Bug-fix: "convolutions": len(network["channels"]) - 1 instead of `-…

… len(network["nodes"])`.

breakout/DQN.py

@@ -175,7 +175,7 @@ def __init__(self,
             "discount": other.get("discount", 0.99),
             "gamma": other.get("gamma", 0.95),
 
-            "convolutions": len(network["channels"]) - len(network.get("nodes", [])),
+            "convolutions": len(network["channels"]) - 1,
 
             "optimizer": optimizer["optimizer"](self.parameters(), lr=optimizer["lr"],
                                                 **optimizer.get("hyperparameters", {}))
@@ -250,12 +250,6 @@ def preprocess(self, state):
         -------
         output : torch.Tensor
         """
-        # state = (torch.tensor(state,
-        #                       dtype=torch.float32).view(self.shape["original"]) /
-        #          torch.tensor(255,
-        #                       dtype=torch.float32))[:, :, self.shape["height"], self.shape["width"]]
-        # state = torch.nn.functional.max_pool2d(state, self.shape["max_pooling"])
-
         state = torch.tensor(state, dtype=torch.float32).view(self.shape["original"])
         state = state[:, :, self.shape["height"], self.shape["width"]] / 255.0
 
@@ -396,7 +390,7 @@ def learn(self, network, clamp=None):
             # BACKPROPAGATION
             # --------------------------------------------------------------------------------------
 
-            loss = torch.nn.functional.huber_loss(actual, optimal, reduction="mean")
+            loss = torch.nn.functional.mse_loss(actual, optimal)
 
         self.parameter["optimizer"].zero_grad()
         loss.backward()

enduro/DQN.py

@@ -37,6 +37,8 @@ def __init__(self,
                 Kernel size for each layer.
             channels : list, optional
                 Number of channels for each hidden layer.
+            nodes : list of int, optional
+                Number of nodes in fully-connected layer(s).
         optimizer : dict
             Contains the optimizer for the model and its hyperparameters. The dictionary must
             contain the following keys:
@@ -174,7 +176,7 @@ def __init__(self,
             "discount": other.get("discount", 0.99),
             "gamma": other.get("gamma", 0.95),
 
-            "convolutions": len(network["channels"]) - len(network.get("nodes", [])),
+            "convolutions": len(network["channels"]) - 1,
 
             "optimizer": optimizer["optimizer"](self.parameters(), lr=optimizer["lr"],
                                                 **optimizer.get("hyperparameters", {}))

tetris/DQN.py

@@ -176,7 +176,7 @@ def __init__(self,
             "discount": other.get("discount", 0.99),
             "gamma": other.get("gamma", 0.95),
 
-            "convolutions": len(network["channels"]) - len(network.get("nodes", [])),
+            "convolutions": len(network["channels"]) - 1,
 
             "optimizer": optimizer["optimizer"](self.parameters(), lr=optimizer["lr"],
                                                 **optimizer.get("hyperparameters", {}))