ResNet图像分类预测报错：无法将数组重塑为(1,224,224,3)

Hey there! Let's break down why you're hitting this ValueError and get your prediction code working smoothly.

What's Causing the Error? #

First, let's unpack the error message:

ValueError: cannot reshape array of size 50176 into shape (1,224,224,3)

Do the quick math:

• Your target shape requires 224 × 224 × 3 = 150528 elements (for a 3-channel RGB image with a batch dimension)
• The array you're trying to reshape only has 50176 = 224 × 224 elements—this is exactly the size of a single-channel grayscale image.

This means your img_data is a grayscale (1-channel) array, but your ResNet model was trained to expect 3-channel RGB input (since you set the input size to 224×224×3 during training).

How to Fix It #

Follow these steps to align your image data with the model's requirements:

1. Convert single-channel grayscale to 3-channel RGB
   Before reshaping, replicate the single channel three times to mimic an RGB image. Use this numpy operation:

   # Convert (224,224) grayscale array to (224,224,3) RGB
   img_data = np.stack((img_data,) * 3, axis=-1)
   

   If you're loading images with PIL, you can skip the numpy step by converting to RGB directly when opening:

   from PIL import Image
   img = Image.open("your_test_image.jpg").convert("RGB").resize((224,224))
   img_data = np.array(img)
   

2. Reshape to the model's expected batch input shape
   Now that you have a 3-channel image, add the batch dimension (the 1 at the start, since we're predicting one image at a time):

   data = img_data.reshape(1, 224, 224, 3)
   

3. Fix the prediction line
   In your original code, you passed orig (the original grayscale array) to predict_classes—this would still cause a mismatch. Instead, use the reshaped data:

   # For older Keras versions with predict_classes
   model_out = model.predict_classes(data)[0]
   # For newer Keras versions (predict_classes is deprecated)
   predictions = model.predict(data)
   model_out = np.argmax(predictions, axis=1)[0]
   

Full Corrected Code Snippet #

Here's your loop with all fixes applied:

fig = plt.figure()
for num, data in enumerate(test_data):
    img_num = data[1]
    img_data = data[0]
    
    # Convert grayscale to 3-channel RGB if needed
    if len(img_data.shape) == 2:
        img_data = np.stack((img_data,) * 3, axis=-1)
    
    y = fig.add_subplot(9, 3, num+1)
    orig = img_data
    
    # Reshape for model input
    data = img_data.reshape(1, 224, 224, 3)
    
    # Run prediction
    predictions = model.predict(data)
    model_out = np.argmax(predictions, axis=1)[0]
    
    str_label = 'Dog' if model_out == 1 else 'Cat'
    
    y.imshow(orig)
    plt.title(str_label)
    y.axes.get_xaxis().set_visible(False)
    y.axes.get_yaxis().set_visible(False)

plt.show()
plt.savefig('test_labeled.jpg')

内容的提问来源于stack exchange，提问作者Madhi