I've created a little PyTorch script to display the activation maps of a specific, or all the CNN's layers. Since I'm learning PyTorch, I thought it would be interesting to do this little project to get used to handling torch objects and tensors.

• Python 3.8
• PyTorch
• Matplotlib

To get a local copy up and running, follow these simple example steps.

• Python ⩾ 3.8

  sudo apt install python3 python3-pip

1. Clone the repo

   git clone https://github.com/Clement-W/Activation-Maps-Visualiser-PT.git
   cd Activation-Maps-Visualiser-PT/

2. Create and activate a virtual environment

   pip3 install virtualenv --upgrade
   virtualenv venv
   source venv/bin/activate

3. Install the requirements

   pip3 install -r requirements.txt

There is two main ways to use this python script :

• Call the function show_activation_maps(model, layer_name, image). This function shows the activation maps for a specific layer in the model.
• Call the function save_all_activation_maps(model, image,path_to_directory). This function saves the activation maps of every layers of the model in the specified directory.

Check the next section to see an example.

Import modules :

from torchvision import models, transforms, datasets
import torch
from PIL import Image
from ActivationMapsExtractor import save_all_activation_maps, show_activation_maps

Set torch device :

device = torch.device("cuda:0"  if torch.cuda.is_available() else  "cpu")

Load pretrained RestNet-18 model from PyTorch :

model = models.resnet18(pretrained=True)

Load an image to feed the ResNet-18 model :

image = Image.open("imageNet/sample.JPEG")
transform = transforms.Compose([transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor()])
image = transform(image).to(device)

If you want to see the activation maps of a specific layer by it's name for that input image :

show_activation_maps(model,"conv1",image)

If you want to save the activation maps of every layers for that input image :

save_all_activation_maps(model, image,"./ResNet-activation-maps")

I'm still learning PyTorch, so feel free to use Issues or PR to report errors and/or propose additions or corrections to my code. Any contributions you make are greatly appreciated.

Distributed under the MIT License. See LICENSE for more information.

• Add a Jupyter notebook demo