How to *actually* read CSV data in TensorFlow?

I think you are mixing up imperative and graph-construction parts here. The operation tf.train.shuffle_batch creates a new queue node, and a single node can be used to process the entire dataset. So I think you are hanging because you created a bunch of shuffle_batch queues in your for loop and didn’t start queue runners for them.

Normal input pipeline usage looks like this:

1. Add nodes like shuffle_batch to input pipeline
2. (optional, to prevent unintentional graph modification) finalize graph

— end of graph construction, beginning of imperative programming —

3. tf.start_queue_runners
4. while(True): session.run()

To be more scalable (to avoid Python GIL), you could generate all of your data using TensorFlow pipeline. However, if performance is not critical, you can hook up a numpy array to an input pipeline by using slice_input_producer. Here’s an example with some Print nodes to see what’s going on (messages in Print go to stdout when node is run)

tf.reset_default_graph()

num_examples = 5
num_features = 2
data = np.reshape(np.arange(num_examples*num_features), (num_examples, num_features))
print data

(data_node,) = tf.slice_input_producer([tf.constant(data)], num_epochs=1, shuffle=False)
data_node_debug = tf.Print(data_node, [data_node], "Dequeueing from data_node ")
data_batch = tf.batch([data_node_debug], batch_size=2)
data_batch_debug = tf.Print(data_batch, [data_batch], "Dequeueing from data_batch ")

sess = tf.InteractiveSession()
sess.run(tf.initialize_all_variables())
tf.get_default_graph().finalize()
tf.start_queue_runners()

try:
  while True:
    print sess.run(data_batch_debug)
except tf.errors.OutOfRangeError as e:
  print "No more inputs."

You should see something like this

[[0 1]
 [2 3]
 [4 5]
 [6 7]
 [8 9]]
[[0 1]
 [2 3]]
[[4 5]
 [6 7]]
No more inputs.

The “8, 9” numbers didn’t fill up the full batch, so they didn’t get produced. Also tf.Print are printed to sys.stdout, so they show up in separately in Terminal for me.

PS: a minimal of connecting batch to a manually initialized queue is in github issue 2193

Also, for debugging purposes you might want to set timeout on your session so that your IPython notebook doesn’t hang on empty queue dequeues. I use this helper function for my sessions

def create_session():
  config = tf.ConfigProto(log_device_placement=True)
  config.gpu_options.per_process_gpu_memory_fraction=0.3 # don't hog all vRAM
  config.operation_timeout_in_ms=60000   # terminate on long hangs
  # create interactive session to register a default session
  sess = tf.InteractiveSession("", config=config)
  return sess

Scalability Notes:

1. tf.constant inlines copy of your data into the Graph. There’s a fundamental limit of 2GB on size of Graph definition so that’s an upper limit on size of data
2. You could get around that limit by using v=tf.Variable and saving the data into there by running v.assign_op with a tf.placeholder on right-hand side and feeding numpy array to the placeholder (feed_dict)
3. That still creates two copies of data, so to save memory you could make your own version of slice_input_producer which operates on numpy arrays, and uploads rows one at a time using feed_dict