如何为 MNIST 时尚构建 CNN 模型并使用来自网络的另一组图像对其进行测试?

本文介绍了如何为 MNIST 时尚构建 CNN 模型并使用来自网络的另一组图像对其进行测试?的处理方法，对大家解决问题具有一定的参考价值，需要的朋友们下面随着小编来一起学习吧！

问题描述

导入数据并分成4份进行测试和训练

x_train=x_train/255.0x_test=x_test/255.0c_trainX = trainX.reshape(x_train.shape[0],28,28,1)#x_train.shape[0] = 60model3 = Sequential() # DNN 的类型model3.add(Conv2D(28, kernel_size=(3,3), input_shape = (28,28,1)))model3.add(MaxPooling2D(pool_size=(2,2)))model3.add(Flatten())model3.add(密集(200，激活=relu"))model3.add(密集(10，激活= tf.nn.softmax))model3.compile(optimizer='adam', loss='sparse_categorical_crossentropy',指标=['准确度'])model3.fit(c_trainX, y_train, epochs=15)model3.evaluate(c_testX, y_test)[0.5343353748321533, 0.9064000248908997]----这是我的验证损失和准确性p = model3.predict(c_testX[:10])使用另一个输入导入 urllib从 PIL 导入图像%matplotlib 内联urllib.request.urlretrieve('https://github.com/antony-joy/Data_sets/blob/main/tes.jpg?raw=true',测试.jpg")img = Image.open("testing.jpg")numpyimgdata = np.asarray(img)导入 cv2numpyimgdata=numpyimgdata/255load_img_rz = np.array(Image.open("testing.jpg").resize((28,28)))Image.fromarray(load_img_rz).save('r_kolala.jpeg')打印(调整大小后:"，load_img_rz.shape)numpyimgdata_reshape_grey = cv2.cvtColor(load_img_rz, cv2.COLOR_BGR2GRAY)your_new_array = np.expand_dims(numpyimgdata_reshape_grey,axis=-1)numpyimgdata_reshape = your_new_array.reshape(-1,28, 28, 1) # 完成后使图像在# 与测试和训练数据的维度相同image_predicted_array = model3.predict(numpyimgdata_reshape)test_pred = np.argmax(image_predicted_array,axis=1)打印(预测:"，test_pred)

[5]

这实际上是错误的.它应该打印为一条裤子，用 1 表示mnist 数据集 5 中的原因是凉鞋标签说明

0 T 恤/上衣
1 条裤子
2 套头衫
3 连衣裙
4 大衣
5 凉鞋
6 衬衫
7 运动鞋
8 袋
9 踝靴

我尝试了不同的图像，当我尝试一些靴子或帆布鞋时，我得到了 5 号(凉鞋).这里的实际错误似乎是什么?

解决方案

主要问题很简单.下面我就给大家完整的实现一下你的程序.请注意，我可能会更改模型定义和图像预处理步骤.好的，开始吧.

时尚 MNIST

获取数据 - 进行一些预处理 - 可视化样本.

 from tensorflow.keras.datasets import fashion_mnist(x_train, y_train), (x_test, y_test) = fashion_mnist.load_data()x_train = tf.expand_dims(x_train, -1) # 从 28 x 28 到 28 x 28 x 1x_train = tf.divide(x_train, 255) # 归一化y_train = tf.one_hot(y_train , depth=10) # 使目标成为 One-Hotx_test = tf.expand_dims(x_test, -1) # 从 28 x 28 到 28 x 28 x 1x_test = tf.divide(x_test, 255) # 归一化y_test = tf.one_hot(y_test , depth=10) # 使目标成为 One-Hotx_train.shape, y_train.shape, x_test.shape, y_test.shape(TensorShape([60000, 28, 28, 1]),TensorShape([60000, 10]),TensorShape([10000, 28, 28, 1]),TensorShape([10000, 10]))

[奖励]:看，这些是 28 和灰度图像.现在，无论出于何种原因，如果我们想要调整大小和/或想要使其成为 RGB(3 通道)，我们也可以这样做.在

注意，主要对象为白色，背景为黑色.

模型和训练

我认为最好使用预训练的权重.但是，这里有一个要训练的玩具模型.

model = Sequential()model.add(Conv2D(16, kernel_size=(3,3), input_shape = (28,28,1)))model.add(Conv2D(32, kernel_size=(3,3), activation=relu"))model.add(Conv2D(64, kernel_size=(3,3), activation=relu"))model.add(Conv2D(128, kernel_size=(3,3), activation=relu"))模型.add(GlobalAveragePooling2D())模型.添加(辍学(0.5))模型.添加(密集(10，激活= tf.nn.softmax))模型摘要()# 与你不同，我使用 categorical_crossentropy# 因为我 one_hot 编码了我的 y_train 和 y_test模型编译(优化器='亚当'，损失='categorical_crossentropy',指标=['准确度'])模型拟合(x_train，y_train，batch_size=256，epochs=15，validation_data=(x_test, y_test))

........纪元:15:损失:0.4552 - 准确度:0.8370 - val_loss:0.4008 - val_accuracy:0.8606

预测

让我们对网络搜索的样本进行一些预测.在此之前，让我们先定义一个函数来进行必要的预处理.

