在Cloud ML部署InceptionV3模型时遇base64无效字符错误求助

Let's break down what's causing this error and walk through how to fix it:

Root Causes #

Your issue stems from two key mismatches:

1. Request format doesn't match your model's signature: When you saved the model, you defined the input key as inputs, but your request is using image_bytes — Cloud ML can't map this to the model's input tensor. Additionally, you're passing a URL-safe Base64 string which isn't compatible with TensorFlow's decoding logic.
2. URL-safe Base64 isn't supported: TensorFlow's DecodeBase64 op only recognizes standard Base64 characters (+/ instead of -_). Your use of base64.urlsafe_b64encode introduces invalid characters for the decoder.

Step-by-Step Fixes #

1. Switch to Standard Base64 Encoding #

Replace the URL-safe encoding with standard Base64 in both your local test code and request code:

# Replace this line:
# encoded_string = str(base64.urlsafe_b64encode(image_file.read()),'ascii')
# With this:
encoded_string = str(base64.b64encode(image_file.read()), 'ascii')

2. Align Request Body with Model Signature #

Your model's serving signature expects an input key named inputs (matching the tensor input_b64:0). Adjust your request body to match this structure:

Option 1: Basic Request (Single Input)

Since your input tensor has shape (1,), pass the encoded string as a single-element array:

request_body = json.dumps({
    "inputs": [encoded_string]
})

Option 2: Instance Format (Good for Batch Requests)

If you need to include identifiers or send multiple images, use the instances format:

request_body = json.dumps({
    "instances": [
        {"inputs": encoded_string}
    ]
})

3. Verify Input Tensor Shape (Optional) #

If you're still having issues, double-check your input tensor's shape to ensure your request matches:

# Add this to your local test code after loading the model
input_tensor = tf.get_default_graph().get_tensor_by_name('input_b64:0')
print(f"Input tensor shape: {input_tensor.shape}")

• If the shape is (1,), keep passing an array [encoded_string]
• If it's a scalar (), pass the string directly without the array wrapper

4. Redeploy the Updated Model #

Make sure you regenerate your SavedModel with the corrected encoding logic, then redeploy it to Cloud ML Engine — old model caches can sometimes cause unexpected behavior.

Full Working Request Example #

Here's a complete snippet of the corrected request code:

import base64
import json
import requests

# Encode image with standard Base64
with open('MEL_BE_0.jpg', 'rb') as image_file:
    encoded_string = str(base64.b64encode(image_file.read()), 'ascii')

# Build properly formatted request body
request_body = json.dumps({
    "inputs": [encoded_string]
})

# Send request to your Cloud ML endpoint
response = requests.post(
    "YOUR_CLOUD_ML_ENDPOINT_URL",
    data=request_body,
    headers={"Content-Type": "application/json"}
)

# Check the response
print(response.json())

内容的提问来源于stack exchange，提问作者sarim zafar