log reward clipping is now -100 (much smaller)

src/gfn/env.py

@@ -23,7 +23,7 @@ def __init__(
         sf: Optional[TT["state_shape", torch.float]] = None,
         device_str: Optional[str] = None,
         preprocessor: Optional[Preprocessor] = None,
-        log_reward_clip: Optional[float] = -20.0,
+        log_reward_clip: Optional[float] = -100.,
     ):
         """Initializes an environment.
 
@@ -221,7 +221,7 @@ def __init__(
         sf: Optional[TT["state_shape", torch.float]] = None,
         device_str: Optional[str] = None,
         preprocessor: Optional[Preprocessor] = None,
-        log_reward_clip: Optional[float] = -20.0,
+        log_reward_clip: Optional[float] = -100.0,
     ):
         """Initializes a discrete environment.
 

src/gfn/gym/box.py

@@ -20,7 +20,7 @@ def __init__(
         R2: float = 2.0,
         epsilon: float = 1e-4,
         device_str: Literal["cpu", "cuda"] = "cpu",
-        log_reward_clip: float = -20.,
+        log_reward_clip: float = -100.,
     ):
         assert 0 < delta <= 1, "delta must be in (0, 1]"
         self.delta = delta

src/gfn/gym/discrete_ebm.py

@@ -48,7 +48,7 @@ def __init__(
         alpha: float = 1.0,
         device_str: Literal["cpu", "cuda"] = "cpu",
         preprocessor_name: Literal["Identity", "Enum"] = "Identity",
-        log_reward_clip: float = -20.,
+        log_reward_clip: float = -100.,
     ):
         """Discrete EBM environment.
 

src/gfn/gym/hypergrid.py

@@ -25,7 +25,7 @@ def __init__(
         reward_cos: bool = False,
         device_str: Literal["cpu", "cuda"] = "cpu",
         preprocessor_name: Literal["KHot", "OneHot", "Identity", "Enum"] = "KHot",
-        log_reward_clip: float = -20.,
+        log_reward_clip: float = -100.,
     ):
         """HyperGrid environment from the GFlowNets paper.
         The states are represented as 1-d tensors of length `ndim` with values in