AttributeError: 'GQNTFRecordDataset' object has no attribute '_graph_attr'

[curryJ]

Hello, I am trying to use your code to perform training again but I got this error
I have downloaded the dataset and place it properly. And the dataset I used is mazes.
The version of my tensorflow is 1.13.1.
Thanks a lot

I change the tensorflow's version to 1.12.0,but another error occurres that 'ValueError: Dimension 1 in both shapes must be equal, but are 84 and 64. Shapes are [?,84,84] and [?,64,64]. for 'GQN/GQN_RNN/Inference/LSTM_inf/concat' (op: 'ConcatV2') with input shapes: [?,84,84,3], [?,64,64,256], [] and with computed input tensors: input[2] = <-1>.'
I think the reason is that the function tf.concat([query_image, canvas], axis=-1),

[ogroth]

Hi @curryJ, sorry for the long silence, I've been busy with other projects. I'm currently in the middle of re-writing the input data pipeline (see commits on the dev-branch) and will get rid of the deprecated data loading functionality used in DeepMind's original release. Stay tuned, the new input pipeline will be merged within the week. I aim for tensorflow 1.12.1. Caveat: I've never tried the maze or jaco datasets provided by DeepMind. But for the sake of completeness, I'd like to add them at some point as well. Feel free to open separate issues for those datasets, but maybe wait until the new data loader is online.
Thanks,
Oliver

[ogroth]

Hi @curryJ , the new input_fn should be able to handle all GQN datasets. Also, both training script and input_fn now come with img_size parameters to resize images from different datasets. Let me know, if you still run into issues. If not, I'll close this issue.