Document model debugging workflow

[andreasjansson]

The workflow for debugging Cog models isn't properly documented yet. You could just run cog predict [...] repeatedly, but that doesn't work if there are bugs due to state being carried over between predict() invocations.

The following is a workaround:

$ cog run python
>>> from predict import Predictor
>>> predictor = Predictor()
>>> predictor.setup()
>>> predictor.predict(arg1="value1", ...) # or if it's a progressively yielding model: next(predictor.predict())

Ideally the workflow should auto-reload, but that requires an ipython shell.

[andreasjansson]

The auto-reloading workflow could look like

$ cog run bash
# apt-get update && apt-get install -y ipython
# ipython --simple-prompt

In [1]: %load_ext autoreload
In [2]: %autoreload 2
In [3]: from predict import Predictor
In [4]: predictor = Predictor()
In [5]: predictor.setup()
In [6] predictor.predict(arg1="value1", ...) # or if it's a progressively yielding model: next(predictor.predict())

Now you can edit the source files and the edits will be visible to the interpreter immediately, without having to exit and run cog run python again.