Captum
Click to visit website
About
Captum is an open-source model interpretability library built on PyTorch, designed to help users understand the predictions of their AI models. It supports interpretability across various modalities, including vision and text, and integrates seamlessly with most PyTorch models with minimal modifications. Captum offers a generic and extensible framework for interpretability research, making it easy to implement and benchmark new algorithms. It provides a range of attribution methods to explain how important each input value is to a particular output scalar value, aiding in debugging and improving model understanding.
Platform
Features
• built on pytorch
• multi-modal support
• extensible for research
FAQs
How do I set the target parameter to an attribution method?
Target selects a scalar output value. For 2D output (N x classes), pass an integer (0,1,2) or tensor ([0,1,0,0]). For >2D, pass a tuple of indices (2,3,2) or a list of tuples.
I am facing Out-Of-Memory (OOM) errors when using Captum. How do I resolve this?
Reduce `n_steps` or use `internal_batch_size` for methods with this arg. For perturbation methods, reduce `perturbations_per_eval`. For others, try smaller input batch sizes.
I am using a perturbation based method, and attributions are taking too long to compute. How can I speed it up?
Increase `perturbations_per_eval` if memory allows. You can also use PyTorch's `DataParallel` or `DistributedDataParallel` if multiple GPUs are available to speed up computations.
Are SmoothGrad or VarGrad supported in Captum?
Yes, SmoothGrad and VarGrad are supported through `NoiseTunnel` in Captum. This functionality can be used with any attribution algorithm available within the Captum library.
How do I use Captum with BERT models?
Captum provides a dedicated tutorial demonstrating the usage of Integrated Gradients specifically for BERT models. Refer to the documentation's "Bert SQUAD Interpret" tutorial for details.
My model inputs or outputs token indices, and when using Captum I see errors relating to gradients, how do I resolve this?
Replace embedding layer with `InterpretableEmbedding` or use `LayerIntegratedGradients` for input token indices. For output token indices, attribute with respect to token score/probability.
Can my model use functional non-linearities (E.g. nn.functional.ReLU) or can reused modules be used with Captum?
Most methods work, but DeepLift-based/Guided Backprop/Deconvolution require `torch.nn.ReLU` modules, not functional ones, and don't reuse modules to avoid issues with multipliers.
Do JIT models, DataParallel models, or DistributedDataParallel models work with Captum?
Yes, all are supported. JIT models do not support hooks, so layer/neuron attribution, DeepLift, Guided Backprop, and Deconvolution are not supported with JIT models.
I am working on a new interpretability or attribution method and would like to add it to Captum. How do I proceed?
Captum offers an "Awesome List" for external projects and considers new algorithms for its `contrib` package. Proposals are reviewed case-by-case based on publication, evaluation, etc.
How can I resolve cudnn RNN backward error for RNN or LSTM network?
To resolve "cudnn RNN backward can only be called in training mode" in eval mode for RNN/LSTM, set `torch.backends.cudnn.enabled=False`. CuDNN doesn't support gradient computation in eval mode.
Job Opportunities
There are currently no job postings for this AI tool.
Ratings & Reviews
No ratings available yet. Be the first to rate this tool!
Alternatives
Interpretable AI
Interpretable AI is a suite of software modules providing comprehensive tools for interpretable machine learning, including optimal trees, imputation, and feature selection.
View DetailsFeatured Tools
adly.news
adly.news is a free platform that simplifies newsletter advertising, connecting businesses with engaged audiences through ad slots, offering bidding, negotiation, and messaging.
View Details
Questie.ai
Questie.ai is an advanced AI gaming companion that watches your actual gameplay in real-time and provides intelligent commentary through natural AI voice chat.
View Details
Gemini Watermark Remover
Gemini Watermark Remover is a client-side tool designed to remove hidden SynthID and other embedded watermarks from your AI-generated images, preserving quality.
View Details
Infatuated.AI
Infatuated.AI is an AI companion platform allowing users to chat, roleplay, and build personalized relationships with AI girlfriends and boyfriends, offering emotional support and secure fantasy sharing.
View Details
ImgGen
ImgGen is the free AI editor that edits photos and turns images into videos in seconds, offering instant creativity all in one place.
View DetailsNano Banana
Nano Banana is a state-of-the-art AI model that revolutionizes text-based image editing and generation with unmatched multi-image fusion and natural language understanding.
View Details
Macaron
Macaron is the world’s first personal AI agent designed to help you live better by focusing on happiness, health, and freedom, unlike typical productivity tools.
View Details
VISBOOM
Visboom is the all-in-one AI fashion content creation platform, enabling brands and e-commerce sellers to generate on-model photoshoots and visual assets quickly.
View Details
Banana AI
Banana AI is an advanced AI photo editor powered by Google’s Nano Banana technology (Gemini 2.5 Flash Image), enabling effortless image editing, restyling, and transformation with simple text prompts.
View Details
twainGPT
twainGPT is a humanizer that transforms any AI-generated text into undetectable, human-like content, trusted by over 2.3 million users.
View Details