如何使用Keras基于训练好的MNIST模型进行预测？

Hey there! Awesome work getting your model trained and saved with solid accuracy—let’s break down exactly how to feed new handwritten digit images into it for predictions. Here’s a step-by-step guide with code:

Step 1: Load Your Saved Model #

First, you’ll need to load the hand_written.h5 model you saved. Make sure you have the necessary Keras imports available:

from keras.models import load_model
import numpy as np

# Load the trained model
model = load_model('hand_written.h5')

Step 2: Prepare Your New Input Data #

Your model expects input data in the exact same format as the training data. That means following these rules:

• The image must be a 28x28 pixel grayscale image (no color channels)
• Pixel values need to be normalized to the range [0, 1] (divided by 255)
• The input shape should include a batch dimension (even for single images, Keras expects this)

Example: Preprocess a Numpy Array Image #

Let’s say you have a numpy array new_digit representing your 28x28 grayscale image (pixel values from 0 to 255). Here’s how to format it:

# Normalize pixel values to [0,1]
new_digit = new_digit.astype('float32') / 255
# Add the single color channel dimension (shape becomes (28,28,1))
new_digit = np.expand_dims(new_digit, axis=-1)
# Add the batch dimension (shape becomes (1,28,28,1))
new_digit = np.expand_dims(new_digit, axis=0)

Example: Load and Preprocess an Image File #

If you’re working with a PNG/JPG file, use PIL to resize and convert it to the right format:

from PIL import Image

# Load image and convert to grayscale
img = Image.open('my_handwritten_digit.png').convert('L')
# Resize to match MNIST's 28x28 size
img = img.resize((28, 28))
# Convert to numpy array
new_digit = np.array(img)

# Apply the same preprocessing steps
new_digit = new_digit.astype('float32') / 255
new_digit = np.expand_dims(new_digit, axis=-1)
new_digit = np.expand_dims(new_digit, axis=0)

Step 3: Run Prediction and Interpret Results #

Once your data is preprocessed, use model.predict() to get the model’s output. The result will be a 1x10 array of probabilities (one for each digit 0-9). To get the predicted digit, take the index of the highest probability with np.argmax():

# Get prediction probabilities (verbose=0 hides unnecessary output)
predictions = model.predict(new_digit, verbose=0)

# Extract the predicted digit
predicted_digit = np.argmax(predictions)

print(f"Predicted Digit: {predicted_digit}")
print(f"Probability for each digit: {predictions[0]}")

Quick Typo Note on Your Training Code #

I spotted a tiny typo in your original code: x_tset=x_test.astype('float32') should be x_test=x_test.astype('float32'). It doesn’t affect your trained model since you normalized x_test correctly afterward, but just something to fix if you re-run training later!

内容的提问来源于stack exchange，提问作者Arjun