arcgis.learn unfreeze model - Which layers/stages?

340
2
10-08-2021 03:48 AM
Labels (1)
shlirkd
New Contributor II

Hello all,

I have a couple of questions regarding Esri Deep-learning apps.

1) Freeze - unfreeze backbone:
I would be interested to know which layers of the backbone are frozen by default. Additionally it would be interesting to know which layers are unfrozen when calling "model.unfreeze()". I could not find any details in the documentation.
As a concrete example:
MaskRCNN with ResNet 18 as backbone
"model = MaskRCNN(data, backbone="resnet18", early_stopping=True)"

2) Transfer learning:

I assume that transfer learning is applied by default. Is it possible to use no pre-trained weights and to train the net completely new?

As a concrete example:
MaskRCNN with ResNet 18 as backbone
"model = MaskRCNN(data, backbone="resnet18", early_stopping=True)"

3) Weight initialization schemes:

There are no Informations regarding this parameter "arcgis.env.type_init_tail_parameters". I would be particularly interested in the following: let's assume i use a tiff with 20 channels. I would now like to use a single Band for training. Which weights are used now, R,G or B? Presumably this parameter only applies if there are more than 3 channels in a training?

 

Thanks!1

 

0 Kudos
2 Replies
SandeepKumar1
Esri Contributor

Hi

1) By Default, The whole trained backbone is frozen, the only unfrozen part is the newly added model head (depends what kind of model it is). In case of multispectral models, additionally the tail or the first layer from bottom of the model (Input side) is also unfrozen. On calling {model}.unfreeze() the whole model is trainable.

2) Currently there is no way to not use pretrained backbones. From our experiments we found these backbones always outperforming the randomly initialized ones on training, hence it was set as the default behavior.

3) This section is not documented and is for internal API configuration only. If you would like to just use one band for training you need to export the training data which has only one band.

Thanks,
Sandeep

0 Kudos
shlirkd
New Contributor II

Thanks Sandeep!

 

0 Kudos