TensorFlow加载H5模型报错TypeError：int()参数为NoneType的解决求助

问题场景 #

使用TensorFlow 2.9.2训练模型，模型定义如下：

import tensorflow as tf

encoder_layers = 1
encoder_bidirectional = False

def get_model():    
  model = tf.keras.Sequential(name='model')
  model.add(tf.keras.layers.Dropout(0.5))

  for _ in range(encoder_layers):
    rnn = tf.keras.layers.LSTM(2**6, return_sequences=True)
    if encoder_bidirectional:
      rnn = tf.keras.layers.Bidirectional(rnn)
    model.add(rnn)

  model.add(tf.keras.layers.Dense(2, activation='softmax'))

  return model


def build_model():
  model = get_model()
  model.build(input_shape=(None, None, 25))
  model.compile(
      loss='sparse_categorical_crossentropy',
      optimizer=tf.keras.optimizers.Adam(0.001),
      metrics=['accuracy']
  )

  model.summary()

  return model

训练代码：

# train model
train, dev, test = get_datasets()

model = build_model()

es = EarlyStopping(
      monitor='val_accuracy',
      mode='max',
      verbose=1,
      patience=10)

mc = ModelCheckpoint(
      'model.h5',
      monitor='val_accuracy',
      mode='max',
      verbose=1,
      save_best_only=True)

with tf.device("/GPU:0"):
  model.fit(
      train,
      epochs=500,
      steps_per_epoch=32,
      validation_data=dev,
      callbacks=[es, mc])

best_model = load_model('model.h5')
best_model.evaluate(test)

执行best_model = load_model('model.h5')时触发错误：

Traceback (most recent call last):
  File "/usr/lib/python3.8/runpy.py", line 194, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/usr/lib/python3.8/runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "/experiments/train.py", line 76, in <module>
    app.run(main)
  File "/usr/local/lib/python3.8/dist-packages/absl/app.py", line 308, in run
    _run_main(main, args)
  File "/usr/local/lib/python3.8/dist-packages/absl/app.py", line 254, in _run_main
    sys.exit(main(argv))
  File "/experiments/train.py", line 70, in main
    best_model = load_model(FLAGS.model_path)
  File "/usr/local/lib/python3.8/dist-packages/keras/utils/traceback_utils.py", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/usr/local/lib/python3.8/dist-packages/keras/initializers/initializers_v2.py", line 1056, in _compute_fans
    return int(fan_in), int(fan_out)
TypeError: int() argument must be a string, a bytes-like object or a number, not 'NoneType'

排查发现model.h5中存在batch_input_shape=[null,null,null]的情况，导致加载时初始化器无法计算参数维度。

原因分析 #

使用model.build(input_shape=(None, None, 25))手动指定输入形状时，TensorFlow在保存H5格式模型时会将None序列化为null，加载时无法正确解析为合法的维度值，进而触发初始化器的类型错误。

解决方法与避免措施 #

方法1：改用显式Input层定义模型输入形状 #

在get_model函数中，先添加Input层明确输入形状，替代后续的model.build调用：

def get_model():    
  model = tf.keras.Sequential(name='model')
  # 添加Input层明确输入形状
  model.add(tf.keras.layers.Input(shape=(None, 25)))
  model.add(tf.keras.layers.Dropout(0.5))

  for _ in range(encoder_layers):
    rnn = tf.keras.layers.LSTM(2**6, return_sequences=True)
    if encoder_bidirectional:
      rnn = tf.keras.layers.Bidirectional(rnn)
    model.add(rnn)

  model.add(tf.keras.layers.Dense(2, activation='softmax'))

  return model

def build_model():
  model = get_model()
  # 无需再调用model.build
  model.compile(
      loss='sparse_categorical_crossentropy',
      optimizer=tf.keras.optimizers.Adam(0.001),
      metrics=['accuracy']
  )

  model.summary()

  return model

方法2：使用SavedModel格式保存模型 #

TensorFlow的SavedModel格式对动态形状的支持更友好，修改ModelCheckpoint的配置，改用SavedModel格式保存：

mc = ModelCheckpoint(
      'best_model',  # 不带.h5后缀，默认用SavedModel格式
      monitor='val_accuracy',
      mode='max',
      verbose=1,
      save_best_only=True)

加载时直接调用：

best_model = tf.keras.models.load_model('best_model')

方法3：加载时手动指定输入形状（临时修复） #

如果已经生成了有问题的H5模型，可以通过先加载模型结构、再手动build的方式修复：

# 先重新构建模型结构
model = build_model()
# 加载权重，不加载模型配置
model.load_weights('model.h5')
# 此时模型已可正常使用
model.evaluate(test)

内容的提问来源于stack exchange，提问作者jessicalfr