v1.0.5 -> optional input clamping

README.md

@@ -53,7 +53,7 @@ See the `examples` and `benchmarks` directories to see how to run training - the
 There are four distinct steps in `fastprop` that define its framework:
  1. Featurization - transform the input molecules (as SMILES strings) into an array of molecular descriptors which are saved
  2. Preprocessing - clean the descriptors by removing or imputing missing values then rescaling the remainder
- 3. Training - send the processed input to the neural network, which is a simple FNN (sequential fully-connected layers with an activation function between)
+ 3. Training - send the processed input to the neural network, which is a simple FNN (sequential fully-connected layers with an activation function between), optionally limiting the inputs to +/-3 standard deviations to aid in extrapolation
  4. Prediction - save the trained model for future use
 
 ## Configurable Parameters
@@ -76,6 +76,7 @@ There are four distinct steps in `fastprop` that define its framework:
     _and_
     - Number of FNN layers (default 2; repeated fully connected layers of hidden size)
     - Hidden Size: number of neurons per FNN layer (default 1800)
+    - Clamp Input: Enable/Disable input clamp to +/-3 to aid in extrapolation (default False).
 
     _or_
     - Hyperparameter optimization: runs hyperparameter optimization identify the optimal number of layers and hidden size

fastprop/model.py

@@ -18,6 +18,18 @@
 logger = init_logger(__name__)
 
 
+class ClampN(torch.nn.Module):
+    def __init__(self, n: float) -> None:
+        super().__init__()
+        self.n = n
+
+    def forward(self, batch: torch.Tensor):
+        return torch.clamp(batch, min=-self.n, max=self.n)
+
+    def extra_repr(self) -> str:
+        return f"n={self.n}"
+
+
 class fastprop(pl.LightningModule):
     def __init__(
         self,
@@ -27,6 +39,7 @@ def __init__(
         num_tasks: int = 1,
         learning_rate: float = 0.001,
         fnn_layers: int = 2,
+        clamp_input: bool = False,
         problem_type: Literal["regression", "binary", "multiclass", "multilabel"] = "regression",
         target_names: List[str] = [],
         feature_means: Optional[torch.Tensor] = None,
@@ -43,6 +56,7 @@ def __init__(
             num_tasks (int, optional): Number of distinct tasks. Defaults to 1.
             learning_rate (float, optional): Learning rate for SGD. Defaults to 0.001.
             fnn_layers (int, optional): Number of hidden layers. Defaults to 2.
+            clamp_input (bool, optional): Clamp inputs to +/-3 to aid extrapolation. Defaults to False.
             problem_type (Literal["regression", "binary", "multiclass", "multilabel"], optional): Type of training task. Defaults to "regression".
             target_names (list[str], optional): Names for targets in dataset, blank for simple integer names. Defaults to [].
             feature_means (Optional[torch.Tensor], optional): Means for scaling features in regression. Defaults to None.
@@ -63,6 +77,7 @@ def __init__(
 
         # fully-connected nn
         layers = OrderedDict()
+        layers["clamp"] = ClampN(3)
         for i in range(fnn_layers):
             layers[f"lin{i+1}"] = torch.nn.Linear(input_size if i == 0 else hidden_size, hidden_size)
             if fnn_layers == 1 or i < (fnn_layers - 1):  # no output activation, unless single layer

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "fastprop"
-version = "1.0.4"
+version = "1.0.5"
 authors = [
     { name = "Jackson Burns" },
 ]