easy-gradcam 0.0.5

GradCAM for torchvision, timm and huggingface model.

easy_gradcam

A lightweight tool to generate Grad-CAM visualizations for image classification models. It supports popular backbones such as ResNet, Vision Transformers (ViT), and Hugging Face Transformers.

This package is built on top of PyTorch for deep learning model implementation, and provides visualization utilities powered by Matplotlib and Seaborn. It is designed to help users easily train, evaluate, and visualize results with clear and customizable plots.

---

Installation

pip install easy_gradcam

Quick Start

1. Import dependencies

# === data preprocess === 
import cv2
from PIL import Image
from pathlib import Path
import torchvision.transforms as transforms

# === model(maybe choose one?) ===
import torchvision.models as models
import timm
from transformers import AutoModelForImageClassification

# === this visualization tool ===
from easy_gradcam.classification import EasyGradCAM
from easy_gradcam.visualization import save_heatmap, save_mix_heatmap

2. Load a model

You can use different backbones:

# Example 1: ResNet-50 (from torchvision)
model = models.resnet50(pretrained=True)   # targets: "layer4"

# Example 2: ViT (from timm)
model = timm.create_model("vit_base_patch16_224_miil", pretrained=True)   # targets: "blocks.10"

# Example 3: DINOv2 (from Hugging Face)
model = AutoModelForImageClassification.from_pretrained(
    "facebook/dinov2-small-imagenet1k-1-layer"
) # targets: "dinov2.encoder.layer.11"

# Example 4: Your own model
model = CustomModel(...)

model.eval()

2.1 Identify target layers

To find the correct target layer names for your model, you can print the model architecture:

print(model)

3. Prepare an image

img_path = Path("./exp1.jpg")
# 1. Use OpenCV to read an image (choose one)
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

# 2. Use Pillow to read an image (choose one)
img = Image.open(img_path).convert("RGB")
img = np.array(img)

totensor = transforms.ToTensor()
resize = transforms.Resize((224, 224))
normalize = transforms.Normalize([0.485, 0.456, 0.406],
                                 [0.229, 0.224, 0.225])

t = totensor(img)
t = resize(t)
t = normalize(t)
t = t.unsqueeze(0)  # add batch dimension

4. Compute Grad-CAM

# Example 1: single target layer (choose one)
gradcam = EasyGradCAM(model, targets="dinov2.encoder.layer.11")

# Example 2: multiple target layers (choose one)
gradcam = EasyGradCAM(model, targets=["dinov2.encoder.layer.10", "dinov2.encoder.layer.11"])  

# Extract features and gradients
feats, grads = gradcam.cal_feat_and_grad(t)

# Generate heatmaps
heats = gradcam.cal_heats(img, feats, grads)

5. Save results

output_path = Path("results")
output_path.mkdir(parents=True, exist_ok=True)

for i in range(len(heats)):
    for name in heats[i]:
        # Save plain heatmap
        save_heatmap(
            save_path=f"{output_path}/{img_path}-{i}-{name}.jpg",
            heat=heats[i][name],
            cmap="jet",
            title="grad-cam"
        )

        # Save overlay with original image
        save_mix_heatmap(
            save_path=f"{output_path}/{img_path}-{i}-{name}-mix.jpg",
            heat=heats[i][name],
            ori_img=img,
            cmap="jet"
        )

Example Output

• results/exp1-0-dinov2.encoder.layer.11.jpg: heatmap only
• results/exp1-0-dinov2.encoder.layer.11-mix.jpg: heatmap overlay on the input image

Notes

• Make sure the target layer you pass matches the internal structure of the model.
• Pretrained models from torchvision, timm, and Hugging Face are supported.
• Heatmaps are saved as .jpg files in the results/ directory.

Bugs/Requests

Please send bug reports and feature requests through github issue tracker.