Fix a bug of Huber loss

This commit is contained in:
Shangtong Zhang
2018-06-12 21:52:15 -06:00
parent 19572d0640
commit a7e070077d
+1 -1
View File
@@ -28,7 +28,7 @@ class QuantileRegressionDQNAgent(BaseAgent):
(2 * np.arange(self.config.num_quantiles) + 1) / (2.0 * self.config.num_quantiles))
def huber(self, x):
cond = (x < 1.0).float().detach()
cond = (x.abs() < 1.0).float().detach()
return 0.5 * x.pow(2) * cond + (x.abs() - 0.5) * (1 - cond)
def evaluation_action(self, state):