Fix broken Captum tutorial — remove retired Insights section, update requirements

[svekars]

Description

The Model Understanding with Captum tutorial has critical breakage: the Captum Insights section uses APIs that were fully removed from Captum in Feb 2026. The tutorial also has stale dependency requirements.

Changes needed

Critical — broken code

• Remove Captum Insights section (lines ~400-499): AttributionVisualizer, Batch, ImageFeature, and visualizer.render() were deprecated in Captum v0.8.0 (Feb 2025) and fully removed from the codebase (Feb 2026). This entire section will crash on current Captum.
• Remove from captum.insights import AttributionVisualizer, Batch
• Remove from captum.insights.attr_vis.features import ImageFeature

Deprecated APIs

• transforms.ToTensor() → v2.ToImage() + v2.ToDtype(torch.float32, scale=True): ToTensor() is deprecated since torchvision 0.18.

Stale requirements

• Python version: 3.6 → 3.9 or higher (Python 3.6 is EOL since Dec 2021; PyTorch 2.x requires 3.9+)
• PyTorch version: Remove 1.2 or higher floor — it's ancient
• TorchVision version: Remove 0.6 or higher floor
• Flask dependency: Remove entirely — Flask was only needed for Captum Insights
• Matplotlib pin: Remove matplotlib==3.3.4 / matplotlib=3.3.4 pin from both pip and conda install commands. The pin was for a Captum bug that was fixed long ago.
• Conda install: Switch to pip-only (Captum v0.8.0 was the last conda release on the PyTorch channel)

Nice to have

• Consider replacing the removed Insights section with an LLM Prompt Attribution example — this is the headline feature of Captum v0.8.0 and would modernize the tutorial to reflect current AI trends.

Files

• beginner_source/introyt/captumyt.py

cc @subramen

[ozgecinko]

/assigntome

[svekars]

Please assign two issues at a time across all repos.

[gtsitsik]

I think the deprecation warning applies to torchvision.transforms.v2.ToTensor, not to the existing torchvision.transforms.ToTensor used here. Since the tutorial currently imports torchvision.transforms as transforms, this appears to be using the v1 transform API. Let me know if I am missing anything.