New auto-df model with 45 seconds or temporal context (from 35) (comm…

…aai#115) * New auto-df model! 45 seconds of temporal context * Update dynamic follow for new auto-df model * whoops, forgot to add a_lead * delete readme backup
stereosix · Jun 18, 2020 · 44083a0 · 44083a0
1 parent d6fbfef
commit 44083a0
Show file tree

Hide file tree

Showing 5 changed files with 7 additions and 28 deletions.
diff --git a/README_backup.md b/README_backup.md
diff --git a/selfdrive/controls/lib/dynamic_follow/__init__.py b/selfdrive/controls/lib/dynamic_follow/__init__.py
@@ -29,10 +29,10 @@ def __init__(self, mpc_id):
 
     # Model variables
     mpc_rate = 1 / 20.
-    self.model_scales = {'v_ego': [-0.06112159043550491, 37.96522521972656], 'v_lead': [0.0, 35.27671432495117], 'x_lead': [2.4600000381469727, 139.52000427246094]}
+    self.model_scales = {'v_ego': [-0.06112159043550491, 37.96522521972656], 'a_lead': [-3.109330892562866, 3.3612186908721924], 'v_lead': [0.0, 35.27671432495117], 'x_lead': [2.4600000381469727, 141.44000244140625]}
     self.predict_rate = 1 / 4.
-    self.skip_every = round(0.2 / mpc_rate)
-    self.model_input_len = round(35 / mpc_rate)  # int: model input time
+    self.skip_every = round(0.25 / mpc_rate)
+    self.model_input_len = round(45 / mpc_rate)
 
     # Dynamic follow variables
     self.default_TR = 1.8
@@ -146,6 +146,7 @@ def _store_df_data(self):
     # Store data for auto-df model
     self.auto_df_model_data.append([self._norm(self.car_data.v_ego, 'v_ego'),
                                     self._norm(self.lead_data.v_lead, 'v_lead'),
+                                    self._norm(self.lead_data.a_lead, 'a_lead'),
                                     self._norm(self.lead_data.x_lead, 'x_lead')])
     while len(self.auto_df_model_data) > self.model_input_len:
       del self.auto_df_model_data[0]

diff --git a/selfdrive/controls/lib/dynamic_follow/auto_df.py b/selfdrive/controls/lib/dynamic_follow/auto_df.py
@@ -16,7 +16,5 @@ def predict(x):
   l1 = np.dot(l0, w[1]) + b[1]
   l1 = np.maximum(0, l1)
   l2 = np.dot(l1, w[2]) + b[2]
-  l2 = np.maximum(0, l2)
-  l3 = np.dot(l2, w[3]) + b[3]
-  l3 = softmax(l3)
-  return l3
+  l2 = softmax(l2)
+  return l2
diff --git a/selfdrive/controls/lib/dynamic_follow/auto_df_weights.npz b/selfdrive/controls/lib/dynamic_follow/auto_df_weights.npz
diff --git a/.../controls/lib/dynamic_follow/benchmark.py → ...s/lib/dynamic_follow/benchmark_auto_df.py b/.../controls/lib/dynamic_follow/benchmark.py → ...s/lib/dynamic_follow/benchmark_auto_df.py
@@ -2,7 +2,7 @@
 import time
 import numpy as np
 
-samples = np.random.rand(2000, 800).astype(np.float32)
+samples = np.random.rand(2000, 720).astype(np.float32)
 
 t = time.time()
 for sample in samples: