我用包含 15 种不同水果的数据集训练了我的模型。一切正常，准确率通常在 90% 以上。
但是如果我想创建关于我的 CNN 的混淆矩阵，这些值没有任何意义，因为它不能反映高精度，而且我不知道为什么，因为我已经尝试了不同的实现方式。
import numpy as np
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten, Conv2D, MaxPooling2D, Dropout, BatchNormalization
from tensorflow.keras.preprocessing import image_dataset_from_directory
from sklearn.metrics import classification_report, confusion_matrix
train_ds = image_dataset_from_directory(
directory='training_data/',
labels='inferred',
label_mode='categorical',
batch_size=32,
image_size=(256, 256))
validation_ds = image_dataset_from_directory(
directory='validation_data/',
labels='inferred',
label_mode='categorical',
batch_size=32,
image_size=(256, 256))
#create model
model = Sequential()
#add 2 convolutional layers
model.add(Conv2D(32, (3, 3), activation="relu", padding="same", input_shape=(256, 256, 3), name = "conv1"))
model.add(Conv2D(32, (3, 3), activation="relu", padding="same", name = "conv2"))
#add the pooling layer
model.add(MaxPooling2D(pool_size=(2,2), name="pooling1"))
#add 2 convolutional layers
model.add(Conv2D(32, (3, 3), activation="relu", padding="same", name = "conv3"))
model.add(Conv2D(32, (3, 3), activation="relu", padding="same", name = "conv4"))
#add the pooling layer
model.add(MaxPooling2D(pool_size=(2,2), name="pooling2"))
#add the flattening layer
model.add(Flatten())
#add the fully connected layer
model.add(Dense(256, activation="relu", name="fc1"))
#add dropout layer
model.add(Dropout(0.2, name="dropout2"))
model.add(Dense(15, activation="softmax"))
model.compile(loss="categorical_crossentropy", optimizer=SGD(lr=0.00001, momentum=0.9), metrics=["accuracy"])
model.fit(train_ds, epochs=10, verbose=1)
score = model.evaluate(validation_ds, verbose=0)
print(f'Test loss: {score[0]} / Test accuracy: {score[1]}')
Y_pred = tf.concat([x for x, y in validation_ds], axis=0)
y_pred = model.predict(Y_pred)
predicted_categories = tf.argmax(y_pred, axis=1)
y_true = tf.concat([y for x, y in validation_ds], axis=0)
true_categories = tf.argmax(y_true, axis=1)
print(confusion_matrix(true_categories, predicted_categories))
print(classification_report(true_categories, predicted_categories,))
在这里您可以看到当前实现的值。准确率是 80%，因为我降低了测试的 epoch 数量：
[[ 9 3 10 2 5 5 8 19 1 12 5 11 1 6 11]
[ 9 1 8 5 5 6 7 10 3 15 5 12 3 8 7]
[11 2 10 5 2 5 7 6 2 15 8 11 9 7 8]
[13 2 7 5 4 5 10 11 1 11 8 7 4 8 6]
[ 9 3 7 6 4 9 5 13 3 16 4 11 3 7 3]
[12 4 6 2 2 4 10 18 2 11 8 10 1 11 7]
[10 2 8 5 3 9 4 8 4 11 7 11 8 11 7]
[12 1 9 6 5 5 8 16 0 15 6 8 5 8 12]
[ 9 2 6 5 2 9 9 8 5 8 3 15 8 12 7]
[10 2 4 6 5 11 9 12 6 6 7 6 5 6 9]
[ 1 5 5 9 1 7 13 11 4 11 9 10 7 7 8]
[10 1 9 1 3 8 8 7 2 13 5 8 11 9 13]
[ 8 1 8 3 5 6 9 11 5 12 8 8 4 7 5]
[ 8 2 5 4 3 8 11 7 5 11 17 8 5 12 6]
[13 2 7 6 3 12 9 12 1 11 7 7 6 4 8]]
precision recall f1-score support
0 0.08 0.06 0.07 144
1 0.01 0.03 0.01 33
2 0.09 0.09 0.09 109
3 0.05 0.07 0.06 70
4 0.04 0.08 0.05 52
5 0.04 0.04 0.04 109
6 0.04 0.03 0.03 127
7 0.14 0.09 0.11 169
8 0.05 0.11 0.07 44
9 0.06 0.03 0.04 178
10 0.08 0.08 0.08 107
11 0.07 0.06 0.06 143
12 0.04 0.05 0.04 80
13 0.11 0.10 0.10 123
14 0.07 0.07 0.07 117
accuracy 0.07 1605
macro avg 0.06 0.07 0.06 1605
weighted avg 0.07 0.07 0.07 1605
Test loss: 0.9647821187973022 / Test accuracy: 0.813707172870636
