TensorFlow中export_inference_graph与freeze_graph的区别及自定义模型选型咨询

Great question! Let's break down the differences between TensorFlow's export_inference_graph and freeze_graph scripts, and clarify how to use them for your custom model training workflow.

These two tools serve distinct, complementary roles in preparing a trained model for deployment. Here's how they differ:

1. Core Purpose #

• export_inference_graph: Think of this as a "graph pruner". Its job is to take your full training graph (which includes extra nodes for things like gradient calculation, optimization, and logging) and strip out everything that's not needed for inference. The result is a clean graph structure that only includes the nodes required to run forward passes (input layers, model layers, output layers). Crucially, this output doesn't contain any of your trained weights—just the blueprint of the model.
• freeze_graph: This is the "graph combiner". It takes the pruned graph from export_inference_graph and merges it with your trained variable weights (stored in checkpoint files like .ckpt). It replaces all variable nodes in the graph with their actual trained values, producing a single, self-contained "frozen" graph file. This file has both the model structure and weights, ready to be deployed for inference without needing access to the original checkpoints.

2. Input & Output #

• export_inference_graph:
  • Input: Usually your training graph definition (from code or a saved graph file) and specifications for which nodes are the input/output of your inference pipeline.
  • Output: A graph file (typically .pb or .pbtxt) with only inference-essential nodes—no weight data.
• freeze_graph:
  • Input: The pruned inference graph from export_inference_graph, plus your trained checkpoint files (which hold the variable weights).
  • Output: A frozen graph file (e.g., frozen_inference_graph.pb) that combines structure and weights into one file.

3. Usage Stage #

• export_inference_graph runs first, right after you finish training your model. It's the step that cleans up your training graph for deployment.
• freeze_graph runs second, taking the cleaned graph and binding it to your trained weights to produce the final deployable model.

Short answer: You don't choose one over the other—you use them together (if you're working with TensorFlow 1.x or legacy workflows).

Here's the typical workflow for a custom trained model:

1. Train your model until you're satisfied with its performance, saving checkpoints along the way.
2. Run export_inference_graph to generate a pruned inference graph (removing training-only nodes).
3. Run freeze_graph to merge this pruned graph with your latest checkpoint weights, creating a frozen graph ready for deployment.

That said, if you're using TensorFlow 2.x, these old scripts are mostly deprecated. Instead, you'll want to use tf.saved_model.save() to export your model as a SavedModel bundle, which wraps the graph structure, weights, and inference signatures all in one package—much simpler than the two-step process above. But if you're maintaining a TF1.x project or working with legacy code, the export_inference_graph → freeze_graph pipeline is still the standard approach.

内容的提问来源于stack exchange，提问作者David Papp