Skip to content

Latest commit

 

History

History
107 lines (79 loc) · 3.14 KB

techniques.md

File metadata and controls

107 lines (79 loc) · 3.14 KB
Raw

XAI Techniques Implementations

This document contains report useful information on the techniques used in the project.

Technique Resources
LRP
LIME LIME paper resources
SHAP
GRAD-CAM XPLIQUE

The Model

The model used in this work is a Sequential model from Tensorflow. The model is subclassed so that it can be adapted to the LIME explainer by implementing the required method predict_proba.

import tensorflow as tf

from tensorflow.keras.models import Sequential
from tensorflow.keras import layers
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.losses import SparseCategoricalCrossentropy
from tensorflow.keras.metrics import SparseCategoricalAccuracy


class DigitClassifier(Sequential):
    def __init__(self):
        super(DigitClassifier, self).__init__([
            layers.Lambda(lambda x: tf.image.rgb_to_grayscale(x)),
            layers.Conv2D(24, (3, 3), activation='relu', input_shape=(28, 28, 1), name='conv_1'),
            layers.MaxPooling2D((2, 2)),
            layers.Dropout(0.25),
            layers.Conv2D(36, (3, 3), activation='relu', name='conv2'),
            layers.MaxPooling2D((2, 2)),
            layers.Dropout(0.25),
            layers.Flatten(),
            layers.Dense(128, activation='relu', name='dense_128'),
            layers.Dense(10, activation='linear', name='visualized')
        ])

        self.compile(
            optimizer=Adam(0.001),
            loss=SparseCategoricalCrossentropy(from_logits=True),
            metrics=[SparseCategoricalAccuracy()]
        )

    def predict_proba(self, x, batch_size=32, verbose=0):
        return self.predict(x)

The Techniques

The following is a list of the techniques with some information about how to use them for the particular case of MNIST digits classification.

LRP

LIME

LIME requires the images to be in rgb format. You can convert images from grayscale to rgb by using Tensorflow.

import tensorflow as tf

# Turn the images into rgb, expected from LIME
train_data_rgb = tf.image.grayscale_to_rgb(tf.expand_dims(train_data, -1))
test_data_rgb = tf.image.grayscale_to_rgb(tf.expand_dims(test_data, -1))

Once the data is in the correct format we can train a classifier.

from utils.DigitClassifier import DigitClassifier

# Create an instance of the classifier and train it
classifier = DigitClassifier()
classifier.fit(
    train_data_rgb,
    train_labels,
    epochs=1
)

The explanations for an image can be computed using the lime library itself.

from lime import lime_image
from lime.wrappers.scikit_image import SegmentationAlgorithm

# Setup the LIME explainer
explainer = lime_image.LimeImageExplainer(verbose = False)
segmenter = SegmentationAlgorithm('quickshift', kernel_size=1, max_dist=5, ratio=0.2)

# Generate the explanation for an image
explanation = explainer.explain_instance(
    test_data[img_num],
    classifier_fn = classifier.predict_proba,
    top_labels=10, hide_color=0, num_samples=1000, segmentation_fn=segmenter
)

SHAP

XPLIQUE