Open ainazHjm opened 5 years ago
The loss function that I'm optimizing on is BCEwithLogitsLoss (The loss is computed for each sample. w
is the weight that is learned. y
is the target which is either 0
or 1
and x
is the prediction.)
-w [y * log(sigmoid(x)) + (1 - y) * log(1 - sigmoid(x))]
As the loss function contains the sigmoid. The model will not have any sigmoids so the output of the model is not a probability. When validating, we need to add the sigmoid to get the probabilities.
As the ratio of the landslide areas is 0.003. The first model is going to predict -9.2, 99.7% of the time and 9.2 0.3% of the time. -9.2 and 9.2 correspond to 0 and 1 because the sigmoid of them is close to 0 and 1. So, from the loss formula the expected loss (likelihood that the model should converge to) should be around 0.03 ~ 0.04.
Also, the purpose of comparing the results from these baselines is to see how much improvement each one of them makes.
The results should be reported on both the validation and the train sets.
Not weighted average loss with no oversampling on training data:
Weighted average loss with no oversampling on training data with weight=50: