レイヤーの読み込みで失敗してしまいます

Question

実現したいこと

Grad-CAM++を利用して注視領域を可視化したいです。
しかし、レイヤーの読み込みで、レイヤーが存在しないとなって失敗してしまいます。
レイヤー名はTensorBoardの出力をを参考にしてます

発生している問題・エラーメッセージ

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-22-4e280938b6d4> in <module>()
      7 
      8     img = img_to_array(load_img(image_path,target_size=(row,col,3)))
----> 9     img_GCAMplusplus = Grad_Cam_plus_plus(model, target_layer, img, row, col)
     10     time = time.ctime()
     11     img_Gplusplusname = image_path+time+"_GCAM++_%s.jpg"

1 frames
/usr/local/lib/python3.7/dist-packages/keras/engine/training.py in get_layer(self, name, index)
   2629         if layer.name == name:
   2630           return layer
-> 2631       raise ValueError(f'No such layer: {name}. Existing layers are '
   2632                        f'{self.layers}.')
   2633     raise ValueError('Provide either a layer name or layer index at '

ValueError: No such layer: batch_normalization. Existing layers are [<keras.engine.functional.Functional object at 0x7f34aca9a1d0>, <keras.layers.pooling.GlobalAveragePooling2D object at 0x7f34b967b090>, <keras.layers.core.dense.Dense object at 0x7f34b967ba90>, <keras.layers.core.dense.Dense object at 0x7f34b967aa10>].

該当のソースコード

import pandas as pd
import numpy as np
import cv2
import argparse
import keras
import time
import sys
from keras import backend as K
from keras.preprocessing.image import array_to_img, img_to_array, load_img
from keras.applications.resnet import ResNet50

K.set_learning_phase(1)

def Grad_Cam_plus_plus(input_model, layer_name, x, row, col):

    model = input_model

    # 前処理
    X = np.expand_dims(x, axis=0)
    X = X.astype('float32')
    preprocessed_input = X / 255.0


    # 予測クラスの算出
    predictions = model.predict(preprocessed_input)
    class_idx = np.argmax(predictions[0])

    #  使用する重みの抽出、高階微分の計算
    class_output = model.layers[-1].output
    conv_output = model.get_layer(layer_name).output
    grads = K.gradients(class_output, conv_output)[0]
    #first_derivative：１階微分
    first_derivative = K.exp(class_output)[0][class_idx] * grads
    #second_derivative：２階微分
    second_derivative = K.exp(class_output)[0][class_idx] * grads * grads
    #third_derivative：３階微分
    third_derivative = K.exp(class_output)[0][class_idx] * grads * grads * grads

    #関数の定義
    gradient_function = K.function([model.input], [conv_output, first_derivative, second_derivative, third_derivative])  # model.inputを入力すると、conv_outputとgradsを出力する関数


    conv_output, conv_first_grad, conv_second_grad, conv_third_grad = gradient_function([preprocessed_input])
    conv_output, conv_first_grad, conv_second_grad, conv_third_grad = conv_output[0], conv_first_grad[0], conv_second_grad[0], conv_third_grad[0]

    #alphaを求める
    global_sum = np.sum(conv_output.reshape((-1, conv_first_grad.shape[2])), axis=0)
    alpha_num = conv_second_grad
    alpha_denom = conv_second_grad*2.0 + conv_third_grad*global_sum.reshape((1,1,conv_first_grad.shape[2]))
    alpha_denom = np.where(alpha_denom!=0.0, alpha_denom, np.ones(alpha_denom.shape))
    alphas = alpha_num / alpha_denom

    #alphaの正規化
    alpha_normalization_constant = np.sum(np.sum(alphas, axis = 0), axis = 0)
    alpha_normalization_constant_processed = np.where(alpha_normalization_constant != 0.0, alpha_normalization_constant, np.ones(alpha_normalization_constant.shape))
    alphas /= alpha_normalization_constant_processed.reshape((1,1,conv_first_grad.shape[2]))

    #wの計算
    weights = np.maximum(conv_first_grad, 0.0)
    deep_linearization_weights = np.sum((weights * alphas).reshape((-1, conv_first_grad.shape[2])))

    #Lの計算
    grad_CAM_map = np.sum(deep_linearization_weights * conv_output, axis=2)
    grad_CAM_map = np.maximum(grad_CAM_map, 0)
    grad_CAM_map = grad_CAM_map / np.max(grad_CAM_map)

    #ヒートマップを描く
    grad_CAM_map = cv2.resize(grad_CAM_map, (row, col), cv2.INTER_LINEAR)
    jetcam = cv2.applyColorMap(np.uint8(255 * grad_CAM_map), cv2.COLORMAP_JET)  # モノクロ画像に疑似的に色をつける
    jetcam = (np.float32(jetcam) + x / 2)   # もとの画像に合成

    return jetcam

if __name__ == '__main__':
    model = keras.models.load_model("/content/drive/MyDrive/ji/saved_model/my_model")
    target_layer = 'batch_normalization'
    image_path = '/content/drive/MyDrive/Cs-N013.jpg'
    row = 224
    col = 224

    img = img_to_array(load_img(image_path,target_size=(row,col,3)))
    img_GCAMplusplus = Grad_Cam_plus_plus(model, target_layer, img, row, col)
    time = time.ctime()
    img_Gplusplusname = image_path+time+"_GCAM++_%s.jpg"
    cv2.imwrite(img_Gplusplusname, img_GCAMplusplus)
    print("Completed.")

試したこと

いくつかレイヤー名を試してdenseの時のみエラーが下記のようになりました

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-23-907724f744ec> in <module>()
      7 
      8     img = img_to_array(load_img(image_path,target_size=(row,col,3)))
----> 9     img_GCAMplusplus = Grad_Cam_plus_plus(model, target_layer, img, row, col)
     10     time = time.ctime()
     11     img_Gplusplusname = image_path+time+"_GCAM++_%s.jpg"

<ipython-input-18-02dcd9cea551> in Grad_Cam_plus_plus(input_model, layer_name, x, row, col)
     16     class_output = model.layers[-1].output
     17     conv_output = model.get_layer(layer_name).output
---> 18     grads = K.gradienttape(class_output, conv_output)[0]
     19     #first_derivative：１階微分
     20     first_derivative = K.exp(class_output)[0][class_idx] * grads

AttributeError: module 'keras.backend' has no attribute 'gradienttape'

Answer 1

tf.summary()

でレイヤー名を確認することでエラー回避

denseのエラー回避は以下のコードを冒頭に記述することで出来る

import tensorflow as tf
tf.compat.v1.disable_eager_execution()