我在做一些NLP练习。我无法在一个hub.KerasLayer层中使用已保存的模型 :(

下面是我完成的步骤：

1. 创建模型 - OK
2. 编译模型 - OK
3. 拟合/训练模型 - OK

import tensorflow as tf
from tensorflow.keras import layers
import tensorflow_hub as hub

# Create a Keras layer using the USE pretrained layer from tensorflow hub
sentence_encoder_layer = hub.KerasLayer("https://www.kaggle.com/models/google/universal-sentence-encoder/TensorFlow2/universal-sentence-encoder/2",
                                        input_shape=[],
                                        dtype=tf.string,
                                        trainable=False,
                                        name="USE")

# Create model using the Sequential API
model_6 = tf.keras.Sequential([
  # sentence_encoder_layer, # take in sentences and then encode them into an embedding
  layers.Lambda(lambda x: sentence_encoder_layer(x)),
  layers.Dense(64, activation="relu"),
  layers.Dense(1, activation="sigmoid"),
], name="model_6_USE")


# Compile the model
model_6.compile(loss="binary_crossentropy",
                optimizer=tf.keras.optimizers.Adam(),
                metrics=["accuracy"])

# Train and fit the model
history_6 = model_6.fit(train_sentences,
                        train_labels,
                        epochs=5,
                        validation_data=(val_sentences, val_labels))

1. 对模型进行预测 - OK

# Make model predictions
model_6_pred_prods = model_6.predict(val_sentences) 
model_6_pred_prods[:10]

1. 保存模型 - OK

# Save model_6 to native Keras format
model_6.save("model_6_saved.keras")

1. 加载模型 - OK

# Re-initialize the hub keraslayer
sentence_encoder_layer = hub.KerasLayer("https://www.kaggle.com/models/google/universal-sentence-encoder/TensorFlow2/universal-sentence-encoder/2",
                                        input_shape=[],
                                        dtype=tf.string,
                                        trainable=False,
                                        name="USE")

# Load the model
loaded_model_6 = tf.keras.models.load_model("model_6_saved.keras",
                                            custom_objects={'KerasLayer':hub.KerasLayer,
 'tf':tf,
 'sentence_encoder_layer': sentence_encoder_layer,
 },
                                            safe_mode=False)

# Verify the loaded model
print(loaded_model_6.summary())

1. 评估加载后的模型 - 失败。

# Evaluate the loaded_model
loaded_model_pred_probs = loaded_model_6.evaluate(val_sentences, val_labels)

下面是错误信息..它说hub层未定义，尽管我在同一个代码单元中已经重新初始化过...在加载模型时它也被定义过...

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
/tmp/ipykernel_2313/2915160527.py in <cell line: 0>()
      1 # Evaluate the loaded_model
----> 2 loaded_model_pred_probs = loaded_model_6.evaluate(val_sentences, val_labels)

1 frames
/usr/local/lib/python3.12/dist-packages/keras/src/utils/python_utils.py in <lambda>(x)
     13 model_6 = tf.keras.Sequential([
     14   # sentence_encoder_layer, # take in sentences and then encode them into an embedding
---> 15   layers.Lambda(lambda x: sentence_encoder_layer(x)),
     16   layers.Dense(64, activation="relu"),
     17   layers.Dense(1, activation="sigmoid"),

NameError: Exception encountered when calling Lambda.call().

name 'sentence_encoder_layer' is not defined

Arguments received by Lambda.call():
  • inputs=tf.Tensor(shape=(None,), dtype=string)
  • mask=None
  • training=False

解决方案

名称错误的原因是Keras无法序列化被Lambda层包装的外部变量。你可以通过去掉Lambda包装，将 hub.KerasLayer 直接输入到你的模型中来修复它。同时使用 tf_keras 以保持与更新的TensorFlow版本的兼容性。我已经用示例数据复现实验过代码，请告诉我这是否对你有用。

import os
os.environ["TF_USE_LEGACY_KERAS"] = "1"

import tensorflow as tf
import tf_keras as keras
from tf_keras import layers
import tensorflow_hub as hub
import numpy as np

# Dummy Data for testing
train_sentences = np.array(["I loved it", "Terrible movie", "It was okay"])
train_labels = np.array([1, 0, 0])
val_sentences = np.array(["Best thing ever", "I hated it"])
val_labels = np.array([1, 0])

# Define the TF Hub Layer
sentence_encoder_layer = hub.KerasLayer(
    "https://www.kaggle.com/models/google/universal-sentence-encoder/TensorFlow2/universal-sentence-encoder/2",
    input_shape=[],
    dtype=tf.string,
    trainable=False,
    name="USE"
)

# Build Model (Pass the layer DIRECTLY to tf_keras Sequential)
model_6 = keras.Sequential([
    sentence_encoder_layer, 
    layers.Dense(64, activation="relu"),
    layers.Dense(1, activation="sigmoid"),
], name="model_6_USE")

# Compile the model
model_6.compile(
    loss="binary_crossentropy",
    optimizer=keras.optimizers.Adam(),
    metrics=["accuracy"]
)

# Train the model
history_6 = model_6.fit(
    train_sentences,
    train_labels,
    epochs=5,
    validation_data=(val_sentences, val_labels)
)

# Predict & Save (Use the native SavedModel format)
model_6_pred_prods = model_6.predict(val_sentences)
model_6.save("model_6_saved")

# Load the model
loaded_model_6 = keras.models.load_model(
    "model_6_saved",
    custom_objects={'KerasLayer': hub.KerasLayer}
)

loaded_model_pred_probs = loaded_model_6.evaluate(val_sentences, val_labels)