feature: add new net model with one hidden layer

A new sequential neural network model is added
with a single hidden layer. Linear activations are used.

agent/backend/models.py

@@ -11,7 +11,22 @@ class Perceptron(nn.Module):
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         out = self.layer(x)
         return out
-    
+
+
+class NetSingleHiddenLayer(nn.Module):
+    def __init__(self, in_features: int, out_features: int, hidden_size):
+        super().__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.hidden_size = hidden_size
+        self.layer1 = nn.Sequential(nn.Linear(in_features=self.in_features,out_features=self.hidden_size))
+        self.layer2 = nn.Sequential(nn.Linear(in_features=self.hidden_size,out_features=self.out_features))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = self.layer1(x)
+        out = self.layer2(out)
+        return out
+
 
 class PolicyGradientNetwork(nn.Module):
     def __init__(self, num_inputs, num_actions, hidden_size):

agent/backend/utils.py

@@ -8,7 +8,7 @@ import itertools
 import solara
 
 from .data import ExplorationDataset
-from .models import Perceptron
+from .models import Perceptron, NetSingleHiddenLayer
 from .loss import loss_mape
 
 def predict_dict(model, ds, inputs: Dict[str, Union[List[int], List[float]]]):
@@ -67,6 +67,10 @@ def train(ds: ExplorationDataset, model_name, trn_ratio,
     df = ds.df
     if model_name == "Perceptron":
         model = Perceptron(in_features=len(input_cols), out_features=len(output_cols))
+    elif model_name == "NetSingleHiddenLayer":
+        # TODO: make hidden_size adjustable in ui
+        model = NetSingleHiddenLayer(in_features=len(input_cols), out_features=len(output_cols), hidden_size=10)
+
     if loss_name == "mape":
         loss_fn = loss_mape
 

agent/dashboard/training.py

@@ -83,13 +83,14 @@ def ExecutePanel(df):
         max_epoch = local_state.value['max_epoch'].value
         loss_name = local_state.value['loss_name'].value
         seed = local_state.value['seed'].value
+        model_name = local_state.value['model_name'].value
 
         epoch_list = []
         trn_loss_list = []
         val_loss_list = []
         ds = ExplorationDataset(dff, input_cols, output_cols)
         local_state.value['ds'].set(ds)
-        for epoch, trn_loss, val_loss, model in train(ds, "Perceptron", trn_ratio,
+        for epoch, trn_loss, val_loss, model in train(ds, model_name, trn_ratio,
               batch_size_trn, batch_size_val, optimizer_name, learning_rate,
               max_epoch, loss_name, seed):
             epoch_list.append(epoch)
@@ -125,7 +126,7 @@ def ParameterSelection(df):
                             value=local_state.value['optimizer_name'].value,
                             on_value=local_state.value['optimizer_name'].set)
                 
-                solara.Select(label="Model", values=["Perceptron"], 
+                solara.Select(label="Model", values=["Perceptron","NetSingleHiddenLayer"],
                             value=local_state.value['model_name'].value,
                             on_value=local_state.value['model_name'].set)