tensorflow 2.x HeatMap(Grad-CAM)

import numpy as np

import tensorflow as tf

from tensorflow.keras.preprocessing.image import load_img

from keras.models import load_model

from keras.applications.efficientnet import(

    EfficientNetB3,

)

import cv2

img_Path = "cat.jpg"

image = np.array(load_img(img_Path, target_size=(250, 250, 3)))

#cv2.imshow("input", image)

#cv2.waitKey(0)

model = load_model("D:\\Model_02-0.8691406846.h5")

#model.summary()

last_conv_layer = model.get_layer("efficientnetb3").get_layer("top_activation")

last_conv_layer_model = tf.keras.Model(model.get_layer("efficientnetb3").input , last_conv_layer.output)

classifier_input = tf.keras.Input(shape=last_conv_layer.output.shape[1:])

x = classifier_input

for layer_name in ["global_average_pooling2d_1", "dropout_1", "dense_1"]:

    x = model.get_layer(layer_name)(x)

classifier_model = tf.keras.Model(classifier_input, x)

#classifier_model.summary()

with tf.GradientTape() as tape:

    last_conv_layer_output = last_conv_layer_model(image[np.newaxis, ...])

    tape.watch(last_conv_layer_output)

    preds = classifier_model(last_conv_layer_output)

    top_pred_index = tf.argmax(preds[0])

    top_class_channel = preds[:, top_pred_index]

   

grads = tape.gradient(top_class_channel, last_conv_layer_output)[0]

last_conv_layer_output = last_conv_layer_output[0]

guided_grads = (

    tf.cast(last_conv_layer_output > 0, "float32")

    * tf.cast(grads > 0, "float32")

    * grads

)

pooled_guided_grads = tf.reduce_mean(guided_grads, axis=(0, 1))

guided_gradcam = np.ones(last_conv_layer_output.shape[:2], dtype=np.float32)

for i, w in enumerate(pooled_guided_grads):

    guided_gradcam += w * last_conv_layer_output[:, :, i]

guided_gradcam = cv2.resize(guided_gradcam.numpy(), (250, 250))

guided_gradcam = np.clip(guided_gradcam, 0, np.max(guided_gradcam))

guided_gradcam = (guided_gradcam - guided_gradcam.min()) / (

    guided_gradcam.max() - guided_gradcam.min()

)

#plt.imshow(image)

#plt.imshow(guided_gradcam, alpha=0.5)

heatmap = np.uint8(255 * guided_gradcam)

heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET)

img_result = cv2.addWeighted(image, 0.7, heatmap, 0.3, 0)

print(f"image Type {type(image)}, Shape: {image.shape}")

print(f"heatmap Type {type(heatmap)}, Shape: {heatmap.shape}")

cv2.imshow("CAM", heatmap)

cv2.imshow("RESULT", img_result)

cv2.waitKey(0)

cv2.destroyAllWindows()

결과