PathoSpotter - Layer visualizer

Grad-cam Architecture

Single Image Analysis

Setup

pip install -r requirements.txt

Running

usage: single.py [-h] [--image_file IMAGE_FILE] [--model_file MODEL_FILE]
                 [--layer_name LAYER_NAME] [--label LABEL] [--method METHOD]
                 [--output_path OUTPUT_PATH] [--guided] [--no_plot]

Options

Also available running python single.py --help

Option	Description	Default value
image_file	Path to the input image	./examples/without.png
model_file	Path to the model file for the CNN model	./models/glomeruloesclerose
layer_name	Layer to use for grad-CAM. Check Architecture Summary section	all
label	Class label to generate grad-CAM for, -1 = use predicted class	-1
method	Method used to visualize the grad-CAM	CAM_IMAGE_JET
output_path	Path to save images in	./output
guided	Flag to activate guided method	false (deactivated)
no_plot	Flag to Deactivate plot output. Will generate one file for each layer to visualize	false (activated)

Architecture Summary

To check your model Architecture summary you may run model_analysis.py

usage: model_analysis.py [-h] [--model_file MODEL_FILE]

eg.:

$ python3 model_analysis.py --model_file=./models/glomeruloesclerose

This will generate a summary file under the model's path on which you will be able to check the layer's names of you model to specify a layer on --layer_name option

Visualization methods

Useful list of possible methods to be used on --method option

METHOD	PREVIEW - max_pooling2d_33	PREVIEW - conv2d_41
CAM_IMAGE_JET
CAM_IMAGE_BONE
CAM_AS_WEIGHTS
JUST_CAM_JET
JUST_CAM_BONE

Roadmap

• Improve main.py file to deal with folders as input and process multiple input images at once
• Upload report
• Add experiments images on Readme to improve explanation
• Add grad-cam architecture image on Readme
• Add code references

Reference Projects

• Pytorch-grad-cam
• Keras-grad-cam
• Grad-cam