TensorFlow中label_image.py指定input_layer报错：'input'不在计算图中

Hey there, let's work through this issue together—this is a common gotcha when working with retrained TensorFlow graphs, so we'll break down the "why" and the fix step by step.

The error The name 'input' refers to an Operation not in the graph means exactly what it says: when you pass --input_layer=input to label_image.py, TensorFlow looks through your loaded retrained_graph.pb and can't find an operation (node) named input.

Retrained graphs often wrap nodes in a namespace (like import/) or use different default names depending on the training script you used. The key here is to find the exact name of the input and output nodes in your specific graph, not rely on generic suggestions from other issues.

You have two reliable ways to get the correct node names:

1. Use TensorBoard #

Open your TensorBoard instance and look at your graph:

• For the input layer: Find the starting node where image data flows into the graph (this is almost always a Placeholder node). Note its full name—including any prefixes like import/ (you can hover over the node to see its full name).
• For the output layer: Locate the node that outputs your classification scores (usually named something like final_result). Again, note its full name including any namespace prefixes.

2. Inspect the Graph with a Quick Python Script #

If TensorBoard isn't giving you clear names, run this small script to print every operation in your graph:

import tensorflow as tf

# Use tf.compat.v1 for compatibility with older TensorFlow versions
with tf.compat.v1.Session() as sess:
    # Load your retrained graph
    graph_def = tf.compat.v1.GraphDef()
    with open('retrained_graph.pb', 'rb') as f:
        graph_def.ParseFromString(f.read())
        sess.graph.as_default()
        tf.import_graph_def(graph_def, name='')
    
    # Print all operation names in the graph
    print("All graph operations:")
    for op in sess.graph.get_operations():
        print(op.name)

Look for:

• An input node: It'll likely have "Placeholder" in the name (e.g., import/Placeholder)
• An output node: Look for something like import/final_result (this is the node that outputs your class probabilities)

Once you have the correct names, update your command to use them. For example, if your input node is import/Placeholder and output node is import/final_result:0 (the :0 refers to the first tensor output of the operation), your command would be:

py label_image.py --image=test.jpg --graph=retrained_graph.pb --labels=retrained_labels.txt --input_layer=import/Placeholder --output_layer=import/final_result:0 --input_width=160 --input_height=120

Suggestions like Mul/input/placeholder are specific to someone else's graph structure—every retrained graph can have slightly different node names depending on how you set up the training (e.g., custom input sizes, preprocessing steps, or using a different base model). You need to match the names in your graph, not someone else's.

The same logic applies to the output layer: use the exact name from your graph inspection. If you get a similar error for the output layer, just repeat the steps above to find the correct output node name and update the --output_layer parameter.

内容的提问来源于stack exchange，提问作者cristian b