Load_learner error in package (not notebook)

I’m using fastai v1 for a segmentation project and trying to integrate a trained inference learner into its own package. I put together a small python file (Inference.py) that defines the custom loss functions I used in my model, and defines a single class for calling load_learner and predicting on numpy input.

When I import Inference in a new jupyter notebook and use the class to call load_learner, it works fine, but I’m finding when I call it from another python package (trying to integrate into a broader project) it fails…

It seems like when I load the pickled model with load_learner, it’s not seeing the custom loss functions I defined (and looking for them in main), even though they are defined in Inference.py

Has anyone else run into anything similar or have thoughts?


 File "/packages/shatterstar/shatterstar/shatterstar.py", line 44, in detection_function_fastai_resnet_unet
    lm = LoadModel(path_mdl, "export.pkl")
  File "/packages/shatterstar/shatterstar/Inference.py", line 62, in __init__
    self.learn = load_learner(path,mdl)
  File "/anaconda3/envs/pFastai/lib/python3.6/site-packages/fastai/basic_train.py", line 614, in load_learner
    state = torch.load(source, map_location='cpu') if defaults.device == torch.device('cpu') else torch.load(source)
  File "/anaconda3/envs/pFastai/lib/python3.6/site-packages/torch/serialization.py", line 426, in load
    return _load(f, map_location, pickle_module, **pickle_load_args)
  File "/anaconda3/envs/pFastai/lib/python3.6/site-packages/torch/serialization.py", line 613, in _load
    result = unpickler.load()
AttributeError: Can't get attribute 'combo_loss' on <module '__main__' from '/anaconda3/envs/pFastai/bin/shatterstar'>

The one you used when you pickled your model was in the notebook (module main). You need to export your model with all custom loss functions/layers coming from a module (like your inference.py). When in the notebook, you can check where the function comes from just by executing a name with the name of that function;

my_function

will either return something like __main__.my_function or inference.my_function

1 Like

Appreciate the swift response – I’ll try that!

To confirm, you’re saying I should go back to the training notebook and import the loss functions from Inference.py vs defining them in main, and then when I export it will know where to find them.

Yes, exactly. Note that you can use learn.load to load a previously trained model if you have a safe, you don’t need to retrain. That Learner object just needs references to a loss function that is in the module inference, not main.

1 Like