92 lines
3.9 KiB
Python
92 lines
3.9 KiB
Python
from Load_process.file_processing import Process_File
|
||
from keras.models import Model
|
||
from matplotlib import pyplot as plt
|
||
import cv2
|
||
import numpy as np
|
||
from keras import backend as K
|
||
from keras.preprocessing import image
|
||
import tensorflow as tf
|
||
import datetime
|
||
|
||
class Grad_CAM:
|
||
def __init__(self, Label, One_Hot, Experiment_Name, Layer_Name) -> None:
|
||
self.experiment_name = Experiment_Name
|
||
self.Layer_Name = Layer_Name
|
||
self.Label = Label
|
||
self.One_Hot_Label = One_Hot
|
||
self.Save_File_Name = self.Convert_One_Hot_To_int()
|
||
|
||
pass
|
||
|
||
def process_main(self, model, index, images):
|
||
for i in range(len(images)):
|
||
array = np.expand_dims(images[i], axis=0) # 替圖片增加一個維度，代表他的數量
|
||
heatmap = self.gradcam(array, model)
|
||
self.plot_heatmap(heatmap, images[i], self.Save_File_Name[i], index, i)
|
||
|
||
pass
|
||
|
||
def Convert_One_Hot_To_int(self):
|
||
return [np.argmax(Label)for Label in self.One_Hot_Label]
|
||
|
||
|
||
def gradcam(self, Image, model, pred_index = None):
|
||
# 首先，我們創建了一個模型，將輸入圖像映射為最後一個卷積層的激活值以及輸出的預測結果。
|
||
grad_model = Model(
|
||
[model.inputs], [model.get_layer(self.Layer_Name).output, model.output]
|
||
)
|
||
|
||
# 然後，我們計算了對於輸入圖像而言預測類別的最高梯度，並相對於最後一個卷積層的激活值進行了計算。
|
||
with tf.GradientTape() as tape: # 創建一個梯度紀錄器，並做前向傳播
|
||
last_conv_layer_output, preds = grad_model(Image)
|
||
if pred_index is None:
|
||
pred_index = tf.argmax(preds[0])
|
||
class_channel = preds[:, pred_index]
|
||
|
||
# 這是輸出神經元（預測的頂部或所選的）對於最後一個卷積層的輸出特徵圖的梯度。
|
||
grads = tape.gradient(class_channel, last_conv_layer_output)
|
||
|
||
# 這是一個向量，其中每個項目是在特定特徵圖通道上梯度的平均強度。
|
||
pooled_grads = tf.reduce_mean(grads, axis=(0, 1, 2))
|
||
|
||
# 我們將特徵圖陣列中的每個通道乘以「這個通道相對於頂部預測類別的重要性」，然後將所有通道加總起來，以獲得熱圖類別激活。n
|
||
last_conv_layer_output = last_conv_layer_output[0]
|
||
heatmap = last_conv_layer_output @ pooled_grads[..., tf.newaxis]
|
||
heatmap = tf.squeeze(heatmap)
|
||
|
||
# For visualization purpose, we will also normalize the heatmap between 0 & 1
|
||
heatmap = tf.maximum(heatmap, 0) / tf.math.reduce_max(heatmap)
|
||
return heatmap.numpy()
|
||
|
||
def plot_heatmap(self, heatmap, img, Label, index, Title):
|
||
File = Process_File()
|
||
|
||
# ReLU
|
||
heatmap = np.maximum(heatmap, 0)
|
||
# 正規化
|
||
heatmap /= np.max(heatmap)
|
||
|
||
# 讀取影像
|
||
# img = cv2.imread(img)
|
||
fig, ax = plt.subplots()
|
||
# im = cv2.resize(cv2.cvtColor(cv2.imread(img_path), cv2.COLOR_BGR2RGB), (img.shape[1], img.shape[0]))
|
||
|
||
# 拉伸 heatmap
|
||
img_path = cv2.resize(img, (512, 512))
|
||
heatmap = cv2.resize(heatmap, (512, 512))
|
||
heatmap = np.uint8(255 * heatmap)
|
||
|
||
img_path = cv2.cvtColor(img_path, cv2.COLOR_BGR2RGB)
|
||
|
||
# 以 0.6 透明度繪製原始影像
|
||
ax.imshow(img_path, alpha=1)
|
||
# 以 0.4 透明度繪製熱力圖
|
||
ax.imshow(heatmap, cmap='jet', alpha=0.3)
|
||
|
||
save_root = '../Result/CNN_result_of_reading('+ str(datetime.date.today()) + " )/" + str(Label)
|
||
File.JudgeRoot_MakeDir(save_root)
|
||
save_root = File.Make_Save_Root(self.experiment_name + "-" + str(index) + "-" + str(Title) + ".png", save_root)
|
||
# 存plt檔
|
||
plt.savefig(save_root)
|
||
plt.close("all") # 關閉圖表
|
||
pass |