mention imagenet synsets.

docs/tutorial/philosophy/dataflow.md

@@ -152,6 +152,7 @@ or when you need to filter your data on the fly.
    1. You do batch training
    1. You use a constant batch size
    1. Indices are sufficient to determine which samples to batch together
+   1. Multiprocessing is a better parallelization strategy than multithreading.
 
    None of these are necessarily true.
 

tensorpack/dataflow/dataset/ilsvrc.py

@@ -179,6 +179,7 @@ class ILSVRC12Files(RNGDataFlow):
 class ILSVRC12(ILSVRC12Files):
     """
     Produces uint8 ILSVRC12 images of shape [h, w, 3(BGR)], and a label between [0, 999].
+    The label map follows the synsets.txt file in http://dl.caffe.berkeleyvision.org/caffe_ilsvrc12.tar.gz.
     """
     def __init__(self, dir, name, meta_dir=None,
                  shuffle=None, dir_structure=None):