dp_augmentor update #130
dp_augmentor update #130
Conversation
…re settings. Implemented compatibility with multichannel labels.
… original augmentor order.
| self._setup_device(device) | ||
| self.log_file = log_file | ||
|
|
||
| self.set_params( |
There was a problem hiding this comment.
I don't think a set_params method is necessary, the assignments can happen in __init__
There was a problem hiding this comment.
The set_params was a method from the original version of the augmentor. Personally I'd rather have that change be in a separate PR.
There was a problem hiding this comment.
might as well do it in this PR? should be a simple change
src/quantem/core/utils/augment_dp.py
Outdated
| def _apply_bkg_to_multichannel_label(self, label: ArrayLike, probe: ArrayLike | None = None) -> ArrayLike: | ||
| """Apply background to specified channels of multichannel label""" | ||
| if len(label.shape) != 3: | ||
| # Single channel label - shouldn't normally happen |
There was a problem hiding this comment.
should raise an error then
| self.generate_params() | ||
| self._init_log_file() | ||
|
|
||
| def _setup_device(self, device: str) -> None: |
There was a problem hiding this comment.
not sure if it's worth updating this now, but we do have quantem.config for device management and array_funcs for handling torch or numpy more easily. https://github.com/electronmicroscopy/quantem/blob/dev/src/quantem/core/utils/array_funcs.py
| if len(transformed_label.shape) == 3: | ||
| transformed_label = self._apply_bkg_to_multichannel_label(transformed_label, probe) | ||
|
|
||
| if self.add_ellipticity or self.add_shift or self.add_scale: |
There was a problem hiding this comment.
shouldn't adding ellipticity/shift/scaling be done before adding background? Otherwise there will be a hard edge to the noise after the shift/scale. You will need to also update the add_bkg function to account for the shifted center (see Zixi's updated augmentor file I sent you a while back for an implementation of that).
arthurmccray
left a comment
arthurmccray
left a comment
There was a problem hiding this comment.
Overall looks good. Main things are whether the background_label_application flag should exist and the order of when to apply_bkg. Also please attach the updated tutorial notebook to the PR.
|
If I may, would there be a quick and dirty notebook? Love to try this class |
…ion and implemented that functionality in apply_background_to_label, which is now either list or None to reflect this. If None, will not apply, otherwise will apply background to labels according to boolean list. Updated _apply_bkg to incorporate ZiXi's correction for shifts. Changed defaults of gaussian noise to be 0 for mu, 1e-5 for std, and updated docstring to reflect that this scaling is based on total electron dose. Changed application order to be elastic->background. Added aperature related attributes to docstring, and fixed typo in docstring. Added warning if single channel passed to multichannel functions.
|
Heads-up: you'll need to pull dev into the PR to enable the updated automated checks. |
Added an option to apply the background to the labels as well (this was implemented for purpose of adding to the background label image, so amorphous halos + background). The background_label_application is a boolean list (or None) that indicates which label channels to apply the background to if apply_background_to_label is true.
Added gaussian noise, with parameters relative to electron dose.
Implemented ability to work with multichannel labels.
Added preliminary sections for apertures.