update memory notes

README.md

@@ -36,9 +36,8 @@ Multi-GPU training is ready to use by simply changing the trainer.
 pip install --user -r requirements.txt
 pip install --user -r opt-requirements.txt (some optional dependencies)
 ```
++ Use [tcmalloc](http://goog-perftools.sourceforge.net/doc/tcmalloc.html) whenever possible: see [TF issue](https://github.com/tensorflow/tensorflow/issues/2942)
 + allow `import tensorpack` everywhere:
 ```
 export PYTHONPATH=$PYTHONPATH:`readlink -f path/to/tensorpack`
 ```
-
-

examples/DoReFa-Net/svhn-digit-dorefa.py

@@ -20,6 +20,8 @@ http://arxiv.org/abs/1606.06160
 The original experiements are performed on a proprietary framework.
 This is our attempt to reproduce it on tensorpack.
 
+You'll need tcmalloc to avoid large memory consumption: https://github.com/tensorflow/tensorflow/issues/2942
+
 This config, with (W,A,G)=(1,1,4), can reach 3.1~3.2% error after 150 epochs.
 With the GaussianDeform augmentor, it will reach 2.8~2.9%
 (we are not using this augmentor in the paper).