Bug in keras.src.saving.saving_lib._save_model_to_dir

[MegaCreater]

tf.keras.__version__ -> "3.4.1"

If model is already saved then method call by keras.src.models.model.Model.save call keras.src.saving.saving_lib._save_model_to_dir, if model is already saved then asset_store = DiskIOStore(assert_dirpath, mode="w") (Line - 178) raise FileExistsError which error handling and finally clause line - asset_store.close() (Line - 189) causes - UnboundLocalError: local variable 'asset_store' referenced before assignment as asset_store is not define.

FileExistsError                           Traceback (most recent call last)
[/usr/local/lib/python3.10/dist-packages/keras/src/saving/saving_lib.py](https://localhost:8080/#) in _save_model_to_dir(model, dirpath, weights_format)
    139             )
--> 140         asset_store = DiskIOStore(assert_dirpath, mode="w")
    141         _save_state(


FileExistsError: [Errno 17] File exists: '/content/.../model_weights/assets'

During handling of the above exception, another exception occurred:

UnboundLocalError                         Traceback (most recent call last)
[/usr/local/lib/python3.10/dist-packages/keras/src/saving/saving_lib.py](https://localhost:8080/#) in _save_model_to_dir(model, dirpath, weights_format)
    148     finally:
    149         weights_store.close()
--> 150         asset_store.close()
    151 
    152 

UnboundLocalError: local variable 'asset_store' referenced before assignment

Solution to move asset_store.close() from finally clause to try clause or check if asset_store is define then only call asset_store.close() (Update from line 158 to line 189 i.e., https://github.com/keras-team/keras/blob/master/keras/src/saving/saving_lib.py#L158-L189)

def _save_model_to_dir(model, dirpath, weights_format):
    if not file_utils.exists(dirpath):
        file_utils.makedirs(dirpath)
    config_json, metadata_json = _serialize_model_as_json(model)
    with open(file_utils.join(dirpath, _METADATA_FILENAME), "w") as f:
        f.write(metadata_json)
    with open(file_utils.join(dirpath, _CONFIG_FILENAME), "w") as f:
        f.write(config_json)
    weights_filepath = file_utils.join(dirpath, _VARS_FNAME_H5)
    assert_dirpath = file_utils.join(dirpath, _ASSETS_DIRNAME)
    try:
        if weights_format == "h5":
            weights_store = H5IOStore(weights_filepath, mode="w")
        elif weights_format == "npz":
            weights_store = NpzIOStore(weights_filepath, mode="w")
        else:
            raise ValueError(
                "Unknown `weights_format` argument. "
                "Expected 'h5' or 'npz'. "
                f"Received: weights_format={weights_format}"
            )
        asset_store = DiskIOStore(assert_dirpath, mode="w")
        _save_state(
            model,
            weights_store=weights_store,
            assets_store=asset_store,
            inner_path="",
            visited_saveables=set(),
        )
    finally:
        weights_store.close()
        if ('asset_store' in locals()): asset_store.close()                 # check if `asset_store` define then only close 

[MegaCreater]

import tensorflow as tf

class SimpleModel(tf.keras.Model):
    def __init__(self):
        super(SimpleModel, self).__init__()
        # Define layers here
        self.dense1 = tf.keras.layers.Dense(10, activation='relu')
        self.dense2 = tf.keras.layers.Dense(1)  # Output layer with no activation (regression)

    def call(self, inputs):
        # Define forward pass
        x = self.dense1(inputs)
        return self.dense2(x)

# Instantiate the model
model = SimpleModel()

model.compile(
    optimizer='adam',
    loss='mean_squared_error',
    metrics=['accuracy']  # For classification tasks, change to appropriate metrics
)

model.save('sample_model',overwrite=True,zipped=False) # this will work 

model.save('sample_model',overwrite=True,zipped=False) # this will cause ERROR - as model is already saved. 

[Grvzard]

This issue has been fixed in the mainline (#19924, the commit just after the release of v3.4.1).

[ghsanti]

there is a misspelling here:

assert_dirpath = file_utils.join(dirpath, _ASSETS_DIRNAME)

but i shouldn't make a PR for that really, if anyone sees it, they can do it. (assert_dirpath should be asset_dirpath)
it was just slightly confusing upon reading it.

im curious what'd be use-cases for saving unzipped

[sachinprasadhs]

@MegaCreater , This is working fine with the latest Keras 3.5.0 version, attaching the Gist here
for reference, please verify and close the issue. Thanks

[MegaCreater]

@sachinprasadhs thanks. It works fine after update.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No