Keras: ValueError: Unable to create group (Name already exists) with model.save_weights()

Created on 3 Feb 2019 · 15Comments · Source: keras-team/keras

This is a similar issue to https://github.com/keras-team/keras/issues/6005 but I believe it is caused by the way h5py defines groups. In particular, if a layer named foo is in a network after a layer named foo/bar, h5py throws an exception. But the same does not occur if foo comes first. To reproduce, see the snippet below.

from keras import layers, models

# This raises an exception.
input_layer = layers.Input((None, None, 3), name='test_input')
x = layers.Conv2D(1, 1, name='conv1/conv')(input_layer)
x = layers.BatchNormalization(name='conv1/bn')(x)
x = layers.Activation('relu', name='conv1')(x)
models.Model(inputs=input_layer, outputs=x).save_weights('test.h5')

# This doesn't raise an exception
input_layer = layers.Input((None, None, 3), name='test_input')
x = layers.Conv2D(1, 1, name='conv1')(input_layer)
x = layers.BatchNormalization(name='conv1/bn')(x)
x = layers.Activation('relu', name='conv1/relu')(x)
models.Model(inputs=input_layer, outputs=x).save_weights('test.h5')

Perhaps we could provide a more helpful error message in keras/engine/saving.py? For example, changing part of save_weights_to_hdf5_group to the following would help trace the offending layer name.

for layer in layers:
    try:
         g = group.create_group(layer.name)
    except ValueError:
         raise ValueError('An error occurred creating weights group for {0}.'.format(layer.name))
    symbolic_weights = layer.weights
    weight_values = K.batch_get_value(symbolic_weights)

Happy to create PR if this is helpful.

buperformance

Source

faustomorales

👍4 👀2

Most helpful comment

Hello,
I also have this issue with tensorflow 2.0.0
If some of you still want to use the .h5 format, I've found a potential fix.
Since the problem lays in the order of creation of the h5py groups: a group name can't be prefix of a previous group name, it is possible to sort the layers by name before saving them.
This change worked for me :
File: $CUSTOM_PATH/tensorflow_core/python/keras/saving/hdf5_format.py
Function: save_weights_to_hdf5_group

  for layer in layers:
    g = f.create_group(layer.name)
    weights = _legacy_weights(layer)
    weight_values = K.batch_get_value(weights)
    weight_names = [w.name.encode('utf8') for w in weights]
    save_attributes_to_hdf5_group(g, 'weight_names', weight_names)
    for name, val in zip(weight_names, weight_values):
      param_dset = g.create_dataset(name, val.shape, dtype=val.dtype)
      if not val.shape:
        # scalar
        param_dset[()] = val
      else:
        param_dset[:] = val

replaced by:

  sorted_layers = [(layers[i].name, i) for i in range(len(layers))]
  sorted_layers.sort()

  for sorted_layer_index in range(len(sorted_layers)):
    layer = layers[sorted_layers[sorted_layer_index][1]]
    g = f.create_group(layer.name)
    weights = _legacy_weights(layer)
    weight_values = K.batch_get_value(weights)
    weight_names = [w.name.encode('utf8') for w in weights]
    save_attributes_to_hdf5_group(g, 'weight_names', weight_names)
    for name, val in zip(weight_names, weight_values):
      param_dset = g.create_dataset(name, val.shape, dtype=val.dtype)
      if not val.shape:
        # scalar
        param_dset[()] = val
      else:
        param_dset[:] = val

With this modification I was able to save my model in .h5 format and then to load my model from scratch and run inferences.
I can do a pull request if you think it's a good idea

Tbuhet on 21 Oct 2019

👍10 🎉8 🚀2

All 15 comments

I don't really know what we should do in this case. But that's definitely an issue. @farizrahman4u what should we do about this?

gabrieldemarmiesse on 3 Feb 2019

Will get back to this and the pyux PR on Saturday.

farizrahman4u on 4 Feb 2019

I'm experiencing the same issue. Any updates on this?

dorian821 on 22 Aug 2019

I am also facing the same issue, Any updates on it?

lucky-suman on 6 Sep 2019

ghonk on 9 Sep 2019

I met the same error, I solved it by saving the model with .tf instead of .h5
Plus, I am using TensorFlow 2.0, the default saving format is .tf

Salingo on 12 Sep 2019

👍13 🎉2

This issue has a larger effect than suggested in the top of the thread. It actually completely prevents saving the weights of a keras model that uses tf.ones_like.

in_layer = Input((None, None, 3), name="test_input")
ones = K.ones_like(in_layer)
model = Model(inputs=in_layer, outputs=ones)

model.save_weights("tmp/test_save_ones.keras")

Raises: ValueError: Unable to create group (name already exists)

This is because of the issue described above by @faustomorales and can be seen if we look at how ones_like gets added to the keras model:

[(l, l.name) for l in model.layers]

Output:

[(<tensorflow.python.keras.engine.input_layer.InputLayer at 0x7fb6e9fa53c8>,
  'test_input'),
 (<tensorflow.python.keras.engine.base_layer.TensorFlowOpLayer at 0x7fb6e9fa56d8>,
  'tf_op_layer_ones_like_2/Shape'),
 (<tensorflow.python.keras.engine.base_layer.TensorFlowOpLayer at 0x7fb6e9fa59e8>,
  'tf_op_layer_ones_like_2')]

I haven't looked around for other places where this is happening but I assume that there are more cases where this causes problems. Given that this effectively breaks a core part of the Keras API in tensorflow--the ability to save certain models in h5 format--I'd argue that this needs a more serious fix than just an error message.

