保存含Keras模型的变量列表遇h5py报错，需实现压缩保存

问题原因 #

直接用pickle+gzip保存Keras模型对象时，pickle序列化模型的过程中会触发Keras底层的模型保存逻辑，该逻辑依赖h5py。即便你已安装h5py，也可能因为Python环境不匹配（比如h5py安装在其他虚拟环境）导致报错。此外，直接pickle模型对象本身也不是Keras推荐的保存方式。

可行解决方案 #

方案1：单独保存模型为HDF5后打包压缩 #

先将每个模型保存为标准HDF5文件，再把所有HDF5文件打包压缩，这种方式兼容性最好：

保存代码

import gzip
import os
from keras.models import load_model

# 临时保存每个模型
temp_model_paths = []
for idx, model in enumerate(list_of_vars):
    temp_path = f"temp_model_{idx}.h5"
    model.save(temp_path)
    temp_model_paths.append(temp_path)

# 将所有临时模型文件写入压缩包
with gzip.open("compressed_models.gz", "wb") as zip_handle:
    for path in temp_model_paths:
        with open(path, "rb") as model_handle:
            zip_handle.write(model_handle.read())

# 清理临时文件
for path in temp_model_paths:
    os.remove(path)

加载代码

import gzip
import os
from keras.models import load_model

# 从压缩包提取临时模型文件
temp_model_paths = []
with gzip.open("compressed_models.gz", "rb") as zip_handle:
    for idx in range(len(list_of_vars)):
        temp_path = f"temp_model_{idx}.h5"
        temp_model_paths.append(temp_path)
        with open(temp_path, "wb") as model_handle:
            model_handle.write(zip_handle.read())

# 加载模型并清理临时文件
loaded_models = []
for path in temp_model_paths:
    model = load_model(path)
    loaded_models.append(model)
    os.remove(path)

方案2：分离模型结构与权重后序列化保存 #

只保存模型的结构（JSON格式）和权重数组，再用pickle+gzip压缩，避免直接序列化模型对象：

保存代码

import gzip
import pickle
from keras.models import model_from_json

# 提取每个模型的结构和权重
model_info_list = []
for model in list_of_vars:
    model_json = model.to_json()
    model_weights = model.get_weights()
    model_info_list.append((model_json, model_weights))

# 压缩保存结构与权重数据
with gzip.open("models_struct_weights.pkl.gz", "wb") as handle:
    pickle.dump(model_info_list, handle)

加载代码

import gzip
import pickle
from keras.models import model_from_json

# 加载并解压数据
with gzip.open("models_struct_weights.pkl.gz", "rb") as handle:
    model_info_list = pickle.load(handle)

# 重建模型
loaded_models = []
for json_str, weights in model_info_list:
    model = model_from_json(json_str)
    model.set_weights(weights)
    loaded_models.append(model)

方案3：修复h5py环境问题 #

如果坚持要用原方法，先确认当前Python环境是否真的安装了h5py：

1. 运行以下代码检查：

import h5py
print(h5py.__version__)

2. 如果报错，在当前环境重新安装h5py：

# pip环境
pip install h5py
# conda环境
conda install h5py

内容的提问来源于stack exchange，提问作者Morteza