Tensorflow: using an input-pipeline (.csv) as a dictionary for training

问题

I'm trying to train a model on a .csv dataset (5008 columns, 533 rows). I'm using a textreader to parse the data into two tensors, one holding the data to train on [example] and one holding the correct labels [label]:

def read_my_file_format(filename_queue):
    reader = tf.TextLineReader()
    key, record_string = reader.read(filename_queue)
    record_defaults = [[0.5] for row in range(5008)]

    #Left out most of the columns for obvious reasons
    col1, col2, col3, ..., col5008 = tf.decode_csv(record_string, record_defaults=record_defaults)
    example = tf.stack([col1, col2, col3, ..., col5007])
    label = col5008
    return example, label

def input_pipeline(filenames, batch_size, num_epochs=None):
    filename_queue = tf.train.string_input_producer(filenames, num_epochs=num_epochs, shuffle=True)
    example, label = read_my_file_format(filename_queue)
    min_after_dequeue = 10000
    capacity = min_after_dequeue + 3 * batch_size
    example_batch, label_batch = tf.train.shuffle_batch([example, label], batch_size=batch_size, capacity=capacity, min_after_dequeue=min_after_dequeue)
    return example_batch, label_batch

This part is working, when executing something like:

with tf.Session() as sess:
    ex_b, l_b = input_pipeline(["Tensorflow_vectors.csv"], 10, 1)
    print("Test: ",ex_b)

my result is Test: Tensor("shuffle_batch:0", shape=(10, 5007), dtype=float32)

So far this seems fine to me. Next I've created a simple model consising of two hidden layers (512 and 256 nodes respectively). Where things go wrong is when I'm trying to train the model:

batch_x, batch_y = input_pipeline(["Tensorflow_vectors.csv"], batch_size)
_, cost = sess.run([optimizer, cost], feed_dict={x: batch_x.eval(), y: batch_y.eval()})

I've based this approach on this example that uses the MNIST database. However, when I'm executing this, even when I'm just using batch_size = 1, Tensorflow just hangs. If I leave out the .eval() functions that should get the actual data from the tensors, I get the following response:

TypeError: The value of a feed cannot be a tf.Tensor object. Acceptable feed values include Python scalars, strings, lists, or numpy ndarrays.

Now this I can understand, but I don't understand why the program hangs when I do include the .eval() function and I don't know where I could find any information about this issue.

EDIT: I included the most recent version of my entire script here. The program still hangs even though I implemented (as far as I know correctly) the solution that was offered by vijay m

回答1:

As the error says, you are trying to feed a tensor to feed_dict. You have defined a input_pipeline queue and you cant pass it as feed_dict. The proper way for the data to be passed to the model and train is shown in the code below:

 # A queue which will return batches of inputs 
 batch_x, batch_y = input_pipeline(["Tensorflow_vectors.csv"], batch_size)

 # Feed it to your neural network model: 
 # Every time this is called, it will pull data from the queue.
 logits = neural_network(batch_x, batch_y, ...)

 # Define cost and optimizer
 cost = ...
 optimizer = ...

 # Evaluate the graph on a session:
 with tf.Session() as sess:
    init_op = ...
    sess.run(init_op)

    # Start the queues
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)

    # Loop through data and train
    for ( loop through steps ):
        _, cost = sess.run([optimizer, cost])

    coord.request_stop()
    coord.join(threads)

来源：https://stackoverflow.com/questions/44862754/tensorflow-using-an-input-pipeline-csv-as-a-dictionary-for-training

标签

tensorflow

tensorflow-gpu

tensor