![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250814161438-689e0b6ecf3c0.png\")
<\/p>\nDigit Recognizer | Kaggle \u9898\u9762<\/p>\n
Digit Recognizer-CNN | Kaggle \u4e0b\u9762\u4ee3\u7801\u7684kaggle\u7248\u672c<\/p>\n
\u4f7f\u7528CNN\u8fdb\u884c\u624b\u5199\u6570\u5b57\u8bc6\u522b<\/p>\n
\u5b66\u4e60\u5230\u4e86\u7f51\u7edc\u642d\u5efa\u624b\u6cd5+\u5b66\u4e60\u7387\u9000\u706b+\u6570\u636e\u589e\u5e7f \u63d0\u9ad8\u8bad\u7ec3\u6548\u679c\u3002<\/p>\n
\u4f7f\u7528\u6df7\u6dc6\u77e9\u9635 \u4ee5\u53ca\u5bf9\u5206\u7c7b\u51fa\u9519\u6982\u7387\u6700\u5927\u7684\u4f8b\u5b50\u5355\u72ec\u62ce\u51fa\u6765\u5206\u6790\u3002<\/p>\n
<\/p>\n
\u6700\u7ec8\u4ee599.546%\u6b63\u786e\u7387 \u6392\u5728 86\/1035<\/p>\n
\u62c6\u6210X\u548cY \u8f93\u51fa\u6bcf\u4e2a\u6570\u5b57 train\u4e2d\u7684\u6570\u636e\u91cf<\/p>\n
import pandas as pd
\nimport seaborn as sns
\ntrain = pd.read_csv("\/kaggle\/input\/digit-recognizer\/train.csv")
\ntest = pd.read_csv("\/kaggle\/input\/digit-recognizer\/test.csv")
\nY_train = train["label"]
\nX_train = train.drop(labels = ["label"],axis = 1)
\ndel train <\/p>\n
g = sns.countplot(x=Y_train)
\nY_train.value_counts() <\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250814161439-689e0b6f806cb.png\")
<\/p>\n
1.\u786e\u8ba4\u6570\u636e\u5e72\u51c0 \u2192 \u6ca1\u6709\u7f3a\u5931\u503c\u3002<\/p>\n
print(X_train.isnull().any().describe()) # \u4e0b\u9762\u4ee3\u8868\u6240\u6709\u5217 \u90fd\u662fFalse
\nprint(test.isnull().any().describe()) <\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250814161440-689e0b707049a.png\")
<\/p>\n
2.\u7070\u5ea6\u5f52\u4e00\u5316（\u9664\u4ee5255） \u2192 \u63d0\u5347\u8bad\u7ec3\u901f\u5ea6\u548c\u7a33\u5b9a\u6027\u3002<\/p>\n
reshape \u2192 \u8f6c\u6362\u6210 CNN \u6240\u9700\u7684 (height, width, channel) \u683c\u5f0f\u3002 \u00a0<\/p>\n
\u539f\u6765\u662f(\u6837\u672c\u6570, 784) -> (\u6837\u672c\u6570, \u9ad8\u5ea6, \u5bbd\u5ea6, \u901a\u9053\u6570)\u00a0 \u6837\u672c\u6570, 28, 28, 1<\/p>\n
X_train, test = X_train \/ 255.0, test \/ 255.0
\nX_train = X_train.values.reshape(-1, 28, 28, 1)
\ntest = test.values.reshape(-1, 28, 28, 1) <\/p>\n
3.One-Hot \u7f16\u7801 \u53ea\u6709\u4e00\u4e2a\u4f4d\u7f6e\u4e3a1\u7684\u6807\u7b7e \u2192 \u540e\u7eed softmax \u8f93\u51fa10\u4e2a\u7c7b\u522b\u7684\u6982\u7387\u3002<\/p>\n
from tensorflow.keras.utils import to_categorical
\nY_train = to_categorical(Y_train, num_classes=10) <\/p>\n
4.\u5212\u5206\u9a8c\u8bc1\u96c6 \u2192 \u7528\u4e8e\u8c03\u53c2\u3001\u9632\u6b62\u8fc7\u62df\u5408\u3002<\/p>\n
from sklearn.model_selection import train_test_split
\nX_train, X_val, Y_train, Y_val = train_test_split(X_train, Y_train, test_size=0.1, random_state=2) <\/p>\n
