Module `tfhelper.visualization.cam`

Expand source code

import tensorflow as tf
import numpy as np
import cv2


def get_cam_image(model_, x, img_size=(28, 28), layer_idx=None):
    """
    Get CAM image from a model and x

    Args:
        model_ (tf.keras.models.Model): TensorFlow Model
        x (np.ndarray): Test Data
        img_size (tuple): Image size for x
        layer_idx (int): Layer Index to generate CAM.
                         If None, last layer of Conv2D will be chosen.
    Returns:
        np.ndarray: CAM image array

    """
    if layer_idx is None:
        for layer_idx in range(len(model_.layers) - 1, -1, -1):
            if type(model_.layers[layer_idx]) == tf.keras.layers.Conv2D:
                break

    cam_model_ = tf.keras.models.Model(model_.inputs, [model_.layers[layer_idx].output, model_.output])
    conv_out, model_out = cam_model_(x)

    cam_images_ = np.zeros((x.shape[0], img_size[0], img_size[1]))

    for i, outs in enumerate(zip(conv_out, model_out)):
        c_out, m_out = outs
        predict_idx = np.argmax(m_out)
        chosen_weight = model_.layers[-1].weights[0][:, predict_idx]

        cam_img_ = np.zeros(c_out.shape[0:2])

        for j in range(c_out.shape[2]):
            cam_img_ += c_out[:, :, j] * chosen_weight[j]

        cam_images_[i] = cv2.resize(cam_img_.numpy(), img_size)

    return cam_images_

Functions

def get_cam_image(model_, x, img_size=(28, 28), layer_idx=None)

Get CAM image from a model and x

Args

model_ : tf.keras.models.Model: TensorFlow Model
x : np.ndarray: Test Data
img_size : tuple: Image size for x
layer_idx : int: Layer Index to generate CAM. If None, last layer of Conv2D will be chosen.

Returns

np.ndarray: CAM image array

Expand source code

def get_cam_image(model_, x, img_size=(28, 28), layer_idx=None):
    """
    Get CAM image from a model and x

    Args:
        model_ (tf.keras.models.Model): TensorFlow Model
        x (np.ndarray): Test Data
        img_size (tuple): Image size for x
        layer_idx (int): Layer Index to generate CAM.
                         If None, last layer of Conv2D will be chosen.
    Returns:
        np.ndarray: CAM image array

    """
    if layer_idx is None:
        for layer_idx in range(len(model_.layers) - 1, -1, -1):
            if type(model_.layers[layer_idx]) == tf.keras.layers.Conv2D:
                break

    cam_model_ = tf.keras.models.Model(model_.inputs, [model_.layers[layer_idx].output, model_.output])
    conv_out, model_out = cam_model_(x)

    cam_images_ = np.zeros((x.shape[0], img_size[0], img_size[1]))

    for i, outs in enumerate(zip(conv_out, model_out)):
        c_out, m_out = outs
        predict_idx = np.argmax(m_out)
        chosen_weight = model_.layers[-1].weights[0][:, predict_idx]

        cam_img_ = np.zeros(c_out.shape[0:2])

        for j in range(c_out.shape[2]):
            cam_img_ += c_out[:, :, j] * chosen_weight[j]

        cam_images_[i] = cv2.resize(cam_img_.numpy(), img_size)

    return cam_images_