Added default num_timesteps to CropHarvestEval to fix validation while running train.py.
The num_timesteps field is not initialized in the CropHarvestMultiClassValidation subclass that is used for validation, but the finetuning_results method calls truncate_timesteps which accesses the field.
5 made the TreeSat eval task multilabel, causing the sigmoid in FinetuningHead not to be applied anymore. However, TreeSatEval still used pytorch's BCELoss which expects inputs to be in [0, 1]. This PR changes TreeSatEval to use BCEWithLogitsLoss instead.
Further regarding TreeSat having been made multilabel: sklearn's LogisticRegression doesn’t support multilabel tasks, its .fit(X, y) takes y with shape (n_samples,). See the discussion here. To keep using logistic regression, I wrapped the estimator in a sklearn.multioutput.MultiOutputClassifier which will fit one classifier per target.
Added default
num_timestepsto CropHarvestEval to fix validation while runningtrain.py. The num_timesteps field is not initialized in theCropHarvestMultiClassValidationsubclass that is used for validation, but thefinetuning_resultsmethod callstruncate_timestepswhich accesses the field.5 made the TreeSat eval task multilabel, causing the sigmoid in
FinetuningHeadnot to be applied anymore. However, TreeSatEval still used pytorch'sBCELosswhich expects inputs to be in [0, 1]. This PR changes TreeSatEval to useBCEWithLogitsLossinstead.Further regarding TreeSat having been made multilabel: sklearn's
LogisticRegressiondoesn’t support multilabel tasks, its.fit(X, y)takes y with shape(n_samples,). See the discussion here. To keep using logistic regression, I wrapped the estimator in asklearn.multioutput.MultiOutputClassifierwhich will fit one classifier per target.