#一个预处理函数def infer_prec(img, img_size):img = tf.expand_dims(img, -1) # 从 28 x 28 到 28 x 28 x 1img = tf.divide(img, 255) # 归一化img = tf.image.resize(img, # 根据输入调整大小[img_size, img_size])img = tf.reshape(img, # reshape 以添加批量维度[1, img_size, img_size, 1])返回图像

好的，我抓取一些Fashion MNIST看起来相似的数据，让我们打开其中一个.

导入 cv2导入 matplotlib.pyplot 作为 pltimg = cv2.imread('/content/a.jpg', 0) # 读取图像为灰度打印(img.shape)#(300, 231)plt.imshow(img, cmap=灰色")plt.show()img = infer_prec(img, 28) # 调用预处理函数打印(img.shape) # (1, 28, 28, 1)

到目前为止一切都很好，除了现在我们有一个白色背景，这与我们训练模型的训练样本不同.如果我没记错的话，Fashion MNIST 的所有样本都有黑色背景.此时，如果我们将此样本传递给模型进行预测，它将无法做出准确或接近准确的预测.

当我们将 RGB 样本设为灰度时，白色像素保持白色，而另一个彩色像素变为黑色.对于我们的案例来处理这个问题，我们可以在将灰度图像传递给模型进行预测之前对灰度图像使用 bitwise_not 操作.这个 bitwise_not 只是让 0 变成 1，反之亦然.

导入 cv2导入 matplotlib.pyplot 作为 pltimg = cv2.imread('/content/a.jpg', 0) # 读取图像为灰度img = cv2.bitwise_not(img) # <----- bitwise_not打印(img.shape)#(300, 231)plt.imshow(img, cmap=灰色")plt.show()img = infer_prec(img, 28) # 调用预处理函数打印(img.shape) # (1, 28, 28, 1)

现在，将其传递给模型以获得预测概率.

y_pred = model.predict(img)y_pred数组([[3.1869055e-03, 5.6372599e-05, 1.1225128e-01, 2.2242602e-02,7.7411497e-01、5.8861728e-11、8.7906137e-02、6.2964287e-12、2.4166984e-04, 2.0408438e-08]], dtype=float32)

现在我们可以得到预测的标签并比较gt.

tf.argmax(y_pred,axis=-1).numpy()# array([4]) # 涂层

Importing the data and splitting it into 4 for test and train

x_train=x_train/255.0    
x_test=x_test/255.0
c_trainX = trainX.reshape(x_train.shape[0],28,28,1)#x_train.shape[0] = 60
model3 = Sequential() # type of DNN
model3.add(Conv2D(28, kernel_size=(3,3), input_shape = (28,28,1)))  
model3.add(MaxPooling2D(pool_size=(2,2)))
model3.add(Flatten()) 
model3.add(Dense(200, activation="relu"))                
model3.add(Dense(10, activation=tf.nn.softmax))         
model3.compile(optimizer='adam', loss='sparse_categorical_crossentropy', 
metrics=['accuracy'])

model3.fit(c_trainX, y_train, epochs=15)

model3.evaluate(c_testX, y_test)
[0.5343353748321533, 0.9064000248908997]---- This is my validation loss and 
accuracy
p = model3.predict(c_testX[:10])
USING ANOTHER INPUT

import urllib
from PIL import Image
%matplotlib inline
urllib.request.urlretrieve('https://github.com/antony-joy/Data_sets/blob/main/tes.jpg?raw=true', 
"testing.jpg")
img = Image.open("testing.jpg")
numpyimgdata = np.asarray(img)    
import cv2
numpyimgdata=numpyimgdata/255
load_img_rz = np.array(Image.open("testing.jpg").resize((28,28)))
Image.fromarray(load_img_rz).save('r_kolala.jpeg')
print("After resizing:",load_img_rz.shape)
numpyimgdata_reshaped_grey = cv2.cvtColor(load_img_rz, cv2.COLOR_BGR2GRAY)
your_new_array = np.expand_dims(numpyimgdata_reshaped_grey, axis=-1)
       numpyimgdata_reshaped = your_new_array.reshape(-1,28, 28, 1)     # this is done make the image in
                                                  # the same dimension of that of test and train data
image_predicted_array = model3.predict(numpyimgdata_reshaped)
test_pred = np.argmax(image_predicted_array, axis=1)     
print("predicted:",test_pred)

[5]

This is actually wrong. It should be printed as a trouser which is denoted by 1cause in mnist dataset 5 is sandalLabel Description

0 T-shirt/top
1 Trouser
2 Pullover
3 Dress
4 Coat
5 Sandal
6 Shirt
7 Sneaker
8 Bag
9 Ankle boot

I tried with different images, I am getting number 5(sandals) when I try with some boot or canvas shoes even. What seems to be the actual mistake here?

