Update OV-XAI #2394
Update OV-XAI #2394
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:20Z Line #2. data=input_image, we don't want to introduce "Overlay is applied over the image in negvet commented on 2024-09-18T13:07:16Z Right, good point, will update |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:20Z Line #5. overlay=True, # optional, saliency map overlay over the input image, defaults to False saliency map overlays over the input image, default to False |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:22Z Line #3. targets=retriever_class_index, # can be a single target or a container of targets list of targets negvet commented on 2024-09-18T13:06:50Z not only list, tuple or numpy.array also would work here. That is why I use general term - container |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:23Z Line #4. overlay=True, # Saliency map overlay over the original image, False by default, set to True for better visual inspection As I see in this notebook, the inline comments start with small letters (not capital). Can you please check the consistency here and further? negvet commented on 2024-09-18T13:09:11Z Sure, will fix |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:23Z Line #8. # Visualize generated saliency maps for each target class (.plot() supports plotting multiple saliency maps)
|
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:24Z Line #4. print(f"Saliency maps were generated for {len(explanation.targets)} classes: ")
Now I can relate to your comment that output cells are too big. Let's just leave a number of classes for which saliency maps were generated. negvet commented on 2024-09-18T13:14:17Z Will fix |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:25Z Line #13. explanation.save(output) What do you think about simplifying it?: explanation.save(output, "grayscale_") negvet commented on 2024-09-18T13:23:42Z I like this idea. Adopted |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:26Z Line #7. target_layer="MobilenetV3/Conv_1/Conv2D", # Optional, by default insert_xai will try to find target_layer automatically
|
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:26Z Can you fix double spaces between "mode" and "treats" in the first sentence? |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:27Z AISE is used as a default black-box method. AISE formulates saliency map generation as a kernel density estimation (KDE) problem, and adaptively samples the input masks using a derivative-free optimizer to maximize the mask saliency score |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:28Z Please, leave a link to RISE paper negvet commented on 2024-09-18T13:25:57Z there is no link as for now GalyaZalesskaya commented on 2024-09-20T17:57:34Z Are we talking about the same RISE paper? [1806.07421] RISE: Randomized Input Sampling for Explanation of Black-box Models (arxiv.org) negvet commented on 2024-09-23T07:22:29Z Ah, sure, I misunderstood |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:29Z Line #14. explanation.save(output, f"{Path(image_path).stem}_") # pass prefix name with underscore
|
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-17T16:55:30Z Block from L27-L37 can be removed by our great new
negvet commented on 2024-09-18T14:21:13Z Exactly, will use explanation.save() |
| AISE | ||
| AISEClassification | ||
| AISEDetection | ||
| WHITEBOX | ||
| BLACKBOX | ||
| Netron | ||
| KDE |
There was a problem hiding this comment.
Shouldn't these names be organized alphabetically for the simplicity of adding new words?
| @@ -25,22 +25,25 @@ Example: Saliency map for `flat-coated retriever` class for MobileNetV3 classifi | |||
|
|
|||
| The tutorial consists of the following steps: | |||
|
|
|||
| - Run explainer in Auto-mode | |||
| - Run explainer in `Auto` mode | |||
There was a problem hiding this comment.
In L11-L12 here we talk only about IR models:
[OpenVINO™ Explainable AI (XAI)](https://github.com/openvinotoolkit/openvino_xai/) provides a suite of XAI algorithms for visual explanation of
[OpenVINO™](https://github.com/openvinotoolkit/openvino) Intermediate Representation (IR) models.
But now thanks to Songki, we have an option to insert_xai branch to PyTorch as well , even thought not to create Explainer using it.
What do you think, should we mention PyTorch usecase in readmes?
There was a problem hiding this comment.
I mentioned PyTorch in the insert_xai section
|
not only list, tuple or numpy.array also would work here. That is why I use general term - container View entire conversation on ReviewNB |
|
Right, good point, will update View entire conversation on ReviewNB |
|
Sure, will fix View entire conversation on ReviewNB |
|
Will fix View entire conversation on ReviewNB |
|
I like this idea. Adopted View entire conversation on ReviewNB |
|
there is no link as for now View entire conversation on ReviewNB |
|
Exactly, will use explanation.save() View entire conversation on ReviewNB |
|
Are we talking about the same RISE paper? [1806.07421] RISE: Randomized Input Sampling for Explanation of Black-box Models (arxiv.org) View entire conversation on ReviewNB |
|
View / edit / reply to this conversation on ReviewNB GalyaZalesskaya commented on 2024-09-20T18:03:19Z saliency_map_name_prefix = f"{image_name}_{gt_info}_pr_"
saliency_map_name_postfix = "_"
More details to names can be added (but not necessary) negvet commented on 2024-09-23T07:27:03Z Added |
|
Ah, sure, I misunderstood View entire conversation on ReviewNB |
|
Added View entire conversation on ReviewNB |
Align OV-XAI notebook with 1.1.0 release