5.\u53ef\u89c6\u5316\u4e00\u4e2a\u6837\u672c \u2192 \u68c0\u67e5\u6570\u636e\u662f\u5426\u6b63\u786e\u52a0\u8f7d\u3002<\/p>\n
import matplotlib.pyplot as plt
\nplt.imshow(X_train[0][:,:,0], cmap='gray')
\nplt.title("Example Image")
\nplt.show() <\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250814161440-689e0b70a73d5.png\")
<\/p>\n<\/p>\n
\u5377\u79ef\u5757：2\u4e2aConv(5 * 5)+ BN + \u6c60\u5316\u00a0 \u00a0 \u00a0 ~\u00a0 \u00a0 \u00a02\u4e2aConv(3 * 3) + BN + \u6c60\u5316\u3002<\/p>\n
Dropout\u820d\u5f03\u4e00\u4e9b\u795e\u7ecf\u5143\u9632\u6b62\u8fc7\u62df\u5408\u3002<\/p>\n
import numpy as np
\nimport tensorflow as tf<\/p>\n
from tensorflow.keras.models import Sequential
\nfrom tensorflow.keras.layers import Conv2D, MaxPooling2D as MaxPool2D
\nfrom tensorflow.keras.layers import Dense, Dropout, Flatten, BatchNormalization
\nfrom tensorflow.keras.optimizers import Adam
\nfrom tensorflow.keras.callbacks import ReduceLROnPlateau
\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator<\/p>\n
model = Sequential()<\/p>\n
# \u5377\u79ef\u5757 1
\nmodel.add(Conv2D(32, (5,5), padding='same', activation='relu', input_shape=(28,28,1)))
\nmodel.add(BatchNormalization()) # \u6279\u5f52\u4e00\u5316
\nmodel.add(Conv2D(32, (5,5), padding='same', activation='relu'))
\nmodel.add(BatchNormalization())
\nmodel.add(MaxPool2D(pool_size=(2,2)))
\nmodel.add(Dropout(0.25))<\/p>\n
# \u5377\u79ef\u5757 2
\nmodel.add(Conv2D(64, (3,3), padding='same', activation='relu'))
\nmodel.add(BatchNormalization())
\nmodel.add(Conv2D(64, (3,3), padding='same', activation='relu'))
\nmodel.add(BatchNormalization())
\nmodel.add(MaxPool2D(pool_size=(2,2), strides=(2,2)))
\nmodel.add(Dropout(0.25)) <\/p>\n
\u8f93\u51fa\u5c42：Flatten + Dense + Softmax<\/p>\n
model.add(Flatten())
\nmodel.add(Dense(256, activation='relu'))
\nmodel.add(BatchNormalization())
\nmodel.add(Dropout(0.5))
\nmodel.add(Dense(10, activation='softmax')) <\/p>\n
model.compile\u00a0 \u00a0 \u00a0Adam\u4f18\u5316\u5668 + \u4ea4\u53c9\u71b5\u635f\u5931 + accuracy\u8bc4\u4f30<\/p>\n
optimizer = Adam(learning_rate=0.001)<\/p>\n
model.compile(optimizer=optimizer,
\n loss='categorical_crossentropy',
\n metrics=['accuracy']) <\/p>\n
\u5b66\u4e60\u7387\u9000\u706b \u82e5 3 \u4e2a epoch \u5185\u6ca1\u63d0\u5347，\u5c31\u628a lr \u51cf\u534a<\/p>\n
learning_rate_reduction = ReduceLROnPlateau(
\n monitor='val_accuracy', patience=3, verbose=1,
\n factor=0.5, min_lr=1e-5
\n) <\/p>\n
\u6570\u636e\u589e\u5e7f（ImageDataGenerator） \u8f7b\u5fae\u65cb\u8f6c\u3001\u7f29\u653e\u3001\u5e73\u79fb\u7b49\u6269\u5145\u6570\u636e<\/p>\n
datagen = ImageDataGenerator(
\n rotation_range=10,
\n zoom_range=0.1,
\n width_shift_range=0.1,
