**if predictions.get_shape().ndims > 1: TypeError: '>' not supported between instances of 'NoneType' and 'int'** about laco

[kadruko]

@yy-mille
The method get_shape() on predictions returns , because shape was not yet initialized in this training phase.
Therefore ndims equals None.

To resolve this issue, you could choose a different version of tensorflow, but I could not find any version, that fits to the whole repository - other errors occured

The best solution is to manually initialize the shape of the predictions by setting the predictions shape to the number of inputs and label number:
token_label_predictions.set_shape([num_token_labels, num_labels])
This line of code needs to be placed at the begin of the definition of the metric_fn().

[yy-mille]

以上是使用tf1.14存在的问题。tf1.13无experimental, tf1.12 cuda9目前无法使用。

[Sweetsmile123]

This error is reported at the end when I run the first program file "train_main_plcp.py".
Have you encountered the following errors when reproducing?

INFO:tensorflow:*** Features ***
INFO:tensorflow: name = fit_docspace_positions, shape = (?, 266)
INFO:tensorflow: name = fit_labelspace_positions, shape = (?, 54)
INFO:tensorflow: name = input_ids, shape = (?, 320)
INFO:tensorflow: name = input_mask, shape = (?, 320)
INFO:tensorflow: name = is_real_example, shape = (?,)
INFO:tensorflow: name = label_ids, shape = (?, 54)
INFO:tensorflow: name = pair, shape = (?, 2)
INFO:tensorflow: name = pair_target, shape = (?, 2)
INFO:tensorflow: name = segment_ids, shape = (?, 320)
INFO:tensorflow: name = token_label_ids, shape = (?, 320)

[EiraZhang]

Hi, I haven't encountered the similar problems. Please modify the code according to tf version.

[wuyw]

以上是使用tf1.14存在的问题。tf1.13无experimental, tf1.12 cuda9目前无法使用。

https://zhuanlan.zhihu.com/p/268735650