Andrew Audibert
Andrew Audibert
@ematejska PTAL
Reassigning since this appears to be a TF filesystem issue
Keras may want to make autosharding configurable for non-dataset inputs. Assigning to @fchollet for triage
The problem is that the shape passed to `unbatch` is unknown. If `unbatch` can't determine the shape of its inputs, it can't determine its cardinality. ``` ds = tf.data.Dataset.range(10) #...
@bhack I'd prefer to keep the behavior consistent between eager and graph mode, so that putting a @tf.function around eager code won't change the result of cardinality/__len__.
Ah I see, in that case I think your suggestion makes sense for `len`.
From the stack traces, it looks like the program hangs while waiting for the Python generator function to complete. It isn't clear whether the Python function is hanging, or if...
@gadagashwini I haven't been able to reproduce the hang. Can you share your repro methodology?
> @aaudiber, I could replicate the issue with Tf-gpu 2.1 on colab. > Please find the gist [here](https://colab.sandbox.google.com/gist/gadagashwini/b74f86d43ac19829827b59b3a149f876/untitled410.ipynb). At the end session stopped by saying `Your session crashed after using...
One idea is to add print statements to `random_image_generator` to see whether the function hangs.