fastmtl

Multi-task learning utilities for fastai

Install

pip install fastmtl

Usage

Loss

Apply a loss function per model output and get weighted sum of them. For instance, if the first model output is for classification and the second model output is for regression,

from fastmtl.loss import CombinedLoss
loss_func = CombinedLoss(CrossEntropyLossFlat(), MSELossFlat(), weight=[1.0, 3.0])

Metric

Apply metrics for each model output. For instance, if we have a model making classification and regression, we can evaluate each model output with relevant metrics. Assuming that model outputs a tuple of tensors for classification and regression, respectively:

from fastai.metrics import F1Score, R2Score
from fastmtl.metric import mtl_metrics

clf_f1_macro =  F1Score(average='macro')
clf_f1_macro.name = 'clf_f1(macro)'
clf_f1_micro =  F1Score(average='micro')
clf_f1_micro.name = 'clf_f1(micro)'

reg_r2 = R2Score()
reg_r2.name = 'reg_r2'

# metrics for classification in the first list 
# metrics for regression in the second list 
metrics = mtl_metrics([clf_f1_macro, clf_f1_micro], [reg_r2])

learn = Learner(
    ...
    metrics=metrics,
)

Tutorials

Video distortion detection

TODO

• Support tabular learner
• Support fastai>=2.7