解决方案

The main problem is very simple. Here I will give you a complete implementation of your program. Please note that I may change the model definition and image preprocessing step. Ok, let get started.

Fashion MNIST

Get the data - Do some preprocessing - Visualize a sample.

from tensorflow.keras.datasets import fashion_mnist   

(x_train, y_train), (x_test, y_test) = fashion_mnist.load_data()

x_train = tf.expand_dims(x_train, -1)     # from 28 x 28 to 28 x 28 x 1 
x_train = tf.divide(x_train, 255)         # Normalize 
y_train = tf.one_hot(y_train , depth=10)  # Make target One-Hot

x_test = tf.expand_dims(x_test, -1)       # from 28 x 28 to 28 x 28 x 1 
x_test = tf.divide(x_test, 255)           # Normalize
y_test = tf.one_hot(y_test , depth=10)    # Make target One-Hot
 
x_train.shape, y_train.shape, x_test.shape, y_test.shape
(TensorShape([60000, 28, 28, 1]),
 TensorShape([60000, 10]),
 TensorShape([10000, 28, 28, 1]),
 TensorShape([10000, 10]))

Now, let's visualize one of a sample from our preprocessed data.

plt.imshow(x_train[0][:,:,0], cmap="gray")
plt.show()

Observe that, the main object right white and the background is black.

Model and Training

It's better to use pretrained weight I think. However, here is a toy model to train.

model = Sequential() 
model.add(Conv2D(16, kernel_size=(3,3), input_shape = (28,28,1)))  
model.add(Conv2D(32, kernel_size=(3,3), activation="relu"))  
model.add(Conv2D(64, kernel_size=(3,3), activation="relu"))  
model.add(Conv2D(128, kernel_size=(3,3), activation="relu"))  
model.add(GlobalAveragePooling2D())     
model.add(Dropout(0.5))         
model.add(Dense(10, activation=tf.nn.softmax))       
model.summary()

# Unlike you I use categorical_crossentropy
# as because I one_hot encoded my y_train and y_test
model.compile(optimizer='adam', 
              loss='categorical_crossentropy', 
              metrics=['accuracy'])

model.fit(x_train, y_train, batch_size=256, 
             epochs=15, validation_data=(x_test, y_test))

....
....
epoch:15: loss: 0.4552 - accuracy: 0.8370 - val_loss: 0.4008 - val_accuracy: 0.8606

Prediction

Let's make some predictions on the web-searched samples. Before that, let's first define a function that will do the necessary preprocessing.

# a preprocess function 
def infer_prec(img, img_size):
    img = tf.expand_dims(img, -1)       # from 28 x 28 to 28 x 28 x 1 
    img = tf.divide(img, 255)           # normalize 
    img = tf.image.resize(img,          # resize acc to the input
             [img_size, img_size])
    img = tf.reshape(img,               # reshape to add batch dimension 
            [1, img_size, img_size, 1])
    return img

Ok, I scrape some Fashion MNIST looking similar data, let's open one of them.

import cv2
import matplotlib.pyplot as plt

img = cv2.imread('/content/a.jpg', 0)   # read image as gray scale    
print(img.shape)   # (300, 231)

plt.imshow(img, cmap="gray")
plt.show()

img = infer_prec(img, 28)  # call preprocess function 
print(img.shape)   # (1, 28, 28, 1)

All is good so far, except now we have a white background, which is not like our training sample on which our model is trained on. If I'm not wrong, all the samples of Fashion MNIST do have a black background. At this point, if we pass this sample to the model for prediction, it wouldn't make accurate or close accurate predictions.

When we make an RGB sample to Grayscale, the white pixel remains white and the other colorful pixel gets black. For our case to handle this, we can use the bitwise_not operation on the grayscale image before passing it to the model for prediction. This bitwise_not simply makes 0 to 1 and vice-versa.

import cv2
import matplotlib.pyplot as plt

img = cv2.imread('/content/a.jpg', 0)  # read image as gray scale    
img = cv2.bitwise_not(img)             # < ----- bitwise_not
print(img.shape)   # (300, 231)

plt.imshow(img, cmap="gray")
plt.show()

img = infer_prec(img, 28)  # call preprocess function 
print(img.shape)   # (1, 28, 28, 1)

Now, pass it to the model for predicted probabilities.

y_pred = model.predict(img)
y_pred  

array([[3.1869055e-03, 5.6372599e-05, 1.1225128e-01, 2.2242602e-02,
        7.7411497e-01, 5.8861728e-11, 8.7906137e-02, 6.2964287e-12,
        2.4166984e-04, 2.0408438e-08]], dtype=float32)

Now we can get the predicted label and compare gt.

tf.argmax(y_pred, axis=-1).numpy() 
# array([4]) # Coat

这篇关于如何为 MNIST 时尚构建 CNN 模型并使用来自网络的另一组图像对其进行测试?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持！