2dmodels

[BrendenBarbour]

Adding config file control over the model and backbone from the Segmentation Models Repository. Change the name to Unet, FPN, or Linknet in the config file. Not all backbones work for all models.

[rak5216]

@BrendenBarbour within each if statement to set model, can u add if's or assert's to verify that a suitable backbone has also been selected (based on your findings in the spreadsheet)? we made this note at the top^

[CarolinaFurtado]

Added my changes:

• input shape magic trick (because the yaml does not take tuples)
• added asserts to make sure the pretrained model and the model are compatible
• info to the config files, we will need to add the remaining variables to the other config files or else it will fail when you have a pretrained model!!! because it is asserting model parameters that are not defined (input_shape and activation)
• added what was missing from master (no conflicts there)

[CarolinaFurtado]

@BrendenBarbour this looks good!

Did you check that everything is working? As the one pushing the changes, you should be the one to do all the verifications. The reviewers will comment on the code itself and see if structurally it makes sense, but it is important that you do an overall check on what you think might be affected.

A few things that are important to check:

• all the workflow steps run
• that all model names work (you can do a print screen of the output so we easily see that the objective of the PR is satisfied). I checked this and it seems to be working but you should to the final check.
• a few asserts were added in the pretrained models. a simple check would be to have different models for the trained and pretrained models and see if the asserts make it stop for example. Checking if the asserts you added actually work is also a good thing to look at.

Let me know if you need help!

[CarolinaFurtado]

@BrendenBarbour @rak5216 this looks good!

I'll aprove after you create a template dataset and go through the workflow steps and make sure all are running as we discussed yesterday, @BrendenBarbour

[CarolinaFurtado]

added the possibilities of optimizers and loss functions:

any specific loss function you want to add, @rak5216 ?

[CarolinaFurtado]

Overall checks - all working as expected

Checking that all the workflow steps work:

infer - ok

dataset generation - ok

train- ok

without pretrained model - ok

with pretrained model compatible - ok

train threshold - ok

test - ok

infer - ok

Pretrained models asserts

with pretrained model NOT COMPATIBLE BACKBONE - ok, assert error

with pretrained model NOT COMPATIBLE MODEL - ok, assert error

with pretrained model OTHER optimizer- ok, NO ERROR, IT'S UP TO THE USER TO KNOW

with pretrained model OTHER loss- ok, NO ERROR, IT'S UP TO THE USER TO KNOW

2 model configurations

Unet - ok , verified above

FPN model - ok

and runs till the end (with batch size of 8)

Linknet model - ok

and runs till the end