frgfm/torch-cam

TorchCAM: class activation explorer

English | 简体中文

Explain PyTorch decisions: activation maps for image classifiers, detectors, segmenters and VLMs, and token contributions for text classifiers.

Debugging one surprising classifier result? Use the predicted-versus-expected agent workflow, install the portable skill in your coding agent with npx skills add frgfm/torch-cam, or start from llms.txt.

Source: image from woopets (activation maps created with a pretrained Resnet-18)

Why TorchCAM

• 12 CAM methods plus VLM token maps: from CAM and Grad-CAM to LeGrad, RefineCAM and TAM.
• CNNs and Vision Transformers: automatic target-layer resolution for CNNs, LeGrad and reshape transforms for ViTs.
• Beyond classification: explain detections, segmentation masks and embeddings with the same extractors.
• Text decision models: fast TokenAblation measures each token’s effect; TokenIG provides embedding Integrated Gradients with numerical checks. Both return class-specific signed maps, including for Laya’s decision architecture.
• Lean: fully typed, with only 4 runtime dependencies (PyTorch, NumPy, Pillow, Matplotlib).
• Measurable: built-in faithfulness metrics (average drop, increase in confidence, deletion/insertion).
• Agent-ready: explain() saves a manifest-backed evidence bundle that humans and coding agents can verify.

Quick Tour

Explain a prediction

from urllib.request import urlretrieve
from PIL import Image
from torchvision.models import ResNet18_Weights, resnet18
from torchvision.transforms.functional import center_crop, resize
from torchcam.explain import explain

urlretrieve("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
image = Image.open("dog.jpg").convert("RGB")
weights = ResNet18_Weights.DEFAULT
model = resnet18(weights=weights).eval()

preprocess = weights.transforms()
input_tensor = preprocess(image).unsqueeze(0)
model_image = center_crop(
    resize(image, preprocess.resize_size, interpolation=preprocess.interpolation, antialias=preprocess.antialias),
    preprocess.crop_size,
)
result = explain(model, input_tensor, class_names=weights.meta["categories"])
result.save("torchcam-explanation", model_image)  # CAMs, heatmaps, overlays and manifest.json

Pass expected_class_idx to compare the prediction with the class you expected. See Debug one prediction for the full contract.

Explain a text decision

Use your existing classifier and tokenizer; TorchCAM adds no Transformers or Laya dependency. TokenAblation and TokenIG are available in the development version (install from source below).

from torchcam.methods import TokenAblation

# model and tokenizer are your loaded text classifier and its tokenizer.
inputs = tokenizer("I was charged twice. Please refund the duplicate payment.", return_tensors="pt", return_special_tokens_mask=True)
inputs = {key: value.to(next(model.parameters()).device) for key, value in inputs.items()}
token_mask = inputs["attention_mask"].bool() & ~inputs.pop("special_tokens_mask").bool()
extractor = TokenAblation(model, output_transform=lambda output: output.logits)
result = extractor(inputs, replacement_id=tokenizer.mask_token_id, token_mask=token_mask)

probabilities = result.scores.softmax(-1)  # Use the model's calibration instead when applicable.
contributions = result.attributions      # (batch, classes, input_tokens), signed class-logit changes

Positive effects mean replacing a token lowers that class score; other tokens stay fixed. Effects can interact and need not add up. Use a replacement ID appropriate to the model, and exclude special/context tokens with token_mask. For additive embedding-path contributions, use TokenIG. Token effects are not probabilities; highlight positive values beside each decision’s probability for a simple CAM-style display. See Text decision models for masks, baselines, efficiency, the result contract, and Laya integration.

Setting your CAM

TorchCAM leverages PyTorch hooking mechanisms to seamlessly retrieve all required information to produce the class activation without additional efforts from the user. Each CAM object acts as a wrapper around your model.

