bug fix & update speed numbers

examples/Atari2600/DQN.py

@@ -38,7 +38,7 @@ EXPLORATION_EPOCH_ANNEAL = 0.01
 END_EXPLORATION = 0.1
 
 MEMORY_SIZE = 1e6
-INIT_MEMORY_SIZE = 50000
+INIT_MEMORY_SIZE = 5e4
 STEP_PER_EPOCH = 10000
 EVAL_EPISODE = 50
 

examples/Atari2600/README.md

@@ -17,7 +17,7 @@ DQN was trained on 1 GPU and it typically took 2~3 days of training to reach a s
 My Batch-A3C implementation only took <2 hours with 2 GPUs (one for training and one for simulation).
 This is probably the fastest RL trainer you'd find.
 
-The x-axis is the number of iterations not wall time. The iteration speed is 6.7it/s for B-A3C and 7.3it/s for D-DQN.
+The x-axis is the number of iterations not wall time, but iteration speed is about 7.8it/s for both models.
 
 A demo trained with Double-DQN on breakout is available at [youtube](https://youtu.be/o21mddZtE5Y).
 

examples/cifar-convnet.py

@@ -18,7 +18,7 @@ A small convnet model for Cifar10 or Cifar100 dataset.
 Cifar10:
     90% validation accuracy after 40k step.
     91% accuracy after 80k step.
-    18.8 step/s on TitanX
+    19.3 step/s on Tesla M40
 
 Not a good for Cifar100, just for demonstration.
 """

examples/svhn-digit-convnet.py

@@ -14,7 +14,9 @@ from tensorpack.tfutils.summary import *
 
 """
 SVHN convnet.
-About 3.0% validation error after 70 epoch.  2.5% after 130 epoch.
+About 3.0% validation error after 70 epoch. 2.5% after 130 epoch.
+
+Each epoch is set to 4721 iterations. The speed is about 44 it/s on a Tesla M30
 """
 
 class Model(ModelDesc):

tensorpack/RL/expreplay.py

@@ -77,7 +77,7 @@ class ExpReplay(DataFlow, Callback):
 
         with tqdm(total=self.init_memory_size) as pbar:
             while len(self.mem) < self.init_memory_size:
-                #from copy import deepcopy  # for debug
+                #from copy import deepcopy  # quickly fill the memory for debug
                 #self.mem.append(deepcopy(self.mem[0]))
                 self._populate_exp()
                 pbar.update()

tensorpack/train/trainer.py

@@ -182,7 +182,7 @@ class QueueInputTrainer(Trainer):
 
         self.train_op = tf.group(
             self.config.optimizer.apply_gradients(grads, get_global_step_var()),
-            summary_moving_average(), 'train_op')
+            summary_moving_average(), name='train_op')
 
         self.main_loop()
 

tensorpack/utils/gpu.py

@@ -3,7 +3,7 @@
 # File: gpu.py
 # Author: Yuxin Wu <ppwwyyxxc@gmail.com>
 
-
+import os
 from .utils import change_env
 
 __all__ = ['change_gpu', 'get_nr_gpu', 'get_gpus']