search

WillianFuks / tfcausalimpact

Python Causal Impact Implementation Based on Google's R Package. Built using TensorFlow Probability.
Apache License 2.0
600 stars 72 forks source link

Error when replicating example #4

Open raffo85h opened 3 years ago

raffo85h commented 3 years ago

Hi there, thanks for this package and intro notebooks, all very informative!

I've tried to replicate an example using your dataset arma_data.csv. I get an error raised by tensorflow here it is: https://pastebin.com/SVP2yigy

Similar error when I replicate the random walk process example (in one of your notebook). This 2nd one tells me that tensorflow_probability seems not to cause problems (i.e. I can simulate data using tf-probability).

WillianFuks commented 3 years ago

Hi @raffo85h , thanks for opening this issue!

You're right, I've just confirmed here and what seems to be the cause is the newest version of tensorflow probability 0.12.0 (it was published yesterday a few days after I've tested the whole package using version 0.11).

For now, the best recommendation would be to downgrade to version 0.11.1:

pip install tensorflow_probability==0.11.1

I've opened this issue on their repository asking for guidance on what's happening (it's related to a bijector that tfcausalimpact uses for modeling the prior on local level), hopefully we'll get an assistance soon.

If you still have any issues please let me know.

Best,

Will

WillianFuks commented 3 years ago

Hi @raffo85h ,

I've just fixed this issue :)

You can just run pip install -U tfcausalimpact and the newer version will work just fine.

Thanks once again for opening this issue. Please let me know if this is solved for you before I close it.

Best,

Will

rpanai commented 3 years ago

Hi @WillianFuks do you mind to update the setup.py accordingly? As you might know (see this) I'm trying to build a conda package out of this. As today I have some spare time I was working on it but I received some errors.

WillianFuks commented 3 years ago

Hi @rpanai ,

Not sure what's wrong with setup.py, as it is today the PyPI API accepts it appropriately. If you want you can start a new issue reporting the errors you are getting and we work from there.

Best,

Will