Drop the bins entry from plot_param_updates' docstring

[VenishPaneliya]

SkorchDoctor.plot_param_updates documents a bins parameter it doesn't accept:

def plot_param_updates(self, match_fn=None, axes=None, figsize=None, **kwargs):

        bins : np.ndarray or None (default=None)
          Bins to use for the histogram. If left as ``None``, they are inferred
          from the data.

It's a line plot of relative parameter updates over time, not a histogram, so there are no bins to configure. The entry looks carried over from the methods that genuinely have it — plot_activations, plot_gradients, plot_activations_over_time and plot_gradient_over_time all take a real bins argument and use it for np.linspace(...). plot_param_updates is the only one of the six that documents it without having it.

Following the docstring doesn't just get ignored, it fails: bins falls through **kwargs into ax.plot(xvec, values, label=key, **kwargs), and matplotlib rejects it.

>>> ax.plot(np.arange(3), np.arange(3), label="x", bins=None)
AttributeError: Line2D.set() got an unexpected keyword argument 'bins'

Docs only — just the four lines removed, no code touched. skorch/tests/test_doctor.py passes (42 tests).

One thing I left alone: the **kwargs entry here (and in the sibling plot methods) describes figsize as something you override through **kwargs, even though figsize is an explicit parameter in each of those signatures. That's consistent across the methods, so it reads like a deliberate house phrasing rather than a mistake, and I didn't want to churn several docstrings on my own reading of it. Happy to tidy that up separately if you'd like it changed.

I also skipped a CHANGES.md entry since this doesn't change behaviour — let me know if you'd prefer one under Fixed.

[BenjaminBossan]

Thanks for fixing the docstring, including "bins" there was indeed a mistake.

One thing I left alone: the **kwargs entry here (and in the sibling plot methods) describes figsize as something you override through **kwargs, even though figsize is an explicit parameter in each of those signatures.

This was certainly not intentional, more of an oversight. I think it would make sense to mention another argument there instead of "figsize", something like "lw" (line width). If you're up to it, please update the docstrings there as well.

[VenishPaneliya]

Done — though lw turned out to only be safe in one of them, so I picked per method based on what each one already forwards.

plot_activations and plot_gradients pass lw explicitly into ax.hist (along with histtype, bins, density), so suggesting lw there would land users in the same trap as figsize:

>>> ax.hist(y, label=k, histtype='step', lw=2, bins=5, density=True, **{'lw': 3})
TypeError: Axes.hist() got multiple values for keyword argument 'lw'

So:

method	forwards to	already passes	example now used
plot_activations, plot_gradients	ax.hist	histtype, lw, bins, density	alpha
plot_param_updates	ax.plot	label	lw
plot_activations_over_time, plot_gradient_over_time	ax.fill_between	alpha, color, lw	linestyle

I checked each replacement actually goes through rather than assuming — alpha on hist, lw on plot and linestyle on fill_between all work, and alpha on fill_between raises the same duplicate-keyword error, which is why the two *_over_time methods got linestyle instead.

skorch/tests/test_doctor.py still passes (42 tests). The rewrap also happens to fix one pre-existing E501 on the old figsize line; the three flake8 warnings left in that file are on untouched lines and are present on master too.

[BenjaminBossan]

Thanks for the update. I went through the docstrings again and found that the arguments are still very inconsistently documented. Ideally, we would have documentation exactly for the arguments that are indeed present. So as an example, in plot_gradient_over_time, lw, figsize, and color are not documented. Is that something you'd like to tackle?

[VenishPaneliya]

Happy to — done. Every plot method now documents exactly the arguments it takes.

method	added
plot_loss	had no Parameters section at all — added ax, figsize, **kwargs, plus a Returns section to match its siblings
plot_activations	histtype, lw, density, figsize
plot_gradients	histtype, lw, density, figsize
plot_param_updates	figsize
plot_activations_over_time	lw, figsize, color
plot_gradient_over_time	lw, figsize, color

I checked it both directions rather than by eye — comparing each signature against its parsed Parameters section, there are now no missing entries and no documented-but-absent ones across all six. numpydoc also parses each of them with zero warnings, and the parameter counts line up with the signatures (e.g. plot_activations → 9).

One small wording note: I described figsize as "only used when a new plot is created, i.e. when axes is None", since _get_axes returns the passed axes untouched and ignores figsize in that case — that seemed worth stating explicitly, but say the word if you'd rather keep it shorter.

skorch/tests/test_doctor.py passes (42 tests), and flake8 output on the file is unchanged from master.

[VenishPaneliya]

Note on the red CI: the failures are the model-convergence assertions (assert 0.71 > 0.75, assert np.float64(7.85) < 1.0), and they show up across Python 3.10 / 3.11 / 3.13 and several torch builds in the same run. This PR changes only docstring text in skorch/_doctor.py (1 file, no code lines), so it can't affect training accuracy — looks like flakiness rather than anything from here. Happy to rebase if you'd like a clean run.

[BenjaminBossan]

Thanks for bringing all the docstrings in line, LGTM.