Python机器学习手写数字识别：跟随网站代码实践技术咨询

Hey there! It’s awesome to see you diving into CNN-based handwritten digit recognition with Python—MNIST is such a classic, foundational project, and the tutorial you’re following has solid bones. Let’s break down key steps, fix the incomplete code you shared, and cover common pitfalls to keep your project on track:

Key Fixes & Completions for Your Code Snippet #

First, your import line cuts off: from keras.layers.convolutional imp... is almost certainly missing MaxPooling2D—a critical layer for downsampling in CNNs. Here’s the full, corrected import block:

from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers.convolutional import Conv2D, MaxPooling2D
import numpy as np
from matplotlib import pyplot as plt

Essential Preprocessing Steps You Can’t Skip #

MNIST data needs reshaping and normalization before training—this prevents unstable gradients and speeds up convergence:

# Load the dataset
(X_train, y_train), (X_test, y_test) = mnist.load_data()

# Reshape images to match CNN input format: [samples][width][height][channels]
# MNIST is grayscale, so channels = 1
X_train = X_train.reshape(X_train.shape[0], 28, 28, 1).astype('float32')
X_test = X_test.reshape(X_test.shape[0], 28, 28, 1).astype('float32')

# Normalize pixel values from 0-255 to 0-1 (standard practice for image data)
X_train = X_train / 255.0
X_test = X_test / 255.0

Building a Functional CNN Model #

Using the imports you have, here’s a complete, tutorial-aligned model structure that works well for MNIST:

# Initialize sequential model
model = Sequential()

# Add convolutional layer with 32 filters, 3x3 kernel, ReLU activation
model.add(Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)))
# Add max pooling layer to reduce spatial dimensions
model.add(MaxPooling2D(pool_size=(2, 2)))
# Dropout for regularization (prevents overfitting)
model.add(Dropout(0.2))
# Flatten 2D features to 1D for dense layers
model.add(Flatten())
# Fully connected layer with 128 units
model.add(Dense(128, activation='relu'))
# Output layer: 10 units (one per digit) with softmax for multi-class classification
model.add(Dense(10, activation='softmax'))

# Compile the model
# Using sparse_categorical_crossentropy because MNIST labels are integers (no one-hot needed)
model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

Training & Visualizing Results #

Once the model is built, train it and use Matplotlib to track performance:

# Train the model
history = model.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=10, batch_size=200)

# Plot training & validation accuracy
plt.figure(figsize=(10, 4))
plt.subplot(1, 2, 1)
plt.plot(history.history['accuracy'], label='Train Accuracy')
plt.plot(history.history['val_accuracy'], label='Test Accuracy')
plt.title('Model Accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend()

# Plot training & validation loss
plt.subplot(1, 2, 2)
plt.plot(history.history['loss'], label='Train Loss')
plt.plot(history.history['val_loss'], label='Test Loss')
plt.title('Model Loss')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend()

plt.tight_layout()
plt.show()

Common Issues to Watch For #

• Import Errors: If you’re using newer TensorFlow versions, Keras is integrated into TensorFlow, so you might need to replace keras. with tensorflow.keras. (e.g., from tensorflow.keras.datasets import mnist).
• Overfitting: If training accuracy is much higher than test accuracy, increase the dropout rate (e.g., to 0.3) or add more convolutional layers.
• Slow Training: Adjust the batch size (larger batches speed up training but use more memory) or reduce the number of units in dense layers.

If you run into specific errors (like shape mismatches, low accuracy, or runtime crashes), share the exact details and I’ll help you troubleshoot further!

内容的提问来源于stack exchange，提问作者rw100