Can't able to train the num_classes=1

[kethan52]

model =SemanticSegmentationTask(
    segmentation_model="fcn",
    encoder_name="resnet18",
    encoder_weights="imagenet",
    in_channels=9,
    num_classes=1,
    num_filters=128,
    loss="jaccard",
    ignore_zeros=True,
    learning_rate=0.1,
    learning_rate_schedule_patience=0.05,
    
)

trainer = pl.Trainer(

                  gpus=0,
                  callbacks=[checkpoint_callback, early_stopping_callback],
                  logger=[csv_logger],
                  min_epochs=10,
                  max_epochs=100,
                  precision=32,
                  log_every_n_steps=0.01,
                  max_steps=8,
                  
               )

trainer.fit(model,dl,dl)

Hi, as I want to do the binary sematic segmentation I need num_clasess=1 can I have any scope of doing with single class

[isaaccorley]

This seems to be a duplicate of your other issues #513 and #514. Can you add some context to why this is a separate issue?

[calebrob6]

num_classes=1 will need to be trained with a different loss than PyTorch's cross entropy, which isn't currently implemented in SemanticSegmentationTask (you would need to use this https://pytorch.org/docs/stable/generated/torch.nn.BCELoss.html#torch.nn.BCELoss). Maybe we should add an assert that num_classes >= 2?

[adamjstewart]

Could we add a conditional that automatically switches to BCE for num_classes=1?

[calebrob6]

Yeah probably, I just don't have time to play around with it :)

…

On Fri, May 6, 2022 at 1:10 PM Adam J. Stewart ***@***.***> wrote: Could we add a conditional that automatically switches to BCE for num_classes=1? — Reply to this email directly, view it on GitHub <#521 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAIJUTXTHM4OVMN4WU6PBYDVIV4E5ANCNFSM5U3GS2RQ> . You are receiving this because you commented.Message ID: ***@***.***>

[adamjstewart]

I think this is also related to #245. So it's not just loss, it's also IoU.