使用Caffe预训练模型预测时报错：TypeError: unhashable type: 'numpy.ndarray'

Let's break down why you're hitting this error and how to fix it step by step:

Root Cause #

The error comes from how you're passing the image to net.forward():

res = net.forward({image})

Here, {image} tries to create a dictionary where the key is the numpy array image—but numpy arrays are unhashable, which is why Python throws that TypeError. Caffe expects you to pass a dictionary where the key is the name of your network's input blob (like 'data' for LeNet) and the value is the preprocessed image data.

Step-by-Step Fix #

1. Use the Correct Input Blob Name

First, check your lenet_train_test.prototxt—the input layer is almost certainly named data. That's the key you need to use in the dictionary for net.forward().

2. Preprocess the Image to Match Caffe's Requirements

Caffe expects input data in a specific format (different from OpenCV's default):

• Channel-first order (CHW instead of OpenCV's HWC)
• Batch dimension added (since networks are designed for batch processing)
• Normalization/matching the mean value used during training (critical for accurate predictions)

3. Corrected Code

Here's the fixed version with proper preprocessing:

import caffe
import cv2
import numpy as np

# Set Caffe mode (CPU or GPU)
caffe.set_mode_cpu()
# caffe.set_mode_gpu()  # Uncomment if you have GPU support

# Load model and weights (note: for prediction, it's better to use deploy.prototxt instead of train_test.prototxt)
model = 'lenet_train_test.prototxt'
weights = 'lenet_iter_10000.caffemodel'
net = caffe.Net(model, weights, caffe.TEST)

# Load and preprocess image
image = cv2.imread('test1.png', cv2.IMREAD_GRAYSCALE)  # LeNet uses grayscale images
image = image.astype(np.float32)

# Resize to match LeNet's input size (typically 28x28)
image = cv2.resize(image, (28, 28))

# Normalize (adjust based on your training mean/scale; example assumes mean 0, scale 1/255)
image /= 255.0

# Add channel and batch dimensions: (H, W) -> (1, 1, H, W) (1 channel for grayscale, 1 batch)
image = image[np.newaxis, np.newaxis, :, :]

# Forward pass with correct input blob name
net.blobs['data'].data[...] = image  # Alternative way to set input
res = net.forward()

# Get the prediction result
prediction = res['prob'].argmax()  # Assuming your output blob is named 'prob'
print(f"Predicted class: {prediction}")

Additional Notes #

• Use deploy.prototxt: Your current code uses lenet_train_test.prototxt, which includes training layers like Data or Loss. For inference, it's better to use a deploy.prototxt that removes training-specific layers and uses Input layer instead.
• Mean Subtraction: If you used a mean file during training, add this step after loading the image:

  mean_file = 'mean.binaryproto'
  mean_blob = caffe.proto.caffe_pb2.BlobProto()
  with open(mean_file, 'rb') as f:
      mean_blob.ParseFromString(f.read())
  mean = caffe.io.blobproto_to_array(mean_blob)[0]
  image -= mean[0]  # Subtract grayscale mean
  

内容的提问来源于stack exchange，提问作者Zhe Ma