Output Not Captured about reticulate HOT 16 CLOSED

Seems like probably an issue regarding how TF tells whether RStudio is an interactive environment. https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/platform/tf_logging.py#L55
Perhaps we can also have a function for capturing standard error which is where logs in non-interactive environment go to.

from reticulate.

Yes, py_capture_output only captures output. We do however already have some hooks to interact with the TF logging system (we use them in py_suppress_warnings. If you let me know exactly the behavior you want here (perhaps with a simple repro) I can make some combination of py_capture_output and py_suppress_warnings make it happen.

from reticulate.

Oh I see. Yea I am expecting the following as part of the output:

INFO:tensorflow:Create CheckpointSaverHook.
INFO:tensorflow:Saving checkpoints for 2 into /var/folders/ln/c7r_x26x46b7vzpftbfh64zc0000gp/T//RtmpbD66eR/tflearn_a32c5e4d2404/model.ckpt.
INFO:tensorflow:loss = 438.822, step = 2
INFO:tensorflow:global_step = 2
INFO:tensorflow:global_step = 6 (0.016 sec)
INFO:tensorflow:global_step = 10 (0.002 sec)

from reticulate.

I think the problem may be here:

https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/platform/tf_logging.py#L39-L56

You can that TF is trying to determine whether the session is interactive. In the case that it is it will pipe output to stdout (which we could then capture) however in the case where it's not it sends it to stderr. That relies on the state of sys.flags.interactive which we maybe able to set by passing the correct command line flag to startup.

from reticulate.

Okay, the function is now renamed py_capture_output. By default it captures both stdout and stderr but it can be configured to capture just one or the other:

2071de8

This should get you what you are looking for, let me know if it doesn't.

from reticulate.

Thanks, will try!

from reticulate.

Closing this. Feel free to re-open if you have additional issues.

from reticulate.

Just gave it a try. Still looks like not able to capture those output. Not a blocker though.

from reticulate.

from reticulate.

Here it is:

> output <- reticulate::py_capture_output(tf$logging$warn("asdf"))
WARNING:tensorflow:asdf
> output
[1] ""

from reticulate.

Interesting, for me this actually successfully captures the output! Here's my exact code as executed (I think it's the same):

library(tensorflow)
output <- reticulate::py_capture_output(tf$logging$warn("asdf"))

Are you certain you've upgraded to the latest versions of tensorflow and reticulate?

from reticulate.

Yes, I've removed all packages and reinstalled everything. Otherwise py_capture_output won't be available in old versions.

from reticulate.

from reticulate.

Okay apologize that I didn't clean up my environment properly. That one worked as expected. However, here's what's weird:

> library(tensorflow)
> tf$logging$get_verbosity()
[1] 30
> tf$logging$WARN
[1] 30
> tf$logging$set_verbosity(tf$logging$INFO)
> tf$logging$get_verbosity()
[1] 20
> output <- reticulate::py_capture_output(tf$logging$info("asdf"))
INFO:tensorflow:asdf
> output
[1] ""

Verbosity is set to info correctly but the output is not captured.

Could you try if you can reproduce it?

from reticulate.

It looks to me like the INFO stream is bypassing stderr. If I write this then the output is correctly captured:

output <- reticulate::py_capture_output({
  sys <- import("sys")
  sys$stderr$write("foo")  
})

Are you just trying to prevent INFO messages from being printed at all? If so we could do something like we do for py_suppress_warnings:

https://github.com/rstudio/tensorflow/blob/master/R/package.R#L35-L45

from reticulate.

I am trying to capture a regex in the output. But it's not super important. I think this is an issue on the Python API end. I'll reopen if this turns out to be very critical.

from reticulate.