TensorFlow训练后模型恢复求助：get_tensor_by_name命名困惑

Got it, let's break down how get_tensor_by_name works when restoring TensorFlow models—this is a super common gotcha, so I’ll walk you through the naming rules using your code as concrete examples.

The key thing to remember is that every tensor in your graph has a full name made up of two parts:

[Node Name]:[Output Index]

Let’s break that down:

• Node Name: This is the name you explicitly set with the name parameter when creating an operation (like your placeholders). If you don’t set a name, TensorFlow auto-generates one (e.g., Placeholder_0), which is messy to work with later.
• Output Index: Most TensorFlow operations (like placeholder, Variable, or matmul) only produce one output tensor, so this index is almost always 0. For operations that return multiple tensors (e.g., tf.split), you’d use 0, 1, etc., to pick the specific output you want.

Applying This to Your Code #

Looking at the code snippet you shared, here’s what each tensor’s full name would be:

• For x_hat = tf.placeholder(..., name='input_img'): The full tensor name is input_img:0
• For x = tf.placeholder(..., name='target_img'): Full name is target_img:0
• For keep_prob = tf.placeholder(..., name='keep_prob'): Full name is keep_prob:0
• For z_in = tf.placeholder(tf.float32,...): If you set a name like z_input, the full name becomes z_input:0 (always add that :0!)

How to Double-Check Tensor Names #

If you’re ever unsure about a tensor’s exact name, use one of these quick methods:

• Print the name during training: Right after creating a tensor, print its .name attribute:

  print(x_hat.name)  # Will output "input_img:0"
  

• List all tensors after restoring the model: Once you load your saved graph, you can loop through all tensors to see their names:

  saver = tf.train.Saver()
  with tf.Session() as sess:
      saver.restore(sess, "path/to/your/model")
      # Iterate through all tensors in the graph
      for op in tf.get_default_graph().get_operations():
          for tensor in op.outputs:
              print(tensor.name)
  

• Use TensorBoard: Visualize your graph in TensorBoard—you can click on any node to see its associated tensors and their full names.

Example: Restoring and Accessing Tensors #

Here’s how you’d use get_tensor_by_name to grab your placeholders (or any other tensor) after restoring the model:

saver = tf.train.Saver()
with tf.Session() as sess:
    # Restore the saved model
    saver.restore(sess, "./your_saved_model_directory")
    
    # Grab the input placeholder tensor using its full name
    input_img_tensor = tf.get_default_graph().get_tensor_by_name("input_img:0")
    
    # Now you can use it in a feed_dict for inference
    test_data = ...  # Your input data
    feed_dict = {input_img_tensor: test_data}
    # Run inference on some output tensor (replace with your actual output name)
    predictions = sess.run("output_tensor_name:0", feed_dict=feed_dict)

Common Mistakes to Avoid #

• Forgetting the :0: This is the #1 mistake! Using just input_img instead of input_img:0 will throw a "tensor not found" error—always include the output index.
• Auto-generated names: If you skip setting the name parameter when creating tensors, TensorFlow will name them things like Placeholder_1 or Variable_3. This makes restoring a hassle, so always explicitly name your key tensors/operations.
• Variable scopes: If you used tf.variable_scope in your graph (e.g., with tf.variable_scope("encoder"):), the node name will include the scope prefix. For example, a variable named weight inside the encoder scope would have a full name of encoder/weight:0.

内容的提问来源于stack exchange，提问作者Shideh Rezaeifar