onnx-tensorflow
onnx-tensorflow copied to clipboard
GRU support with linear_before_reset
Similar to #227 but without adding the backend, uses the CudnnCompatibleGRUCell implementation in TensorFlow instead.
Looks like https://www.tensorflow.org/api_docs/python/tf/contrib/cudnn_rnn/CudnnCompatibleGRUCell is deprecated. In general, we should avoid using APIs in contrib. We would need to replace it soon anyway so could you find alternative ways?
It looks like the new GRU layer is CuDNN compatible and supports reset_after
so I think it should just be a case of swapping tf.contrib.cudnn_rnn.CudnnCompatibleGRUCell
for tf.keras.layers.GRU
. I will have a proper look tomorrow.
Needs #460 fixing first.
Yeah, we will need to revisit this once we get to TF 2.0
Hi, does this feature still on progress? I met this problem when I transfer my ONNX model. I'd like to fix this, if there's no implement yet
Hi, does this feature still on progress? I met this problem when I transfer my ONNX model. I'd like to fix this, if there's no implement yet
Any update on this? I am also facing the same issue while converting ONNX to tensorflow.
The patch I submitted worked for me at time submission but I don't have time to (re)evaluate currently @alamnasim @dianyo give it a try.
Any update on this?
@iiSeymour, when I use your gru-reset-after branch,I found the value of the argument 'name' in function _custom_getter() which defined in gru.py is 'GRU_9c8f1236/rnn/multi_rnn_cell/cell_0/cudnn_compatible_gru_cell/candidate/hidden_projection/bias'. Can you tell me how is the argement name generated, I have not found it in the context, thanks.