Custom prediction problem

fastai
1

Hi, I create the learner by the pytorch model and data. How should I achieve the prediction by the learner?
data = Dataset(TRN)
trn, val = train_test_split(data, test_size=0.2)
trn = DataLoader(trn, batch_size=4, shuffle=True)
val = DataLoader(val, batch_size=4, shuffle=False)
data = DataBunch(trn, val)

learn = Learner(data, U_Net(), opt_func=torch.optim.Adam, metrics=[jaccard_dis, jaccard_coef_int])
learn.loss_func = nn.BCELoss()

The trn and val both are pytorch dataloader. When I try to run learn.predict, the error is as follows:
AttributeError: 'list' object has no attribute 'set_item'

2

As far as I know, learn.predict requires the DataBunch to be created with the data_block api. In addition, the argument you pass to it must inherit from ItemBase, such as the Image class. Just to be clear, are you trying to get a single prediction or all train/validation set predictions? If it is the latter, you want to use learn.get_preds.

I’d recommend loading your data with the data_block api, so that you can use the full array of learner functions that depend on it. However, if you just want to get a single prediction with your current setup you can just grab the input tensor and directly pass it to learn.model.

3

Thanks for your reply. It is difficult to create the DataBunch in my datasets. I will try the method you recommended.