I don't have a specific problem at the moment, but it keeps coming up that I have a bug in my loss function and the error printouts are not sufficient to localize the problem to a specific line of code. For example expected 'int32' but got 'float32' or something like that. Is there a way to know which line of code in the loss function is the source of the problem?
I'll note that sometimes the error comes during compilation, in which case print statements have been helpful. But I have not identified a way to find the problem (outside of guessing or commenting out sections) if it happens only during training, since printouts are not displayed.
You can use one of following option to debug custom loss function or any other deep learning works.
tf.print
tfbdg
TensorBoard debugger
Related
I am new to Tensorflow and am currently working on Tensorflow2. I'm still having a hard time writing code, because I don't have the possibility to debug.
I already tried to get further with the line:
tf.executing_eagerly()
and
tf.print()
but this is only a small help compared to the "normal" debugging in python.
Is there a better possibility to debug the code and to view the content of variables?
The only thing I currently get is this view, but that doesn't give me any insight into the actual variables either:
Rather than using tf.print, use the normal Python's print if you are eagerly executing. You will be able to see the contents of the variables.
I state that I am a beginner, here is my problem.
My model works perfectly when I run it with a normal simulation. Now I'm trying to optimize some parameters using the optimization experiment, I've followed all the steps of the official tutorial, but it doesn't work because I get "Exception during discrete event execution:
Truncated class file". The strange thing is that, looking into the console displaying the error, I see that some lines are referred to an old version of my model, for example:
java.lang.ClassFormatError: Truncated class file
at coffe_maker.Main._m1_1_delayTime_xjal(Main.java:14070)
The current model's name is coffee_maker_v2_6 so I don't understand why I get this kind of error, do you know if it is normal? What am I doing wrong?
The most likely cause is that you have Java code left in an 'unused' configuration of a Delay block's "Delay time" expression (e.g., it now has a static value but you had Java code in the now-switched-out dynamic value).
Unfortunately, AnyLogic sometimes still includes the switched-out code in the compiled class, and this can sometimes cause strange runtime errors such as that one.
If this does look to be the case, temporarily switch to the offending switched-out configuration and delete it before switching back to the correct one.
I have resolved: the problem was that, in every delay block of my model, the delay time was linked with a database reference (type code), now I am trying by tyiping the probability distributions in the delays directly and now the optimization works
When I run a prediction with saved_model_cli it executes properly, but the same inputs give me an error when I try to run it through tensorflow-serving.
It tells me that an item is being both fed and fetched (because it is in the inputs and the outputs) but it works fine from the command line. Does anybody know why it works with one and not the other or how to go about fixing it?
Looks similar to the issue mentioned in the below link :
Placeholder_2:0 is both fed and fetched
Please refer to this and see if it resolves your issue. If not, please share the code snippet to further debug the issue.
I'm using scipy.ndimage.zoom and I get this annoying warning:
UserWarning: From scipy 0.13.0, the output shape of zoom() is calculated with round() instead of int() - for these inputs the size of the returned array has changed.
I'm not sure what I should get from it, I started using it with SciPy 1.0.0 so I don't believe it really affects me.
I guess calling it UserWarning is a bit questionable given it's not intended for user consumption, but maybe the intended user is the developer importing the library.
I'm using multiprocessing and I get one warning per process, even more annoying.
Is there a sane way to silent it?
It was easier than I thought, leaving the question for future reference in case anyone needs this.
import warnings
warnings.filterwarnings('ignore', '.*output shape of zoom.*')
Your proposed solution did not work for me. But what does work is:
import warnings
# other nice code
with warnings.catch_warnings():
warnings.simplefilter("ignore")
x = scipy.ndimage.interpolation.zoom(...)
It may be a basic question, but how can I view the log of the solver after the model is optimized. I came across the option to hide the outputs by specifying model.hideOutput(). Is there an option to show the outputs (by default, I don't see any output log. I am using Spyder IDE).
Printing the SCIP output is activated by default in PySCIPOpt. As you already wrote, you would have to manually disable it by calling hideOutput(). Have you tried running your code in a terminal outside of Spyder? Is the model even solved, so did you check whether there is a solution in the end?
To complete the other answer: if your PySCIPOpt model has been created by third-party code, it is possible the output has been turned off upstream.
You can still call model.hideOutput(False) -- see the documentation of hideOutput().