899035486798b1cf3c49ce850291d1f1e76c
…torch#575) Summary: As I understand, Captum uses Python 3, which means that classes don't need to inherit from object as it done implicitly already. Pull Request resolved: meta-pytorch#575 Reviewed By: vivekmig, bilalsal Differential Revision: D25716435 Pulled By: NarineK fbshipit-source-id: 4983e375dfc81a6c03388b009778f90b9ca34a6d
Summary: Since raised in the issue, meta-pytorch#564, adding a solution about it to FAQ. Pull Request resolved: meta-pytorch#576 Reviewed By: edward-io Differential Revision: D25725280 Pulled By: NarineK fbshipit-source-id: 376fc3c98f4cc742242842bb2a1e8df1828d7b4d
Summary: Adding DLRM tutorial and the KDD presentation slides related to it. The actual model is 2.1GB, which it pretty big. Git allows 100MB maximum. Either we need to override max size or put the model elsewhere. Update ............ Storing the model on aws S3 Pull Request resolved: meta-pytorch#531 Reviewed By: vivekmig Differential Revision: D25730733 Pulled By: NarineK fbshipit-source-id: 318c7b606f3fb98d245d9a381fc92c4e21819209
Summary: - Clean up some JS warnings - remove unpkg - switch to plotlyjs-basic-dist-min Pull Request resolved: meta-pytorch#556 Reviewed By: NarineK Differential Revision: D25506078 Pulled By: edward-io fbshipit-source-id: f133975579c1ab3376f243499aca6520aafbb568
Summary: Pull Request resolved: meta-pytorch#588 Reviewed By: miguelmartin75 Differential Revision: D25920600 Pulled By: bilalsal fbshipit-source-id: 6e99e4da74e3ffb4a3a4cc21d39e1115ef9d7938
Summary: This commit reduces the size of Captum Insights by - Replacing the old graphing library with a more lightweight one - In the standalone app, using compression in the Flask server - In the notebook extension, excluding unused dependencies-of-dependencies  For the standalone app, it reduces the size significantly:  For the notebook extension, there's a similar size reduction of `index.js` from 1090 KB to 449 KB. Testing: I used `titanic.py` to test this change, making sure that the graphs are working as before and that the other functionality is unnafected. Pull Request resolved: meta-pytorch#562 Reviewed By: edward-io Differential Revision: D25628623 Pulled By: Reubend fbshipit-source-id: ef8a0d9ec8c7e0df6955b69dd7a96656defc37e8
Summary: CircleCI conda tests are currently failing due to the missing dependency of flask-compress; this adds flask-compress to conda test setup. Pull Request resolved: meta-pytorch#589 Reviewed By: Reubend Differential Revision: D25962860 Pulled By: vivekmig fbshipit-source-id: e66a79f4598bf9392566f90b874b5cfc8162f90c
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Creating Captum v0.3.1 to patch release with recent minor updates. Includes all changes since v0.3.0 except LRP.