Bidirectional LTSM example throws shape error

[colinskow]

Running the following file throws an error:
https://github.com/tflearn/tflearn/blob/master/examples/nlp/bidirectional_lstm.py

ValueError: Shape (128, ?) must have rank at least 3

Setup:

• MacOS Sierra (10.12)
• Python 2.7
• Tensorflow v1.2.0
• TFLearn v0.3.2

Traceback (most recent call last):
  File "bidirectional_lstm.py", line 47, in <module>
    net = bidirectional_rnn(net, BasicLSTMCell(128), BasicLSTMCell(128))
  File "/Users/colin/tensorflow/lib/python2.7/site-packages/tflearn/layers/recurrent.py", line 374, in bidirectional_rnn
    dtype=tf.float32)
  File "/Users/colin/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 375, in bidirectional_dynamic_rnn
    time_major=time_major, scope=fw_scope)
  File "/Users/colin/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 574, in dynamic_rnn
    dtype=dtype)
  File "/Users/colin/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 637, in _dynamic_rnn_loop
    for input_ in flat_input)
  File "/Users/colin/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 637, in <genexpr>
    for input_ in flat_input)
  File "/Users/colin/tensorflow/lib/python2.7/site-packages/tensorflow/python/framework/tensor_shape.py", line 649, in with_rank_at_least
    raise ValueError("Shape %s must have rank at least %d" % (self, rank))
ValueError: Shape (128, ?) must have rank at least 3

[willduan]

Please try to install tensorflow 1.0 or 1.1. Since it is not very compatible with tensorflow 1.2 version currently.

[colinskow]

I can confirm this does work fine with TensorFlow 1.1.0. That means there was a breaking change in tensorflow.python.ops.rnn.bidirectional_dynamic_rnn. I would like to get this working in the latest version since I am relying on new 1.2.0 features.

[colinskow]

I found the problem:
On TF <= 1.1.0 recurrent.py imports static_bidirectional_rnn which works.
On TF >= 1.2.0 recurrent.py imports bidirectional_dynamic_rnn which is not compatible with the current implementation.

Removing the import for bidirectional_dynamic_rnn solves it. PR inbound.

[stepsma]

Hi @colinskow

Could you please re-visit this problem one more time, I am currently still getting this error with tensorflow 1.4. Other people seems seeing this problem as well with more recent tensorflow version (988 issue). Thank you for your concern.

Best,

[FTAsr]

I'm getting the same error using the following packages on a linux GPU machine, a conda environment with python 2.7 (same code gives no error on my MacBook with tensorflow cpu):

tensorflow 1.3.0 0
tensorflow-base 1.3.0 py27h0dbb4d0_1
tensorflow-gpu 1.3.0 0
tensorflow-gpu-base 1.3.0 py27cuda8.0cudnn6.0_1
tensorflow-tensorboard 0.1.5 py27_0
tflearn 0.3.2

I was initially using Keras together with the same Tensorflow-gpu package and got segmentation fault trying to add any type of recurrent layers, and then moved to using tflearn but again having problem with this type of network. Please fix the problem in all tensorflow distributions. I wonder why the GPU version is always one step behind the CPU version. Thank you!

[MrKZZ]

Oh, I have this problem too! Seems that this problem hasn't fixed so far.Neither tensorflow or tflearn team wants to do this job! But I wander to know what's the problem means?
ValueError: Shape (128, ?) must have rank at least 3

[fiorinin]

Same problem here, some insight would be nice.

[thomas-young-2013]

same problem found! I am working with tensorflow_gpu 1.8.0, please revisit the issue again!

[maxiaomu]

Oh, i've found the same problem with tf.version
'1.6.0-rc0'.

[signalprime]

aye, here too... tensorflow 1.8.0

[g-laz77]

Same issue again. Please revisit this!

[sandove]

same error... tensorflow 1.9.0

[dyq0811]

same error here!...

[zhvankh]

d01a0b9#diff-5885c0e8432a40cf7a3e4b2448302e3a
this is best answer for this problem

[Vani261196]

same error here ....... tensorflow 1.14

[Syauri]

Getting the same error in Tensorflow 1.15! Please revisit!

[yunfanye]

FWIW, I met this issue today too and it turned out it was a different root cause that led to the same error message. Given that there was a comment saying it was fixed in TF >= 1.2, I'd bet that the issue was indeed fixed and all the later errors were caused by other issues.

It was so frustrating initially to see all the comments here complaining a issue 5 years ago without answers, so I figured I'd shed some light.

The error message basically says: the code expects a 3-dimensional input, i.e. (128, ?, ?) instead of (128, ?). For my specific error, it was caused by an upstream bug where one dimension was gone, i.e. it should have been (128, 256, ?).