Add Visualization of Classification Results with GradCam

[faizanahemad]

Proposed Features

• Add Grid-like visualization for seeing Classification Results.
• Support for Gradcam in same visualizations.
  • add a parameter flag which enables/disables gradcam
• Visualize any model layer with Gradcam.
• Support both Loop style and Keras model.fit style models.
• Choose N samples - examples parameter
• Filter visualisations by:
  • visualise particular actual class only
  • visualise particular predicted class only
  • visualise particular actual class predicted as particular predicted class only
  • visualise randomly
  • misclassified-only
  • correct-only
• show histogram of prediction probabilities Example Section:"Showing 25 Misclassified Examples".

How the API looks like?

def visualize_results(model, dataset, class_names, layer=None, enable_gradcam=False,
                       actual_class=None, predicted_class=None,show_misclassified=False,show_correct=False,
                       examples=25,examples_per_row=5, image_size_multiplier=3,
                       show_actual=True, show_heatmap=False, show_superimposed=True, show_probability_histogram=False):
    ## Code here

Intended Usage Once model is trained this function can be called with Model and dataset to visualize.

Limitations

Online References/Documentation was referred while making this API

• Gradcam Examples for correct and misclassified

• Gradcam examples of layers

• Assignment 11 Gradcam

• Showing 25 Misclassified Examples

• Code Gradcam

• Code Showing 25 Misclassified Examples

• Put these under the inspect module https://github.com/faizanahemad/FastNet/tree/master/fastnet/inspect

• While we will expose only 1 interface outside, make the inner functions modular. See the Code examples for some modularization that I had done.

Branch Name (The github branch in which you are working on this).

Note: Before raising a PR make sure that you have made an issue like this with tag: enhancement, ensure your code is tested and documented.

[faizanahemad]

Note: this task requires reading a lot of code

[unography]

I just had a suggestion, can we use Python 3's type annotation everywhere? That would make functions little easier to read?

something like -

import tensorflow as tf
from typing import List

def visualize_results(model: tf.keras.Model, class_names: List[str]):
    return

[faizanahemad]

Yes @unography that would be good. You can edit the original task with these annotations

[faizanahemad]

Break this into Subtasks if this seems too big to do by 1 person or in 1 week.