freeze_backbone and freeze_decoder in Trainers #1290
Conversation
isaaccorley
changed the title
github-actions
bot
added
testing
isaaccorley
requested review from
adamjstewart,
calebrob6 and
nilsleh
and removed request for
adamjstewart and
calebrob6
|
Can you do other trainers too? For this paper, all of our tasks will be semantic segmentation, but might as well do all trainers if it's easy. |
isaaccorley
force-pushed
the
trainers/linear-probe
branch
from
be4cd0d to
4fa65fb
Compare
isaaccorley
changed the title
| @@ -110,7 +124,9 @@ class and used with 'ce' loss | |||
| *encoder_weights* to *weights*. | |||
|
|
|||
| .. versionadded: 0.5 | |||
| The *class_weights* parameter. | |||
| The *class_weights*, *freeze_backbone*, | |||
| and *freeze_decoder* parameters. | |||
|
Can you reference some papers that freeze the entire UNet and add fine-tune an MLP on top? I have never seen this before. Also it would make more sense to fine tune a conv layer not a MLP. Could you also suggest better parameter names? |
|
I'm only talking about classification/regression. No idea what the usual practice is for linear probing of semantic segmentation. Will think about parameter names, don't have a better name off the top of my head. I still kind of like linear_probing because it makes it clear why you would want to use the parameter. |
|
Disagree about the naming. Freeze_backbone is the action we are performing. Linear probing is just the downstream task a user may choose to perform with the classifier with a frozen backbone. There may be other potential reasons for freezing the backbone other than linear probing. |
This PR adds a
freeze_backbonearg to all trainers and afreeze_decoderarg to theSemanticSegmentationTaskandPixelwiseRegressionTask. This is particularly useful if wanting to fine-tune (aka linear probe) only the classifier head of a pretrained model or only the decoder or segmentation head in an encoder/decoder architecture.