\n height_shift_range=0.1
\n) <\/p>\n
\u8bad\u7ec3 \u8bbe\u7f6e epoch\u00a0 batch_size\u00a0 steps<\/p>\n
datagen.fit(X_train)<\/p>\n
# ===== \u8bad\u7ec3 =====
\nepochs = 30
\nbatch_size = 86
\nsteps = int(np.ceil(len(X_train) \/ batch_size)) # \u4e00\u6b21\u8bad\u7ec3\u8986\u76d6\u6240\u6709\u6837\u672c<\/p>\n
history = model.fit(
\n datagen.flow(X_train, Y_train, batch_size=batch_size), # \u6570\u636e\u589e\u5f3a
\n epochs=epochs,
\n steps_per_epoch=steps,
\n validation_data=(X_val, Y_val), # \u9a8c\u8bc1\u6570\u636e
\n callbacks=[learning_rate_reduction], # \u56de\u8c03\u5b66\u4e60\u7387
\n verbose=2 # \u6bcf\u4e2a epoch \u8f93\u51fa\u4e00\u6b21
\n)
\n440\/440 – 144s – 327ms\/step – accuracy: 0.9950 – loss: 0.0170 – val_accuracy: 0.9957 – val_loss: 0.0134 – learning_rate: 6.2500e-05
\nEpoch 30\/30<\/p>\n
Epoch 30: ReduceLROnPlateau reducing learning rate to 3.125000148429535e-05.
\n440\/440 – 145s – 329ms\/step – accuracy: 0.9952 – loss: 0.0153 – val_accuracy: 0.9955 – val_loss: 0.0141 – learning_rate: 6.2500e-05 <\/p>\n
\u8fd9\u662f\u8bad\u7ec3\u65e5\u5fd7\u7684\u6700\u540e\u4e00\u90e8\u5206\u8f93\u51fa \u51c6\u786e\u7387\u8fbe\u5230 99.5%<\/p>\n<\/p>\n
1. history\u62a5\u544a\u4e2d \u8bad\u7ec3\u96c6\u548c\u9a8c\u8bc1\u96c6\u7684 Loss \u548c Accuracy \u53ef\u89c6\u5316\u3002<\/p>\n
val\u4e3a\u9a8c\u8bc1，\u9a8c\u8bc1\u4e0d\u6bd4\u8bad\u7ec3\u5dee \u8bf4\u660e\u6ca1\u6709\u592a\u8fc7\u62df\u5408\u3002<\/p>\n
fig, ax = plt.subplots(2,1)<\/p>\n
# \u7ed8\u5236\u8bad\u7ec3\u96c6\u548c\u9a8c\u8bc1\u96c6\u7684 Loss
\nax[0].plot(history.history['loss'], color='b', label="Training loss")
\nax[0].plot(history.history['val_loss'], color='r', label="validation loss")
\nax[0].legend(loc='best', shadow=True)<\/p>\n
# \u7ed8\u5236\u8bad\u7ec3\u96c6\u548c\u9a8c\u8bc1\u96c6\u7684 Accuracy
\nax[1].plot(history.history['accuracy'], color='b', label="Training accuracy")
\nax[1].plot(history.history['val_accuracy'], color='r', label="Validation accuracy")
\nax[1].legend(loc='best', shadow=True) <\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250814161441-689e0b7126ff3.png\")
<\/p>\n<\/p>\n
2. \u6df7\u6dc6\u77e9\u9635 confusion_matrix<\/p>\n
from sklearn.metrics import confusion_matrix
\nimport itertools<\/p>\n
def plot_confusion_matrix(cm, classes,
\n normalize=False,
\n title='Confusion matrix',
\n cmap=plt.cm.Blues):<\/p>\n
plt.imshow(cm, interpolation='nearest', cmap=cmap)
\n plt.title(title)
\n plt.colorbar()
\n tick_marks = np.arange(len(classes))
\n plt.xticks(tick_marks, classes, rotation=45)
\n plt.yticks(tick_marks, classes)<\/p>\n
if normalize:
\n cm = cm.astype('float') \/ cm.sum(axis=1)[:, np.newaxis]<\/p>\n
thresh = cm.max() \/ 2.
