(Easy) Add eps to Adam (facebookresearch#858)

Summary:
Pull Request resolved: facebookresearch#858

The Adam optimizer has a parameter called epsilon which is useful for numeric stability. It's useful to explicitly set epsilon especially for unit-tests. This diff adds epsilon to Adam's config

Differential Revision: D16573250

fbshipit-source-id: adc1eca53d190ac4f851c1bc57b71852f499df61

pytext/optimizer/optimizers.py

@@ -34,13 +34,14 @@ class Adam(torch.optim.Adam, Optimizer):
     class Config(Optimizer.Config):
         lr: float = 0.001
         weight_decay: float = 0.00001
+        eps: float = 1e-8
 
-    def __init__(self, parameters, lr, weight_decay):
-        super().__init__(parameters, lr=lr, weight_decay=weight_decay)
+    def __init__(self, parameters, lr, weight_decay, eps):
+        super().__init__(parameters, lr=lr, weight_decay=weight_decay, eps=eps)
 
     @classmethod
     def from_config(cls, config: Config, model: torch.nn.Module):
-        return cls(model.parameters(), config.lr, config.weight_decay)
+        return cls(model.parameters(), config.lr, config.weight_decay, config.eps)
 
 
 class SGD(torch.optim.SGD, Optimizer):