huber loss

examples/Atari2600/common.py

@@ -90,7 +90,7 @@ class Evaluator(Callback):
         self.output_names = output_names
 
     def _setup_graph(self):
-        NR_PROC = min(multiprocessing.cpu_count() // 2, 8)
+        NR_PROC = min(multiprocessing.cpu_count() // 2, 20)
         self.pred_funcs = [self.trainer.get_predict_func(
             self.input_names, self.output_names)] * NR_PROC
 

tensorpack/RL/simulator.py

@@ -126,7 +126,6 @@ class SimulatorMaster(threading.Thread):
 
         def f():
             msg = self.send_queue.get()
-            # slow
             self.s2c_socket.send_multipart(msg, copy=False)
         self.send_thread = LoopThread(f)
         self.send_thread.daemon = True
@@ -142,11 +141,7 @@ class SimulatorMaster(threading.Thread):
 
     def run(self):
         self.clients = defaultdict(self.ClientState)
-        #cnt = 0
         while True:
-            #cnt += 1
-            #if cnt % 3000 == 0:
-                #print_total_timer()
             msg = loads(self.c2s_socket.recv(copy=False).bytes)
             ident, state, reward, isOver = msg
             client = self.clients[ident]
@@ -179,7 +174,6 @@ class SimulatorMaster(threading.Thread):
     def __del__(self):
         self.context.destroy(linger=0)
 
-
 class SimulatorProcessDF(SimulatorProcessBase):
     """ A simulator which contains a forward model itself, allowing
     it to produce data points directly """
@@ -208,7 +202,6 @@ class SimulatorProcessDF(SimulatorProcessBase):
     def get_data(self):
         pass
 
-
 class SimulatorProcessSharedWeight(SimulatorProcessDF):
     """ A simulator process with an extra thread waiting for event,
     and take shared weight from shm.

tensorpack/callbacks/stat.py

 # -*- coding: utf-8 -*-
-# File: summary.py
+# File: stat.py
 # Author: Yuxin Wu <ppwwyyxxc@gmail.com>
 
 import tensorflow as tf

tensorpack/tfutils/symbolic_functions.py

@@ -78,12 +78,16 @@ def rms(x, name=None):
             return tf.sqrt(tf.reduce_mean(tf.square(x)), name=name)
     return tf.sqrt(tf.reduce_mean(tf.square(x)), name=name)
 
-def clipped_l2_loss(x, name=None):
+def huber_loss(x, delta=1, name=None):
     if name is None:
-        name = 'clipped_l2_loss'
+        name = 'huber_loss'
     sqrcost = tf.square(x)
     abscost = tf.abs(x)
-    return tf.select(abscost < 1, sqrcost, abscost, name=name)
+    return tf.reduce_sum(
+            tf.select(abscost < delta,
+                sqrcost * 0.5,
+                abscost * delta - 0.5 * delta ** 2),
+            name=name)
 
 def get_scalar_var(name, init_value):
     return tf.get_variable(name, shape=[],