\n for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):
\n plt.text(j, i, cm[i, j],
\n horizontalalignment="center",
\n color="white" if cm[i, j] > thresh else "black")<\/p>\n
plt.tight_layout()
\n plt.ylabel('True label')
\n plt.xlabel('Predicted label')<\/p>\n
Y_pred = model.predict(X_val) # \u9884\u6d4b\u503c
\nY_pred_classes = np.argmax(Y_pred, axis = 1) # \u9884\u6d4b\u5bf9\u5e94\u7c7b\u522b
\nY_true = np.argmax(Y_val,axis = 1) # \u771f\u5b9e\u7c7b\u522b<\/p>\n
confusion_mtx = confusion_matrix(Y_true, Y_pred_classes)
\nplot_confusion_matrix(confusion_mtx, classes = range(10)) <\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250814161441-689e0b71a4be6.png\")
<\/p>\n<\/p>\n
3. \u5bf9\u4e8e\u5206\u7c7b\u9519\u8bef\u7684\u4f4d\u7f6e \u7528\u9884\u6d4b（\u9519\u8bef\u6570\u5b57\u7684\u6982\u7387 – \u6b63\u786e\u6807\u7b7e\u7684\u6982\u7387）<\/p>\n
\u8f93\u51fa\u9884\u6d4b\u9519\u8bef\u5dee\u522b\u6700\u5927\u7684\u56fe\u7247 \u662f\u4ec0\u4e48\u6837\u7684\u3002<\/p>\n
# \u627e\u51fa\u9884\u6d4b\u9519\u8bef\u7684\u4f4d\u7f6e \u5e03\u5c14\u6570\u7ec4
\nerrors = (Y_pred_classes != Y_true )<\/p>\n
# \u63d0\u53d6\u9519\u8bef\u9884\u6d4b\u7684\u8be6\u7ec6\u6570\u636e
\nY_pred_classes_errors = Y_pred_classes[errors]
\nY_pred_errors = Y_pred[errors]
\nY_true_errors = Y_true[errors]
\nX_val_errors = X_val[errors]<\/p>\n
# \u9519\u8bef\u9884\u6d4b\u7684\u6700\u5927\u6982\u7387
\nY_pred_errors_prob = np.max(Y_pred_errors,axis = 1)<\/p>\n
# \u6b63\u786e\u6807\u7b7e\u5bf9\u5e94\u7684\u9884\u6d4b\u6982\u7387
\ntrue_prob_errors = np.diagonal(np.take(Y_pred_errors, Y_true_errors, axis=1))<\/p>\n
# \u5dee\u503c：\u9884\u6d4b\u7684\u6700\u5927\u6982\u7387 – \u6b63\u786e\u6807\u7b7e\u6982\u7387
\ndelta_pred_true_errors = Y_pred_errors_prob – true_prob_errors<\/p>\n
# \u627e\u51fa\u6982\u7387\u5dee\u6700\u5927\u7684\u9519\u8bef
\nmost_important_errors = np.argsort(delta_pred_true_errors)[-6:]<\/p>\n
def display_errors(errors_index,img_errors,pred_errors, obs_errors):
\n """ This function shows 6 images with their predicted and real labels"""
\n n = 0
\n nrows = 2
\n ncols = 3
\n fig, ax = plt.subplots(nrows,ncols,sharex=True,sharey=True)
\n for row in range(nrows):
\n for col in range(ncols):
\n error = errors_index[n]
\n ax[row,col].imshow((img_errors[error]).reshape((28,28)))
\n ax[row,col].set_title("Predicted label :{}\\\\nTrue label :{}".format(pred_errors[error],obs_errors[error]))
