Enabling dropout during trainer.predict

[standard-aaron]

I want to enable dropout during .predict and tried implementing the following:

model.eval() 
for m in model.modules():
    if m.__class__.__name__.startswith('Dropout'):
        m.train()
                
...

trainer.predict(
    model, dataloaders=data_loader, return_predictions=True
)

It seems like .predict is overriding this because I get identical predictions with different seeds.

Can someone explain how to accomplish this, or point me to the relevant docs? (Couldn't find them & tried looking for while)

Thank you!

[rohitgr7]

hey @35ajstern !

you can enable this inside predict_step itself. Check this out: https://github.com/PyTorchLightning/pytorch-lightning/blob/f35e2210e240b443fd4dafed8fe2e30ee7d579ea/docs/source/common/production_inference.rst#prediction-api

this is part of a PR, will be available in the docs once merged.

[standard-aaron]

Thank you @rohitgr7!