TensorFlow重训练模型保存失败求助：加载正常但重训后保存出错

Hey there! Let's troubleshoot why your retrained TensorFlow model is refusing to save properly—especially weird since the initial save went off without a hitch. Here are the most likely issues and fixes to try out:

1. Did you modify the model structure after loading? #

If you added new layers, changed input/output shapes, or adjusted layer properties (like freezing/unfreezing) post-loading, this can break the model's serialization. TensorFlow’s SavedModel format relies on a consistent computation graph, so even small structural changes can cause save errors.

For example, if you added a new dense layer after loading, double-check that the input shape matches the previous layer’s output. Also, if you altered trainable flags, make sure you recompile the model before saving:

model = tf.keras.models.load_model("original_model")

# Example: Unfreeze top layers and add a new classifier
for layer in model.layers[:-5]:
    layer.trainable = False
model.add(tf.keras.layers.Dense(10, activation="softmax"))

# Critical: Recompile after structural changes
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=1e-5),
              loss="categorical_crossentropy")

# Now save
model.save("retrained_model")

2. Are custom layers/objects properly registered? #

If your original model used custom layers, loss functions, or metrics, you need to ensure these objects are serializable and registered when loading—and they stay accessible when saving. The most common mistake is forgetting to implement get_config() for custom layers, which TensorFlow needs to save/load the layer’s state.

Here’s how to fix it:

# Define your custom layer with get_config()
class CustomNormalization(tf.keras.layers.Layer):
    def __init__(self, mean=0.0, std=1.0, **kwargs):
        super().__init__(**kwargs)
        self.mean = mean
        self.std = std

    def call(self, inputs):
        return (inputs - self.mean) / self.std

    # Required for serialization
    def get_config(self):
        config = super().get_config()
        config.update({"mean": self.mean, "std": self.std})
        return config

# Load the model with custom_objects specified
model = tf.keras.models.load_model("original_model",
                                   custom_objects={"CustomNormalization": CustomNormalization})

# Now saving should work without issues
model.save("retrained_model")

3. Check save path permissions and existing files #

TensorFlow can throw errors if the target save path already exists (and you don’t allow overwriting) or if your user account lacks write permissions for the directory.

Try these fixes:

• Delete any existing directory/file with the same name as your target save path.
• Use overwrite=True to force overwriting (TF 2.x+):

  model.save("retrained_model", overwrite=True)
  

• Save as an HDF5 file instead (single file, avoids directory conflicts):

  model.save("retrained_model.h5", save_format="h5")
  

4. TensorFlow version mismatches #

If you saved the original model with one TF version and are loading/retraining with another, serialization incompatibilities can crop up. For example, TF 1.x SavedModels might have issues in TF 2.x, or newer TF versions might change how certain layers are serialized.

Fixes:

• Use the same TF version for saving, loading, and retraining.
• If migrating from TF 1.x to 2.x, try loading with compatibility mode:

  model = tf.compat.v1.keras.models.load_model("original_model")
  

5. In-place modifications breaking the computation graph #

Sometimes ad-hoc changes during retraining (like manually editing layer weights or altering the model’s input pipeline without updating the graph) can leave the model in an unserializable state.

Make sure:

• You don’t directly modify layer weights without using TensorFlow’s built-in methods.
• Any changes to the model’s input shape are reflected in the entire graph (e.g., if you resized images, update the input layer accordingly).

If none of these fix the issue, sharing the exact error traceback would help narrow down the problem further!

内容的提问来源于stack exchange，提问作者Mojo Jojo