\n n += 1<\/p>\n
display_errors(most_important_errors, X_val_errors, Y_pred_classes_errors, Y_true_errors) <\/p>\n
\u53d1\u73b0\u9884\u6d4b\u9519\u8bef\u7684\u51e0\u5f20\u56fe\u7247 \u672c\u8eab\u5c31\u5f88\u5bb9\u6613\u8bef\u89e3<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250814161442-689e0b7252134.png\")
<\/p>\n<\/p>\n
4. \u9884\u6d4b&\u63d0\u4ea4\u7ed3\u679c<\/p>\n
results = model.predict(test)
\nresults = np.argmax(results,axis = 1) # \u6982\u7387\u8f6c\u7c7b\u522b
\nresults = pd.Series(results,name="Label")<\/p>\n
submission = pd.concat([pd.Series(range(1,28001),name = "ImageId"),results],axis = 1)
\nsubmission.to_csv("cnn_mnist_datagen.csv",index=False) <\/p>\n","protected":false},"excerpt":{"rendered":"
\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb173\u6b21\u3002\u672c\u6587\u4ecb\u7ecd\u4e86\u4f7f\u7528\u5377\u79ef\u795e\u7ecf\u7f51\u7edc(CNN)\u8fdb\u884c\u624b\u5199\u6570\u5b57\u8bc6\u522b\u7684\u5b8c\u6574\u6d41\u7a0b\u3002\u9996\u5148\u5bf9MNIST\u6570\u636e\u96c6\u8fdb\u884c\u9884\u5904\u7406\uff0c\u5305\u62ec\u5f52\u4e00\u5316\u3001reshape\u548cone-hot\u7f16\u7801\u3002\u7136\u540e\u6784\u5efa\u5305\u542b\u4e24\u4e2a\u5377\u79ef\u5757(5×5\u548c3×3\u5377\u79ef\u6838)\u7684CNN\u6a21\u578b\uff0c\u91c7\u7528\u6279\u5f52\u4e00\u5316\u548cDropout\u9632\u6b62\u8fc7\u62df\u5408\u3002\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4f7f\u7528\u5b66\u4e60\u7387\u9000\u706b\u548c\u6570\u636e\u589e\u5f3a\u7b56\u7565\uff0c\u6700\u7ec8\u5728\u9a8c\u8bc1\u96c6\u4e0a\u8fbe\u5230\u4e8699.5%\u7684\u51c6\u786e\u7387\u3002\u6587\u7ae0\u8fd8\u5c55\u793a\u4e86\u635f\u5931\/\u51c6\u786e\u7387\u66f2\u7ebf\u3001\u6df7\u6dc6\u77e9\u9635\u5206\u6790\u4ee5\u53ca\u5178\u578b\u9519\u8bef\u6837\u672c\u7684\u53ef\u89c6\u5316\uff0c\u6700\u540e\u751f\u6210\u6d4b\u8bd5\u96c6\u9884\u6d4b\u7ed3\u679c\u5e76\u63d0\u4ea4\u3002\u6574\u4e2a\u6d41\u7a0b\u6db5\u76d6\u4e86\u6570\u636e\u9884\u5904\u7406\u3001\u6a21\u578b\u6784\u5efa\u3001\u8bad\u7ec3\u4f18\u5316\u548c\u7ed3\u679c\u8bc4\u4f30\u7b49\u5173\u952e\u73af\u8282\u3002<\/p>\n","protected":false},"author":2,"featured_media":56863,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[5347,50,5349,2863,5923,5924,207,5346],"topic":[],"class_list":["post-56870","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-kaggle","tag-50","tag-5349","tag-2863","tag-5923","tag-5924","tag-207","tag-5346"],"yoast_head":"\n