(udc_test.py) Incompatible shapes: [80,1] vs. [160,1] error even after changing the batch size

[dennybritz]

(udc_test.py) Incompatible shapes: [80,1] vs. [160,1] error even after changing the batch size

[Sangheli]

if move in udc_model.py

batch_size = targets.get_shape().as_list()[0]

to

if mode == tf.contrib.learn.ModeKeys.EVAL:
_____batch_size = targets.get_shape().as_list()[0]

code work correctly on tf 0.12 in predict mode, but you need change call of predict function, as in example https://github.com/Sangheli/chatbot-retrieval/blob/master/udc_predict.py#L68
without flag as_iterable

[jugador]

In my case I changed both.
In udc_hparams.py
tf.flags.DEFINE_integer("batch_size", 64, "Batch size during training") tf.flags.DEFINE_integer("eval_batch_size", 8, "Batch size during evaluation")
In udc_test.py
tf.flags.DEFINE_integer("test_batch_size", 8, "Batch size for testing")

After that it started working. I had to rerun the train script. Also I changed my tf to version .11

[Sangheli]

for udc_test.py you can set test_batch_size to 8, as a batch_size in udc_hparams.py
https://github.com/dennybritz/chatbot-retrieval/blob/master/udc_test.py#L15

[rtbins]

Sangheli thanks, it worked, but now while running udc_predict.py am getting error
File "/usr/local/lib/python3.5/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 1047, in _call_model_fn model_fn_results = self._model_fn(features, labels, mode=mode) File "/home/rohitg/Documents/chatbot-retrieval-master/udc_model.py", line 29, in model_fn batch_size = targets.get_shape().as_list()[0] AttributeError: 'dict' object has no attribute 'get_shape'

[Sangheli]

wich version of tensorflow you use? in tensorflow 0.11 all works correct (python 3.5)

[rtbins]

The version am using is 0.12.0 and python 3.5

[Sangheli]

i have some problems while running this code in tf 0.12 or 0.10

[rtbins]

Do I need to switch to .11?

[Sangheli]

yes, this helped me

[rtbins]

Do I need to train data again or I can use the same output data there?

[Sangheli]

I don't know, I'm trained every time I trying to solve problems, on gt 770 4gb train time about 1-2 hours(16 000 steps), to get perfect results on ubuntu text corpus, in other issues author says he train 2000 steps, to get good result

[rtbins]

Thanks for helping me out :D

[rtbins]

Cool it worked with .11, I wonder what is the prob with .12

[basma-b]

@Sangheli please I am getting the same error on training, should I change batch_size or eval_batch_size in udc_hparams.py to 8 and test_batch_size in udc_test.py to 8 too ?

[hunkim]

Can any of you make a pull request to fix this issue?

[yogesh-0586]

udc_predict.py return only POTENTIAL_RESPONSES zero position value