keras-conv-visualizer 0.0.3

Package allows visualize convolutional layers from keras models.

Keras Conv Visualizer

Package allows visualize convolutional layers from keras models.

Table of contents

• General info
• Libraries
• Setup
• Documentation
• PyPi
• TODO
• Development
• Status
• Contact

General info

This package is a set of tools for visualizing convolutional layers from keras models. At this moment includes:

• Filters visualization
• Grad-CAM activation visualization
• Intermediate activations visualization

Libraries

• Keras - version 2.4.3
• Matplotlib - version 3.3.3
• NumPy - version 1.19.4
• OpenCV - version 4.4.0.46
• TensorFlow - version 2.4.0rc1

Setup

• Install from PyPi: pip install keras-conv-visualizer

Documentation

Status: in progress

Filters visualization

import matplotlib.pyplot as plt
from tensorflow.keras.applications import VGG16
from keras_conv_visualizer.filters import FilterVisualization

# Model has to have standarized input (std=0, var=1)!
model = VGG16(weights="imagenet", include_top=False, input_shape=(224, 224, 3))
layer_name = "block5_conv3"

# First parameter - trained keras model, second - input_size
fv = FilterVisualization(model)
# First parameter - layer feature index (ex. block1_conv1 has (224, 224, 64) index is from 0 to 63)
# Second parameter - layer name
loss, img = fv.visualize_filter(0, layer_name)
plt.imshow(img)

Result:

Grad-CAM activation visualization

from tensorflow.keras.preprocessing.image import load_img, img_to_array
from tensorflow.keras.applications import VGG16, imagenet_utils
import matplotlib.pyplot as plt
from keras_conv_visualizer.gradcam import GradCAM

img_path = 'elephant.jpg'

# load the input image from disk (in Keras/TensorFlow format) and preprocess it
image = load_img(img_path, target_size=(224, 224))
image = img_to_array(image)
image = imagenet_utils.preprocess_input(image)

model = VGG16(weights="imagenet", input_shape=(224, 224, 3))

cam = GradCAM(model)
# First parameter - image tensor, second - image path, third - alpha value for heatmap (transparency)
heatmap, output = cam.make_superimposed_img(image, img_path, alpha=0.6)

plt.imshow(heatmap)
plt.imshow(output)

Results:

Input image	Heatmap	Superimposed image

Intermediate activations visualization

from keras.models import load_model
from keras.preprocessing import image
from keras_conv_visualization.intermediate_activations import IntermediateActivations

# load the input image from disk (in Keras/TensorFlow format) and preprocess it
img = image.load_img('some_image.png', target_size=(96, 96), color_mode='grayscale')
img_tensor = image.img_to_array(img)
img_tensor /= 255

model = load_model('some_model.h5')

int_activations = IntermediateActivations(model)
int_activations.plot_intermediate_activations(img_tensor)

Input image:

Results:

PyPi

keras-conv-visualizer

TODO

• Add shap values
• Automatically recognition input size for FilterVisualization

Development

Want to contribute? Great!

To fix a bug or enhance an existing module, follow these steps:

• Fork the repo
• Create a new branch (git checkout -b improve-feature)
• Make the appropriate changes in the files
• Verify if they are correct
• Add changes to reflect the changes made
• Commit changes
• Push to the branch (git push origin improve-feature)
• Create a Pull Request

Status

Library is: in progress

Contact

[e-mail] - feel free to contact me!