Area: This model is a Bernouli distribution that will calculate the ratio of landslides (pos. samples) in Veneto and use that as the prob. for the distribution. For Veneto this ratio is almost 0.003 so 0.003 of the time it will return logits(1) and 0.997 of the time logits(0).
NoNghbr: This model is a simple CNN with k=(1,1) that will only look at the pixel values.
Nghbr: This model is a CNN with k=(3,3) that considered the 3x3 neighborhood for prediction as well as the pixel value itself.
I tried both the weighted version of BCEwithLogitsLoss and not weighted version:
Not weighted:
Area: 0.055448
No neighbors: 0.049747
3x3 neighbors: 0.022946
Weighted:
Area: 1.431502
No neighbors: 0.32223
3x3 neighborhood: 0.343558
**
the reason that the loss is bigger than 1 for Area in the weighted version is that weighted BCEwithLogitsLoss is not a normalized version of the log likelihood so only the positive features are multiplied by some weight and not the negative ones. Also, I didn't try the other two model enough, the loss would've been smaller if trained enough.
I have three baseline models for Veneto:
I tried both the weighted version of BCEwithLogitsLoss and not weighted version: Not weighted:
Weighted:
** the reason that the loss is bigger than 1 for Area in the weighted version is that weighted BCEwithLogitsLoss is not a normalized version of the log likelihood so only the positive features are multiplied by some weight and not the negative ones. Also, I didn't try the other two model enough, the loss would've been smaller if trained enough.