ELind77 on 30 Sep 2019

👍1

hello, I also get trouble with this issue.
If you guys are using tensorflow 2.0, you can change ".h5" to ".tf" and everything should be saved.

cuongth95 on 11 Oct 2019

  for layer in layers:
    g = f.create_group(layer.name)
    weights = _legacy_weights(layer)
    weight_values = K.batch_get_value(weights)
    weight_names = [w.name.encode('utf8') for w in weights]
    save_attributes_to_hdf5_group(g, 'weight_names', weight_names)
    for name, val in zip(weight_names, weight_values):
      param_dset = g.create_dataset(name, val.shape, dtype=val.dtype)
      if not val.shape:
        # scalar
        param_dset[()] = val
      else:
        param_dset[:] = val

replaced by:

  sorted_layers = [(layers[i].name, i) for i in range(len(layers))]
  sorted_layers.sort()

  for sorted_layer_index in range(len(sorted_layers)):
    layer = layers[sorted_layers[sorted_layer_index][1]]
    g = f.create_group(layer.name)
    weights = _legacy_weights(layer)
    weight_values = K.batch_get_value(weights)
    weight_names = [w.name.encode('utf8') for w in weights]
    save_attributes_to_hdf5_group(g, 'weight_names', weight_names)
    for name, val in zip(weight_names, weight_values):
      param_dset = g.create_dataset(name, val.shape, dtype=val.dtype)
      if not val.shape:
        # scalar
        param_dset[()] = val
      else:
        param_dset[:] = val

With this modification I was able to save my model in .h5 format and then to load my model from scratch and run inferences.
I can do a pull request if you think it's a good idea

Tbuhet on 21 Oct 2019

👍10 🎉8 🚀2

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

tensorflow-butler[bot] on 22 Oct 2019

wow! I seconded @Salingo! By changing weight file name from .h5 to .tf it should work! No idea why it is the case, though!
My case: Tensorflow Keras (version 1.14)

hoangcuong2011 on 8 Dec 2019

This issue has a larger effect than suggested in the top of the thread. It actually completely prevents saving the weights of a keras model that uses tf.ones_like.
in_layer = Input((None, None, 3), name="test_input")
ones = K.ones_like(in_layer)
model = Model(inputs=in_layer, outputs=ones)

model.save_weights("tmp/test_save_ones.keras")
Raises: ValueError: Unable to create group (name already exists)

This is because of the issue described above by @faustomorales and can be seen if we look at how ones_like gets added to the keras model:
[(l, l.name) for l in model.layers]
Output:
[(<tensorflow.python.keras.engine.input_layer.InputLayer at 0x7fb6e9fa53c8>,
  'test_input'),
 (<tensorflow.python.keras.engine.base_layer.TensorFlowOpLayer at 0x7fb6e9fa56d8>,
  'tf_op_layer_ones_like_2/Shape'),
 (<tensorflow.python.keras.engine.base_layer.TensorFlowOpLayer at 0x7fb6e9fa59e8>,
  'tf_op_layer_ones_like_2')]
I haven't looked around for other places where this is happening but I assume that there are more cases where this causes problems. Given that this effectively breaks a core part of the Keras API in tensorflow--the ability to save certain models in h5 format--I'd argue that this needs a more serious fix than just an error message.

One walk around is to use the Lambda layer to wrap these type of operations

ysyyork on 9 Apr 2020

👍1

I'm still getting this issue no matter what I try. Any ideas on what I'm doing wrong? I'm trying to save the YAM net model so I can then convert it to tensorflow.js.

      cls._yamnet = yamnet.yamnet_frames_model(params)
      cls._yamnet.load_weights('yamnet.h5')
      cls._yamnet_classes = yamnet.class_names('yamnet_class_map.csv')
      YAMNetTest._yamnet.save('./foo/yamnet.h5') # Unable to create group (name already exists)
    # YAMNetTest._yamnet.save('./foo/yamnet') # 'Tensor' object has no attribute '_datatype_enum'
    # tfjs.converters.save_keras_model(YAMNetTest._yamnet, './foo') # Unable to create group (name already exists)
    # YAMNetTest._yamnet.save('./foo/foo.tf') # 'Tensor' object has no attribute '_datatype_enum'
    # tf.saved_model.save(YAMNetTest._yamnet, './foo') # 'Tensor' object has no attribute '_datatype_enum'

Versions:

tensorboard          2.1.1      
tensorflow           2.1.0      
tensorflow-cpu       2.1.0      
tensorflow-estimator 2.1.0      
tensorflow-hub       0.7.0      
tensorflowjs         1.7.4