From 3c9b388b04e0c3f6317c60d4fed81e671862fc95 Mon Sep 17 00:00:00 2001
From: "Daniel.Frisinghelli" <daniel.frisinghelli@eurac.edu>
Date: Thu, 7 Oct 2021 15:40:58 +0200
Subject: [PATCH] Implemented Bernoulli-Weibull loss.
---
climax/core/dataset.py | 6 ++++--
climax/core/loss.py | 25 ++++++++++++-------------
climax/core/predict.py | 2 +-
3 files changed, 17 insertions(+), 16 deletions(-)
diff --git a/climax/core/dataset.py b/climax/core/dataset.py
index 17f8366..302960c 100644
--- a/climax/core/dataset.py
+++ b/climax/core/dataset.py
@@ -19,7 +19,8 @@ import dask.array as da
from climax.core.constants import (ERA5_VARIABLES, ERA5_PRESSURE_LEVELS,
ERA5_P_VARIABLE_NAME, ERA5_S_VARIABLE_NAME,
PROJECTION)
-from climax.core.loss import BernoulliGammaLoss, BernoulliGenParetoLoss
+from climax.core.loss import (BernoulliGammaLoss, BernoulliGenParetoLoss,
+ BernoulliWeibullLoss)
from pysegcnn.core.utils import search_files, img2np
from pysegcnn.core.trainer import LogConfig
@@ -126,7 +127,8 @@ class EoDataset(torch.utils.data.Dataset):
# check which loss function is used
if predictand == 'pr':
if (isinstance(loss, BernoulliGammaLoss) or
- isinstance(loss, BernoulliGenParetoLoss)):
+ isinstance(loss, BernoulliGenParetoLoss) or
+ isinstance(loss, BernoulliWeibullLoss)):
# adjust state file for precipitation
state_file = '_'.join([state_file, '{}mm'.format(
str(loss.min_amount).replace('.', '')),
diff --git a/climax/core/loss.py b/climax/core/loss.py
index 7fdb74b..abb9ca3 100644
--- a/climax/core/loss.py
+++ b/climax/core/loss.py
@@ -178,15 +178,14 @@ class BernoulliWeibullLoss(NaNLoss):
# clip probabilities to (0, 1)
p_pred = torch.sigmoid(y_pred[:, 0, ...].squeeze()[mask])
- # clip shape and scale to (0, 50)
- # NOTE: in general shape, scale in (0, +infinity), but clipping to
- # finite numbers is required for numerical stability
- # gshape = torch.clamp(
- # torch.exp(y_pred[:, 1, ...].squeeze()[mask]), max=1)
- # gscale = torch.clamp(
- # torch.exp(y_pred[:, 2, ...].squeeze()[mask]), max=20)
- gshape = torch.sigmoid(y_pred[:, 1, ...].squeeze()[mask])
- gscale = torch.exp(y_pred[:, 2, ...].squeeze()[mask])
+ # clip scale to (0, +infinity)
+ scale = torch.exp(y_pred[:, 2, ...].squeeze()[mask])
+
+ # clip shape to (0, 1)
+ # NOTE: in general shape in (0, +infinity), but for precipitation the
+ # shape parameter of the Weibull distribution < 1
+ shape = torch.clamp(
+ torch.exp(y_pred[:, 1, ...].squeeze()[mask]), max=1)
# negative log-likelihood function of Bernoulli-Weibull distribution
@@ -198,10 +197,10 @@ class BernoulliWeibullLoss(NaNLoss):
# Weibull contribution
loss -= p_true * (torch.log(p_pred + self.epsilon) +
- torch.log(gshape + self.epsilon) -
- gshape * torch.log(gscale + self.epsilon) +
- (gshape - 1) * torch.log(y_true + self.epsilon) -
- torch.pow(y_true / (gscale + self.epsilon), gshape)
+ torch.log(shape + self.epsilon) -
+ shape * torch.log(scale + self.epsilon) +
+ (shape - 1) * torch.log(y_true + self.epsilon) -
+ torch.pow(y_true / (scale + self.epsilon), shape)
)
# clip loss to finite values to ensure numerical stability
diff --git a/climax/core/predict.py b/climax/core/predict.py
index b09bf22..4ece682 100644
--- a/climax/core/predict.py
+++ b/climax/core/predict.py
@@ -92,7 +92,7 @@ def predict_ERA5(net, ERA5_ds, predictand, loss, batch_size=16, **kwargs):
# distribution
# pr = p * scale * tau(1 + 1 / shape)
pr = (prob * np.exp(target[:, 2, ...].squeeze()) *
- gamma(1 + 1 / torch.sigmoid(target[:, 1, ...])))
+ gamma(1 + 1 / np.exp(target[:, 1, ...])))
ds = {'prob': prob, 'precipitation': pr}
--
GitLab