You can find the exhaustive list of supported CAM methods in the documentation, then use it as follows:

from torchvision.models import get_model, get_model_weights
from torchcam.methods import LayerCAM

# Define your model
model = get_model("resnet18", weights=get_model_weights("resnet18").DEFAULT).eval()
# Set your CAM extractor
cam_extractor = LayerCAM(model)

Please note that by default, the layer at which the CAM is retrieved is set to the last non-reduced convolutional layer. If you wish to investigate a specific layer, use the target_layer argument in the constructor.

Retrieving the class activation map

Once your CAM extractor is set, you only need to use your model to infer on your data as usual. If any additional information is required, the extractor will get it for you automatically.

from urllib.request import urlretrieve
from torchvision.io import decode_image
from torchvision.models import get_model, get_model_weights

# Get a model and an image
weights = get_model_weights("resnet18").DEFAULT
model = get_model("resnet18", weights=weights).eval()
preprocess = weights.transforms()
urlretrieve("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
img = decode_image("dog.jpg")

input_tensor = preprocess(img)

Compute the class activation map:

from torchcam.methods import LayerCAM

with LayerCAM(model) as cam_extractor:
  out = model(input_tensor.unsqueeze(0))
  # Retrieve the CAM by passing the class index and the model output
  activation_map = cam_extractor(out.squeeze(0).argmax().item(), out)

Here class_idx (the first argument) is the index in the model’s output logits of the class you want to explain — out.squeeze(0).argmax().item() picks the top prediction, but you can pass any class index. The extractor returns one activation map per target layer.

If you want to visualize your heatmap, you only need to cast the CAM to a numpy ndarray:

import matplotlib.pyplot as plt
# Visualize the raw CAM
plt.imshow(activation_map[0].squeeze(0).numpy()); plt.axis('off'); plt.tight_layout(); plt.show()

Or if you wish to overlay it on your input image:

import matplotlib.pyplot as plt
from torchvision.transforms.v2.functional import to_pil_image
from torchvision.transforms.functional import center_crop, resize
from torchcam.utils import overlay_mask

# Use the same spatial transforms as inference, before normalization.
model_image = center_crop(
    resize(img, preprocess.resize_size, interpolation=preprocess.interpolation, antialias=preprocess.antialias),
    preprocess.crop_size,
)
result = overlay_mask(to_pil_image(model_image), to_pil_image(activation_map[0].squeeze(0), mode='F'), alpha=0.5)
plt.imshow(result); plt.axis('off'); plt.tight_layout(); plt.show()

[!TIP] Using your own (non-torchvision) model, a Vision Transformer, 3D/video data, or batched inputs? Read the Advanced usage guide — it also covers how to choose the right target_layer and CAM method.

Hitting a cannot register a hook ... / requires grad error, a NaN, or a blank heatmap? See Troubleshooting.

Setup

Python 3.11 (or higher) and uv/pip are required to install TorchCAM. TorchCAM 0.5.0 requires PyTorch 2.4.1 or higher (torch>=2.4.1,=1.26.4,<3). Use the matching torchvision release.

On macOS, TorchCAM 0.5.0 requires Apple Silicon when using published PyTorch wheels. Intel Mac users can use Python 3.11 or 3.12 and install the previous release with pip install "torchcam==0.4.1".

Stable release

You can install the last stable release of the package using pypi as follows:

pip install torchcam

Latest version

Alternatively, if you wish to use the latest features of the project that haven’t made their way to a release yet, you can install the package from source:

pip install "torchcam @ git+https://github.com/frgfm/torch-cam.git"

CAM Zoo

This project is developed and maintained by its author, but the implementation was based on the following research papers:

