hidden sweep

cfg/hidden_size.cfg

@@ -0,0 +1,131 @@
+[DEFAULT]
+
+alg = dagger
+
+# learning parameters
+batch_size = 20
+buffer_size = 10000
+updates_per_step = 200
+seed = 11
+actor_lr = 5e-5
+
+n_train_episodes = 400
+beta_coeff = 0.993
+test_interval = 40
+n_test_episodes = 20
+
+# architecture parameters
+k = 3
+hidden_size = 32
+gamma = 0.99
+tau = 0.5
+
+# env parameters
+env = FlockingRelative-v0
+v_max = 3.0
+comm_radius = 1.0
+n_agents = 100
+n_actions = 2
+n_states = 6
+debug = False
+dt = 0.01
+
+
+header = n_layers, hidden_size, reward
+
+[1, 4]
+n_layers = 1
+hidden_size = 4
+
+[1, 8]
+n_layers = 1
+hidden_size = 8
+
+[1, 16]
+n_layers = 1
+hidden_size = 16
+
+[1, 32]
+n_layers = 1
+hidden_size = 32
+
+[1, 64]
+n_layers = 1
+hidden_size = 64
+
+[1, 128]
+n_layers = 1
+hidden_size = 128
+
+[2, 4]
+n_layers = 2
+hidden_size = 4
+
+[2, 8]
+n_layers = 2
+hidden_size = 8
+
+[2, 16]
+n_layers = 2
+hidden_size = 16
+
+[2, 32]
+n_layers = 2
+hidden_size = 32
+
+[2, 64]
+n_layers = 2
+hidden_size = 64
+
+[2, 128]
+n_layers = 2
+hidden_size = 128
+
+[3, 4]
+n_layers = 3
+hidden_size = 4
+
+[3, 8]
+n_layers = 3
+hidden_size = 8
+
+[3, 16]
+n_layers = 3
+hidden_size = 16
+
+[3, 32]
+n_layers = 3
+hidden_size = 32
+
+[3, 64]
+n_layers = 3
+hidden_size = 64
+
+[3, 128]
+n_layers = 3
+hidden_size = 128
+
+[4, 4]
+n_layers = 4
+hidden_size = 4
+
+[4, 8]
+n_layers = 4
+hidden_size = 8
+
+[4, 16]
+n_layers = 4
+hidden_size = 16
+
+[4, 32]
+n_layers = 4
+hidden_size = 32
+
+[4, 64]
+n_layers = 4
+hidden_size = 64
+
+[4, 128]
+n_layers = 4
+hidden_size = 128
+

cfg/make_config.py

@@ -1,14 +1,16 @@
 import itertools
 
 baseline = False
-baseline = True
-param = 'vel'
-param = 'dt'
+# baseline = True
+# param = 'vel'
+# param = 'dt'
 # param = 'n'
 # param = 'rad'
 
-params = {}
+param = 'hidden_size'
 
+params = {}
+#
 # params['seed'] = range(10)
 
 if baseline:
@@ -19,8 +21,9 @@
 else:
     default_fname = 'default.cfg'
     out_fname = param + '.cfg'
+    params['n_layers'] = [1, 2, 3, 4]
 
-    params['k'] = [1, 2, 3, 4]
+    # params['k'] = [1, 2, 3, 4]
 
 if param == 'vel':
     params['v_max'] = [0.5, 1.5, 2.5, 3.5, 4.5]
@@ -30,6 +33,8 @@
     params['n_agents'] = [25, 50, 75, 100, 125, 150, 175, 200]
 elif param == 'dt':
     params['dt'] = [0.1, 0.075, 0.05, 0.025, 0.01, 0.0075]
+elif param == 'hidden_size':
+    params['hidden_size'] = [4, 8, 16, 32, 64, 128]
 
 
 out_file = open(out_fname, "w")

learner/gnn_dagger.py

@@ -28,6 +28,7 @@ def __init__(self, device, args, k=None):  # , n_s, n_a, k, device, hidden_size=
         n_a = args.getint('n_actions')
         k = k or args.getint('k')
         hidden_size = args.getint('hidden_size')
+        n_layers = args.getint('n_layers') or 2
         gamma = args.getfloat('gamma')
         tau = args.getfloat('tau')
 
@@ -38,7 +39,7 @@ def __init__(self, device, args, k=None):  # , n_s, n_a, k, device, hidden_size=
         # Device
         self.device = device
 
-        hidden_layers = [hidden_size, hidden_size]
+        hidden_layers = [hidden_size] * n_layers
         ind_agg = 0  # int(len(hidden_layers) / 2)  # aggregate halfway
 
         # Define Networks