Changes to hypernet

-

-

-

experimenting

-

ml-agents/mlagents/trainers/torch/layers.py

 from mlagents.torch_utils import torch
 import abc
+import math
 from typing import Tuple
 from enum import Enum
 
             layer_in_size = layer_size
         flat_output = linear_layer(
             layer_size,
-            input_size * output_size + output_size,
+            input_size * output_size,
-        self.hypernet = torch.nn.Sequential(*layers, flat_output)
+
+        # Re-initializing the weights of the last layer of the hypernetwork
+        bound = math.sqrt(1 / (layer_size * self.input_size))
+        flat_output.weight.data.uniform_(-bound, bound)
+
+        self.hypernet = torch.nn.Sequential(*layers, flat_output, LayerNorm())
+
+        # The hypernetwork will not generate the bias of the main network layer
+        self.bias = torch.nn.Parameter(torch.zeros(output_size))
-        flat_output_weights = self.hypernet(hyper_input)
-        batch_size = input_activation.size(0)
+        output_weights = self.hypernet(hyper_input)
-        output_weights, output_bias = torch.split(
-            flat_output_weights, self.input_size * self.output_size, dim=-1
-        )
-
-        output_weights = output_weights.view(
-            batch_size, self.input_size, self.output_size
-        )
-        output_bias = output_bias.view(batch_size, self.output_size)
-        output = (
+        output_weights = output_weights.view(-1, self.input_size, self.output_size)
+        return (
-            + output_bias
+            + self.bias
-        return output

ml-agents/mlagents/trainers/torch/networks.py

                 self.obs_types[idx] == ObservationType.GOAL
                 and self.conditioning_type != ConditioningType.DEFAULT
             ):
+                if goal_signal is not None:
+                    raise "Error"
                 goal_signal = processed_obs
 
         if len(encodes) == 0: