An open python library for researchers and developers to generate GradCAM explanations for the TensorFlow CNN models.
Other popular python ML frameworks will soon be supported by the cnncam library
The below instructions assume you already have pip installed and exposed to the python environment where you want to run cnncam.
Official instructions for installing pip can be found here
Run the below pip command in a shell of your choice.
pip install cnncam
We currently support two ways of obtaining GradCAM heatmaps:
- Display heatmap. If you're only interested in seeing the heatmap images for your model's prediction, you can run the below script:
from cnncam import display_heatmap
from keras.applications.vgg16 import VGG16
display_heatmap(model=VGG16(), # your keras model
img=img, # your image
predicted_class=pred, # your models prediction for the image
layer_name='block5_conv3', # the layer you would like to see GradCAM for
alpha=0.6 # opacity of heatmap overlayed on image
)Output:
We can observe the differences in convolutional layer behaviours across different models, for example, the below code executes the same function on Xception instead of VGG16:
from cnncam import display_heatmap
from keras.applications.xception import Xception
display_heatmap(model=Xception(), # your keras model
img=img, # your image
predicted_class=pred, # your models prediction for the image
layer_name='block5_conv3', # the layer you would like to see GradCAM for
alpha=1 # opacity of heatmap overlayed on image
)Output:
See /examples for executable examples, including the above application of our implementation of GradCAM on VGG-16 with the owner of this repo's very cute cat, meso.