Recently several people have asked how to debug AllenNLP code using their IDEs. Here is a guide how to do so in both PyCharm and VSCode.
If you have a different preferred debugger, these instructions will hopefully get you most of the way there.
We'll demonstrate the debugger using the Academic Paper Classifier model from our "AllenNLP-as-a-Library" example.
If you'd like to follow along, clone that repo and install its requirements.
Our recommended workflow is using our command-line tool allennlp.
The example repo contains the training command:
allennlp train experiments/venue_classifier.json -s /tmp/your_output_dir_here --include-package my_library -o '{"trainer": {"cuda_device": -1}}'
(I added an override to train on the CPU, since the machine you're running PyCharm on probably doesn't have a GPU.)
After which you can select "Run > Attach to Local Process",
search for the one that's running allennlp,
and get results in the debugger:
Some people find the "attach to local process" workflow to be a nuisance, especially if you are repeatedly debugging things.
Fortunately, it's easy to launch the training process from code,
using allennlp.commands.main().
First, create a wrapper script that puts your command-line args into sys.argv and calls main(). In the academic paper classifier example, I could use the following:
import json
import shutil
import sys
from allennlp.commands import main
config_file = "experiments/venue_classifier.json"
# Use overrides to train on CPU.
overrides = json.dumps({"trainer": {"cuda_device": -1}})
serialization_dir = "/tmp/debugger_train"
# Training will fail if the serialization directory already
# has stuff in it. If you are running the same training loop
# over and over again for debugging purposes, it will.
# Hence we wipe it out in advance.
# BE VERY CAREFUL NOT TO DO THIS FOR ACTUAL TRAINING!
shutil.rmtree(serialization_dir, ignore_errors=True)
# Assemble the command into sys.argv
sys.argv = [
"allennlp", # command name, not used by main
"train",
config_file,
"-s", serialization_dir,
"--include-package", "my_library",
"-o", overrides,
]
main()Then you can set a breakpoint somewhere in your code:
Go back to the wrapper script, right-click in the editor and choose "Debug":
And then once you hit your breakpoint, you can inspect variables or step through code to your heart's content:
You can use the same debug wrapper script as in the previous section.
Then click on the debug icon all the way on the left,
select "Python > Current File" as the debugging profile,
set a breakpoint somewhere,
go back to the wrapper script, and select "Debug > Start Debugging" from the menu.
Your code will run in the integrated terminal, and when you hit your breakpoint you'll be able to inspect variables
and the call stack
and whatever else you need to do to debug your code.
Happy debugging!