人工神经网络的预测大部分时间都是错误的

如何解决人工神经网络的预测大部分时间都是错误的

我的问题是我的人工神经网络预测正确数字中的 10 次中只有大约 2 次，但是当人工神经网络被安装时，它告诉我大约 98% 的准确度。我退出了 ANNS 的初学者，我不知道我是否遗漏了一些明显的东西，或者为什么会这样。 为了测试，我使用了一个包含 81 位数字的表格（总是从 1 到 9 排成一行）。 如果有人可以帮助我，我将不胜感激:)

训练人工神经网络

import tensorflow as tf
import tensorflow_datasets as tfds
from tensorflow.keras.models import save_model

# See available datasets
print(tfds.list_builders())

# Construct a tf.data.Dataset
dataset = tfds.load(name="mnist",split=tfds.Split.TRAIN)

# Build your input pipeline
dataset = dataset.shuffle(1024).batch(32).prefetch(tf.data.experimental.AUTOTUNE)
for features in dataset.take(1):
  image,label = features["image"],features["label"]

# tfds works in both Eager and Graph modes


# See available datasets
print(tfds.list_builders())

# Construct a tf.data.Dataset
dataset = tfds.load(name="mnist",features["label"]

  mnist = tf.keras.datasets.mnist

# Aufteilung in Training- und Testset
(x_train,y_train),(x_test,y_test) = mnist.load_data()
x_train,x_test = x_train / 255.0,x_test / 255.0
print(x_test[0])
# TF Bilderkennungsmodell
model = tf.keras.models.Sequential([
  tf.keras.layers.Flatten(input_shape=(28,28)),tf.keras.layers.Dense(512,activation=tf.nn.relu),tf.keras.layers.Dropout(0.2),tf.keras.layers.Dense(10,activation=tf.nn.softmax)
])

# Crossentropy für die 10 Zahlen Klassen
model.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])

# Modellfitting und Evaluation
model.fit(x_train,y_train,epochs=50)
model.evaluate(x_test,y_test)

filepath = './saved_model2'
save_model(model,filepath)

使用我的表进行测试：

import matplotlib.pyplot as plt
import numpy as np
from tensorflow.keras.models import load_model
import cv2
import PIL


# Model configuration
img_width,img_height = 28,28
batch_size = 250
no_epochs = 25
no_classes = 10
validation_split = 0.2
verbosity = 1


filepath = './saved_model2'

model = load_model(filepath,compile=True)
img = PIL.Image.open("Dataset/raster.jpg")
bilder = []

i= 0
for x in range(0,224,28):
    img1 = img.crop([0,x,28,x+28])
    img2 = img.crop([28,56,x+28])
    img3 = img.crop([56,84,x+28])
    img4 = img.crop([84,112,x+28])
    img5 = img.crop([112,140,x+28])
    img6 = img.crop([140,168,x+28])
    img7 = img.crop([168,196,x+28])
    img8 = img.crop([196,x+28])
    img9 = img.crop([224,252,x+28])
    img1.save("Dataset/eigeneBilder/"+str(i+1)+".jpg")
    img2.save("Dataset/eigeneBilder/"+str(i+2)+".jpg")
    img3.save("Dataset/eigeneBilder/"+str(i+3)+".jpg")
    img4.save("Dataset/eigeneBilder/"+str(i+4)+".jpg")
    img5.save("Dataset/eigeneBilder/"+str(i+5)+".jpg")
    img6.save("Dataset/eigeneBilder/"+str(i+6)+".jpg")
    img7.save("Dataset/eigeneBilder/"+str(i+7)+".jpg")
    img8.save("Dataset/eigeneBilder/"+str(i+8)+".jpg")
    img9.save("Dataset/eigeneBilder/"+str(i+9)+".jpg")
    i += 9
print(bilder)

samples_to_predict = []
for i in range(1,81,1):
    # Generate a plot
    sample = cv2.imread("Dataset/eigeneBilder/"+str(i)+".jpg")
    img = cv2.bitwise_not(sample)
    sample = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
    reshaped_image = sample.reshape((img_width,img_height,1))
    plt.imshow(reshaped_image)
    reshaped_image/255
    # Add sample to array for prediction
    samples_to_predict.append(reshaped_image)

samples_to_predict = np.array(samples_to_predict)

print(samples_to_predict[0])
predictions = model.predict(samples_to_predict)
classes = np.argmax(predictions,axis=1)
print(samples_to_predict)
print(classes)

解决方法

你可以做的一件事是在model.fit中添加validation_data，如下所示：

    model.fit(x_train,y_train,epochs=50,validation_data=(x_test,y_test))

您还可以在展平之前添加一些带池化的 2Dconv 层。或者添加更多神经元。

如果有帮助，请告诉我。