Deep learning Fizzbuzz

Question

Inspired by the job-interview with Joel Grus, the goal of this challenge is to write a tensorflow (or other deep/machine learning) program that learns Fizzbuzz and correctly prints out the answers to the positive integers less than 1000.

You can assume there are files named train.csv and test.csv and each contain a sorted list of sequential integers and the fizzbuzz answer:

 ...
 100, buzz
 101, 101
 102, fizz
 103, 103
 104, 104
 105, buzz
 ...
 150000, fizzbuzz

test.csv spans 1-1000 and train.csv spans 1001-150000.

Rules

1. You must not hard-code the rules to Fizzbuzz anywhere in your program. The output must be from a machine learned representation that is learned while running the code.
2. You must utilize train.csv in the training set and check your output against test.csv. You cannot use test.csv during training.
3. You must get all outputs correct from test.csv (but as is case with deep-learning, we'll allow your code to fail this rule no more than 5% of the time).
4. You may use any language and any external module (eg. python/tensorflow) as long they explicitly perform some kind of learning. Reference both the language and the module in the title of your post.
5. This is a popularity contest, so the submission with the most votes after one week wins.