• Learning Deep Features for Discriminative Localization: the original CAM paper
• Grad-CAM: GradCAM paper, generalizing CAM to models without global average pooling.
• Grad-CAM++: improvement of GradCAM++ for more accurate pixel-level contribution to the activation.
• Smooth Grad-CAM++: SmoothGrad mechanism coupled with GradCAM.
• Score-CAM: score-weighting of class activation for better interpretability.
• SS-CAM: SmoothGrad mechanism coupled with Score-CAM.
• IS-CAM: integration-based variant of Score-CAM.
• XGrad-CAM: improved version of Grad-CAM in terms of sensitivity and conservation.
• Layer-CAM: Grad-CAM alternative leveraging pixel-wise contribution of the gradient to the activation.
• Finer-CAM: contrastive CAM objective highlighting differences between similar classes.
• LeGrad: layerwise positive attention-gradient maps for Vision Transformers.
• RefineCAM: multi-layer refinement producing high-resolution activation maps.

Not sure which one to use? See Choosing a CAM method.

Source: YouTube video (activation maps created by Layer-CAM with a pretrained ResNet-18)

What else

Documentation

The full package documentation is available here for detailed specifications.

Playground app

The Streamlit explorer has three tabs, each with a model selector. Models load only when you request an explanation:

• Image classifiers: pretrained torchvision models, class probabilities, CAMs and overlays.
• VLMs: BLIP, Qwen3-VL 2B and Gemma 4 E2B generate captions. Select a token to view its CAM and next-token probability. See VLM attribution for architecture-specific targets.
• Decision models: Laya and Decider-0.8B score allowed routing choices directly. MiniLM and DeBERTa provide NLI baselines. All use TokenAblation, with one probability and input-token heatmap per option; signed effects are downloadable as JSON.

Uploads accept JPEG/PNG and clear on tab changes; other controls persist. The cache retains one model, with serialized inference per model; active requests can keep an evicted model alive during replacement. A GPU is recommended for the larger VLMs; Gemma requires CUDA or a BF16-accelerated CPU. The VLM displays its actual processed image beside the map. Text explanations are limited to 64 ticket tokens. See text decision models for replacement baselines, native probability calibration and interpretation limits.

Install the optional demo dependencies (Streamlit, Transformers and Laya) to run it locally:

pip install -e ".[demo]"

You can then easily run your app in your default browser by running:

streamlit run demo/app.py

The first computation downloads the selected pretrained checkpoint; later computations use the local model cache. For offline verification, python demo/smoke.py checks navigation, UI state, input alignment and native model adapters using small random-weight fixtures. These fixtures are only tests; the app always requests pretrained checkpoints. python demo/smoke.py --pretrained downloads the deployed models, checks routing examples and verifies that masking high-ranked regions reduces subject-token confidence more than masking low-ranked regions across multiple removal fractions. Gemma is skipped on CPUs without BF16 acceleration. Trusted deployments run it. The deployed version is available on Hugging Face Spaces.

Visualization script

An example script is provided for you to benchmark the heatmaps produced by multiple CAM approaches on the same image:

python scripts/cam_example.py --arch resnet18 --class-idx 232 --rows 2

All script arguments can be checked using python scripts/cam_example.py --help

Performance benchmarks

The purpose of CAM methods is to provide interpretability and they do so by pointing the biggest influence factors on the model outputs. Ideally the CAM should pinpoint all the visual cues that have any influence of the output classification score. For this, we use two metrics:

• Increase in Confidence (higher is better): the fraction of inputs for which masking with the CAM increases the probability of the original predicted class.
• Average Drop (lower is better): the mean relative decrease in that class probability after masking, with increases counted as zero drop.

The table below records historical October 2025 results, obtained with earlier metric and CAM implementations. These values have not been revalidated with the current implementation. The ResNet-18 LayerCAM row has been corrected to match the recorded CSV and original benchmark notebook.

The recorded protocol used the validation set of imagenette2-320, Resize(256) followed by CenterCrop(224), and target layers layer4 for ResNet-18 and features for MobileNet V3 Large. It evaluated the original predicted class. Masking multiplies the ImageNet-normalized input by the CAM: a zero mask therefore corresponds to the ImageNet mean RGB color, not black. No original random seed was recorded; stochastic CAMs and changes to CAM or metric numerics require a fresh benchmark run.

