Tflearn: BiRNN bug

[ruizhaogit]

Hi TFlearn,

There is a small bug in BiRNN:

WARNING:tensorflow:Error encountered when serializing layer_tensor/BiRNN.
Type is unsupported, or the types of the items don't match field type in CollectionDef.
'list' object has no attribute 'name'.

Best,
Rui

[tgallice]

This issue come from the model serialization. On serialization, tensorflow try to get the name property for a list or dict value.

There are severals places where a list or dict is added as collection value :

• https://github.com/tflearn/tflearn/blob/master/tflearn/layers/recurrent.py#L92
• https://github.com/tflearn/tflearn/blob/master/tflearn/layers/recurrent.py#L400
• https://github.com/tflearn/tflearn/blob/master/tflearn/summaries.py#L51

Here tensorflow call the name property which is good for Tensor but not for list or dict. But the _get_kind_type does not make the distinction between list or dict and Tensor.

We can maybe make a workaround with tf.identity ? @aymericdamien what do you think ? It's a misuse in tflearn or an issue for tensorflow ?

[aymericdamien]

Yes, the collections that have dicts (such as summaries) are deleted before being saved into the model file.
I think this is an error linked with tensorflow... before it was working well (you could save dicts without problems) but when they used protobuf, it got issues. But, I do not think TensorFlow will support dict any time soon, so we should find a way to fix that in TFLearn.

[ssriramana93]

Hello, it's been a year has this been fixed?

[saisumit]

has it been resolved ?