如何将TensorFlow训练的CNN模型保存为单个.txt文件及转换checkpoint文件为单txt？

Hey there! Let's tackle your two TensorFlow model saving questions one by one:

1. 将训练完成的CNN模型保存为单个.txt文件 #

TensorFlow doesn't natively support saving full models as .txt files, but you can export either the model's weights, or combine the model structure and weights into a single text file depending on your needs.

选项1：仅导出模型权重为.txt #

If you just need to inspect or share the model's weight parameters, you can iterate through the model's layers, extract weights, and write them to a text file:

import tensorflow as tf

# 假设你已经有训练好的模型实例model
# 如果是从SavedModel加载：model = tf.keras.models.load_model("saved_model_dir")

with open("cnn_weights.txt", "w") as f:
    for layer in model.layers:
        # 只处理包含可训练权重的层
        if layer.trainable_weights:
            f.write(f"### Layer: {layer.name} ###\n")
            for weight in layer.trainable_weights:
                weight_np = weight.numpy()
                f.write(f"  Weight Name: {weight.name}\n")
                f.write(f"  Shape: {weight_np.shape}\n")
                f.write("  Values:\n")
                # 如果你觉得全部数值太占空间，可以只打印前几行
                # f.write(f"{weight_np[:2]}\n\n")
                f.write(f"{weight_np}\n\n")

选项2：导出完整模型（结构+权重）为.txt #

If you need both the model architecture and weights in one .txt file, you can serialize the structure to JSON/YAML and combine it with weight data:

import tensorflow as tf
import json

# 获取模型结构的JSON字符串
model_json = model.to_json()
# 将权重转换为可序列化的列表格式
weights_dict = {w.name: w.numpy().tolist() for w in model.trainable_weights}

with open("full_cnn_model.txt", "w") as f:
    f.write("=== Model Architecture (JSON) ===\n")
    f.write(model_json)
    f.write("\n\n=== Model Weights ===\n")
    # 用JSON格式化权重字典，方便阅读
    f.write(json.dumps(weights_dict, indent=2))

注意：这种方式生成的.txt文件会比原生模型文件大很多，因为文本存储数值效率低。如果后续需要重新加载模型进行推理/训练，更推荐使用SavedModel或.h5格式，而非.txt。

2. 将checkpoint目录下的.data、.meta、.index文件转换为单个.txt文件 #

Checkpoint files (.meta, .data, .index) are TensorFlow's legacy format for session-based models. To convert them to a single .txt file, you'll need to load the checkpoint first, then extract the graph structure and variable weights:

import tensorflow as tf

# 加载checkpoint
sess = tf.Session()
# 导入.meta文件中的图结构
saver = tf.train.import_meta_graph("path/to/your/model.ckpt.meta")
# 恢复权重数据
saver.restore(sess, "path/to/your/model.ckpt")

# 提取图结构并转为文本
graph_def = tf.get_default_graph().as_graph_def()
graph_text = str(graph_def)

# 提取所有变量的权重
variables = tf.global_variables()
weights_text = ""
for var in variables:
    var_value = sess.run(var)
    weights_text += f"### Variable: {var.name} ###\n"
    weights_text += f"  Shape: {var_value.shape}\n"
    weights_text += f"  Values:\n{var_value}\n\n"

# 写入合并后的.txt文件
with open("checkpoint_converted.txt", "w") as f:
    f.write("=== Graph Structure ===\n")
    f.write(graph_text)
    f.write("\n\n=== Variable Weights ===\n")
    f.write(weights_text)

sess.close()

提醒：转换后的.txt文件体积会非常大，且不适合直接用于模型加载。如果只是需要迁移模型，建议将checkpoint转换为SavedModel格式（tf.saved_model.simple_save()），而非.txt。

内容的提问来源于stack exchange，提问作者nk1991