TensorFlow新手将.pb转TFLite遇错：Graph无SerializeToString属性

Hey there! Let's break down why you're seeing this error and how to fix it quickly.

What's Causing the Error? #

The tf.contrib.lite.toco_convert function (or its TF 2.x equivalent) expects a GraphDef object (the serialized definition of your model graph) as input, but you're passing it a full tf.Graph instance (CD_graph in your code). The SerializeToString method is part of GraphDef, not the Graph class—hence the error.

Solution 1: For TensorFlow 1.x #

Let's adjust your code to pass the correct GraphDef instead of the Graph object. You can either modify your load_graph function to return the GraphDef, or extract it from the loaded graph:

Option A: Modify the Loader to Return GraphDef

import numpy as np
import tensorflow as tf
from collections import OrderedDict

def load_graph_def(frozen_graph_filename):
    with tf.gfile.GFile(frozen_graph_filename, "rb") as f:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())
    return graph_def

# Load the GraphDef instead of the full Graph
CD_graph_def = load_graph_def("CD_Check_k.pb")

# Define your input/output tensor names (match your model's nodes)
input_tensor = "prefix/input_node:0"
output_tensor = "prefix/output_node:0"

# Convert to TFLite
tflite_model = tf.contrib.lite.toco_convert(
    CD_graph_def,
    input_tensors=[input_tensor],
    output_tensors=[output_tensor]
)

# Save the converted model
with open('./CD_Check_k.tflite', "wb") as f:
    f.write(tflite_model)

Option B: Extract GraphDef from Your Existing Loaded Graph

If you want to keep using your original load_graph function:

CD_graph = load_graph("CD_Check_k.pb")
CD_input = CD_graph.get_tensor_by_name('prefix/input_node:0')
CD_output = CD_graph.get_tensor_by_name('prefix/output_node:0')

# Extract the GraphDef from the loaded Graph
graph_def = CD_graph.as_graph_def()

with tf.Session(graph=CD_graph) as sess:
    tflite_model = tf.contrib.lite.toco_convert(
        graph_def,
        input_tensors=[CD_input],
        output_tensors=[CD_output]
    )
    with open('./CD_Check_k.tflite', "wb") as f:
        f.write(tflite_model)

Solution 2: For TensorFlow 2.x #

If you're using TF 2.x, the tf.contrib.lite module has been moved to tf.lite, and the API is cleaner with from_frozen_graph:

import tensorflow as tf

# Directly convert the frozen .pb file
converter = tf.lite.TFLiteConverter.from_frozen_graph(
    graph_def_file="CD_Check_k.pb",
    input_arrays=["prefix/input_node"],  # Input node name (without the :0 suffix)
    input_shapes={"prefix/input_node": [1, 256, 256, 3]},  # Match your input shape
    output_arrays=["prefix/output_node"]  # Output node name (without the :0 suffix)
)

tflite_model = converter.convert()

# Save the model
with open('./CD_Check_k.tflite', "wb") as f:
    f.write(tflite_model)

Key Notes #

• Make sure your input/output node names exactly match what's in your frozen graph (use tools like TensorBoard to verify if you're unsure).
• For TF 2.x, you don't need to manually load the graph—from_frozen_graph handles that for you.

内容的提问来源于stack exchange，提问作者Harshit Mishra