You can run a new benchmark on your hardware with an explicit seed and weight version as follows:

uv run --extra scripts python scripts/eval_perf.py ~/Downloads/imagenette2-320 LayerCAM --arch mobilenet_v3_large --seed 0 --weights IMAGENET1K_V2

The optional --deletion-insertion evaluation uses a normalized zero baseline for both curves and 20 perturbation steps by default. It generates CAMs separately from the classification metrics, consuming extra random draws that can change subsequent stochastic classification CAMs even with the same seed; compare stochastic methods using the same flags. This differs from the original RISE protocol, which uses a blurred image as its insertion baseline; scores from different protocols are not directly comparable.

All script arguments can be checked using python scripts/eval_perf.py --help

Latency benchmark

You crave for beautiful activation maps, but you don’t know whether it fits your needs in terms of latency?

The table below preserves historical October 2021 CPU latency measurements (initial forward pass not included), from the original benchmark commit. They have not been revalidated with current implementations. The GPU column has been retired because the original timer did not synchronize CUDA operations; the current scripts/eval_latency.py synchronizes CUDA before and after timing.

*The base CAM method cannot work with architectures that have multiple fully-connected layers

These CPU measurements used 100 iterations on (224, 224) inputs and a laptop with an Intel® Core™ i7-10750H.

You can run this latency benchmark for any CAM method on your hardware as follows:

uv run --extra scripts python scripts/eval_latency.py SmoothGradCAMpp --device cpu --weights none --output latency.json

Each command runs five fresh processes with one CPU thread. It reports the first call after extractor setup, then runs 10 full CAM warm-up calls before collecting 100 samples. Use --repeat, --threads, --warmup, and --it to change these settings. --weights none uses an untrained model and avoids downloads; omit it to use pretrained weights.

The default --scope extractor excludes the initial model forward pass. --scope end-to-end includes the forward pass and CAM extraction; both exclude image loading, preprocessing, device transfers, and extractor setup. The JSON report saves raw samples, median/p95 latency, resolved layers and weights, versions, revision, and script hash. Summary latency statistics are medians of trial statistics; memory uses the highest peak. Peak RSS is the highest RAM use of a worker through setup and measurement, including output checks. CUDA allocated and reserved memory are reported separately. RSS is unavailable on Windows.

ViT and Swin spatial methods preserve channel-first targets and reshape token/channel-last targets. In eval_latency.py, LeGrad needs an explicit transformer block (--target-layer encoder.layers.encoder_layer_11 for vit_b_16), and RefineCAM needs at least two repeated --target-layer arguments. cam_example.py uses --target for both methods: --target encoder.layers.encoder_layer_11 for LeGrad or --target layer3,layer4 for RefineCAM. Targets in one extractor must share a tensor layout.

All script arguments can be checked using python scripts/eval_latency.py --help

Example notebooks

Explore these runnable notebooks, hosted in frgfm/notebooks:

Each notebook includes setup cells that pin a development snapshot with fixes after TorchCAM 0.5.0. Run those cells to use the intended implementation.

Author

TorchCAM is created and maintained by François-Guillaume (F-G) Fernandez (@frgfm).

Citation

If you wish to cite this project, feel free to use this BibTeX reference:

@misc{torcham2020,
    title={TorchCAM: class activation explorer},
    author={François-Guillaume Fernandez},
    year={2020},
    month={March},
    publisher = {GitHub},
    howpublished = {\url{https://github.com/frgfm/torch-cam}}
}

Contributing

Feeling like extending the range of possibilities of CAM? Or perhaps submitting a paper implementation? Any sort of contribution is greatly appreciated!

You can find a short guide in CONTRIBUTING to help grow this project!

License

Distributed under the Apache 2.0 License. See LICENSE for more information.