backup

docs/design/network_inputs.md

@@ -0,0 +1,19 @@
+# Neural network architecture and inputs
+
+## Inputs
+From the observations, we use:
+* joint_position
+* joint_velocity
+* joint_actuator_force
+* actuator_force
+* disk_position
+* disk_rotation
+* disk_linear_velocity
+* disk_angular_velocity
+* unit_xy_direction_to_target
+* xy_distance_to_target
+
+as inputs for the sensor and the feature extractor.
+
+## Sensor architecture
+todo

src/brittle_star_project/MLPs/mlps.py

@@ -2,7 +2,6 @@ from dataclasses import dataclass, fields, field
 
 import flax
 import flax.linen as nn
-import jax.numpy as jnp
 import jax.tree_util
 from typing import Sequence, Callable
 from flax.linen.initializers import constant, orthogonal
@@ -13,18 +12,26 @@ class GenericDenseLayersWithActivation(nn.Module):
     layer_sizes: Sequence[int] = field(default_factory=lambda: [64, 64])
     activation: Callable = nn.tanh
 
+    def setup(self):
+        self.dense_layers = [nn.Dense(size) for size in self.layer_sizes]
+
     @nn.compact
     def __call__(self, x):
-        for size in self.layer_sizes:
+        for layer in self.dense_layers:
-            x = nn.Dense(size, kernel_init=orthogonal(jnp.sqrt(2)))(x)
+            x = layer(x)
             x = self.activation(x)
         return x
 
 
 class OneDenseLayerMLP(nn.Module):
+    feature_dim: int = field(default=1)
+
+    def setup(self):
+        self.dense_layer = nn.Dense(self.feature_dim)
+
     @nn.compact
     def __call__(self, x):
-        return nn.Dense(1, kernel_init=orthogonal(1), bias_init=constant(0.0))(x)
+        return self.dense_layer(x)
 
 
 class Actor(nn.Module):

src/brittle_star_project/trainers/PPOTrainer.py

@@ -1,4 +1,5 @@
 import datetime
+import logging
 import random
 import time
 from dataclasses import asdict, dataclass
@@ -215,6 +216,9 @@ class PPOTrainer:
         self.run_name = run_name
         self.logger = get_logger()
 
+        # TODO: remove
+        self.logger.set_level(logging.DEBUG)
+
         self.key = jax.random.PRNGKey(args.seed)
 
         self.sensor, self.feature_extractor, self.actor, self.critic = self._init_agent()
@@ -262,8 +266,10 @@ class PPOTrainer:
     def _init_agent(self):
         self.logger.info("[AGENT]: Initializing agent...")
 
-        sensor = GenericDenseLayersWithActivation()
+        sizes = [195 // 2, 195 // 4, 195 // 4]
-        feature_extractor = GenericDenseLayersWithActivation()
+
+        sensor = GenericDenseLayersWithActivation(layer_sizes=sizes)
+        feature_extractor = GenericDenseLayersWithActivation(layer_sizes=sizes)
         actor = Actor(
             action_dim=self.env.single_action_space.shape[0]
         )  # continuous actions for MJX