A3C notes about cpu

examples/OpenAIGym/README.md

@@ -6,7 +6,10 @@ Implemented A3C in [Asynchronous Methods for Deep Reinforcement Learning](http:/
 
 `./train-atari.py --env Breakout-v0 --gpu 0`
 
-The pre-trained models are all trained with 4 GPUs.
+It should run at a speed of 6~10 iteration/s on 1 GPU.
+Training with a significant slower speed (e.g. on CPU) will give bad performance,
+probably because of async issues.
+The pre-trained models are all trained with 4 GPUs for about 2 days.
 
 ### To run a pretrained Atari model for 100 episodes:
 

examples/OpenAIGym/train-atari.py

@@ -247,6 +247,7 @@ if __name__ == '__main__':
             logger.info("[BA3C] Train on gpu {} and infer on gpu {}".format(
                 ','.join(map(str, train_tower)), ','.join(map(str, predict_tower))))
         else:
+            logger.warn("Without GPU this model will never learn! CPU is only useful for debug.")
             nr_gpu = 0
             PREDICTOR_THREAD = 1
             predict_tower = [0]

examples/README.md

@@ -11,7 +11,6 @@ Training examples with __reproducible__ and meaningful performance.
 + [ResNet for ImageNet/Cifar10/SVHN](ResNet)
 + [Holistically-Nested Edge Detection](HED)
 + [Spatial Transformer Networks on MNIST addition](SpatialTransformer)
-+ [Generative Adversarial Networks](GAN)
-+ [DisturbLabel -- I don't believe the paper](DisturbLabel)
++ [Generative Adversarial Networks variants](GAN)
 + Reinforcement learning (DQN, A3C) on [Atari games](Atari2600) and [demos on OpenAI Gym](OpenAIGym).
 + [char-rnn for fun](char-rnn)

tensorpack/models/pool.py

@@ -120,9 +120,9 @@ def FixedUnPooling(x, shape, unpool_mat=None):
 @layer_register()
 def BilinearUpSample(x, shape):
     """
-    Non-parametric bilinear upsample the input images.
+    Deterministic bilinear upsample the input images.
     :param x: input NHWC tensor
-    :param shape: an integer
+    :param shape: an integer, the upsample factor
     """
     def bilinear_conv_filler(s):
         """