Introduction

When you run freeze_graph.py, output_node_names must identify the final tensors you want to keep for inference. In a TensorFlow model_with_buckets setup, there is no single universal answer because the graph often produces a list of outputs, one per bucket or decoding step, rather than one obvious final node.

What output_node_names Means

freeze_graph.py removes graph parts that are not needed to compute the outputs you name. So output_node_names should point to the inference endpoints you actually fetch at runtime.

If you pick the wrong names, one of two things happens:

• freezing fails because the nodes do not exist
• freezing succeeds but prunes away tensors you still need

That is why this argument is not a generic TensorFlow setting. It is specific to your graph.

Why Bucketed Models Are Tricky

A model_with_buckets graph usually creates multiple parallel decoder branches to handle different sequence lengths. Each bucket may have its own logits, losses, or decoder outputs.

That means there may not be one built-in node called something obvious like output. Instead, you may have:

• a list of logits tensors
• per-bucket decoder outputs
• training-only loss nodes you do not want
• tensors created without stable names

For freezing, the right approach is usually to create explicit identity nodes for the inference outputs you care about.

Create Explicit Output Nodes

Instead of guessing internal tensor names, attach named identities in the graph.

1import tensorflow as tf
2
3bucket_logits = [
4    tf.constant([[0.1, 0.9]], dtype=tf.float32),
5    tf.constant([[0.8, 0.2]], dtype=tf.float32),
6]
7
8export_nodes = []
9for i, logits in enumerate(bucket_logits):
10    export_nodes.append(tf.identity(logits, name=f"bucket_{i}_logits"))
11
12with tf.compat.v1.Session() as sess:
13    for node in export_nodes:
14        print(node.name)

This pattern gives you stable node names such as bucket_0_logits and bucket_1_logits. Those names are much safer to pass to freeze_graph.py than trying to infer temporary operation names from a large graph dump.

How To Choose The Right Outputs

Ask what your inference program actually consumes. For a sequence model, you usually want decoder logits, predicted token ids, or final probabilities. You usually do not want:

• losses
• optimizer ops
• gradient nodes
• queue runners
• training summaries

If your runtime fetches predicted ids, export those. If it consumes logits, export logits. The freeze step should preserve the tensors used in production, not every tensor the training graph happened to create.

Inspecting The Graph

If you inherited the graph and do not know the output names, inspect it before freezing.

1import tensorflow as tf
2
3with tf.compat.v1.Session() as sess:
4    graph = tf.compat.v1.get_default_graph()
5    for op in graph.get_operations():
6        if "bucket" in op.name or "logits" in op.name:
7            print(op.name)

You can also open the graph in TensorBoard to inspect operation names visually. That is often easier for bucketed seq2seq models because the graph can be large and repetitive.

Example freeze_graph.py Usage

Once you have explicit output names, the freezing command becomes straightforward.

1python freeze_graph.py \
2  --input_graph=graph.pbtxt \
3  --input_checkpoint=model.ckpt \
4  --output_graph=frozen.pb \
5  --output_node_names=bucket_0_logits,bucket_1_logits

If you only serve one bucket at a time, you may export only the corresponding node. If your runtime can select between multiple bucket outputs, include all of them.

A Better Pattern For Old Graphs

In older TensorFlow 1.x code, the biggest source of pain is relying on unnamed tensors. If you control the model code, add explicit tf.identity(..., name="...") nodes wherever you expect to consume outputs after training.

That turns freezing from a graph archaeology exercise into a reproducible export step.

Common Pitfalls

The most common mistake is assuming output_node_names should point at loss nodes or training placeholders. Freezing is for inference, so you need the final prediction-side tensors.

Another issue is expecting one universal answer for every model_with_buckets graph. Bucketed models vary, and the correct names depend on what tensors you expose and use.

Teams also often rely on auto-generated operation names, which change when code changes slightly. Stable identity nodes are a much better export contract.

Finally, remember that freeze_graph.py works on graph structure, not on semantic intent. If you do not name the real outputs clearly, TensorFlow cannot infer which nodes matter to your serving program.

Summary

• 'output_node_names should name the final inference tensors you want preserved in the frozen graph.'
• In a model_with_buckets graph, there is usually no single universal output node name.
• The safest approach is to add named tf.identity nodes for per-bucket inference outputs.
• Inspect the graph or TensorBoard if you inherit a model and do not know the node names.
• Export only the tensors your inference runtime actually consumes.