TensorFlow memory use while running on GPU: why does it look like not all memory is used?

Question

This is a follow-up to my question posted here: Memory error with larger images when running convolutional neural network using TensorFlow on AWS instance g2.2xlarge

I built a CNN model in Python using TensorFlow and run it on NVIDIA GRID K520 GPU. It runs fine with 64x64 images, but produces a memory error with 128x128 images (even when input consists of only 1 image).

The error says Ran out of memory trying to allocate 2.00GiB. 2GiB is the size of my first fully-connected layer (input: 128*128*2(channels) output: 128*128 * 4 bytes = 2.14748 GB = 2.0 GiB).

From here, I can see that GRID K520 has 8GB = 7.45GiB memory. When I start running my code, I also see the output: Total memory: 3.94GiB, Free memory: 3.91GiB.

My question is, what is the relationship between all these numbers: if there are 7.45GiB of memory on GPU, why there are only 3.94GiB of total memory and most importantly, why GPU cannot allocate 2GiB memory, which is just above half of total memory? (I am not a computer scientist, so a detailed answer would be valuable.)

Some more specific information in case it is useful: I tried using allow_growth and per_process_gpu_memory_fraction. Still get the memory error, but also some memory stats (would really appreciate if someone could explain to me these numbers):

allow_growth = True
Stats: 
Limit:                  3878682624
InUse:                  2148557312
MaxInUse:               2148557312
NumAllocs:                      13
MaxAllocSize:           2147483648

allow_growth = False
Stats: 
Limit:                  3878682624
InUse:                  3878682624
MaxInUse:               3878682624
NumAllocs:                      13
MaxAllocSize:           3877822976

per_process_gpu_memory_fraction = 0.5
allow_growth = False
Stats: 
Limit:                  2116026368
InUse:                      859648
MaxInUse:                   859648
NumAllocs:                      12
MaxAllocSize:               409600

per_process_gpu_memory_fraction = 0.5
allow_growth = True
Stats: 
Limit:                  2116026368
InUse:                     1073664
MaxInUse:                  1073664
NumAllocs:                      12
MaxAllocSize:               623616

Minimal working example: with dummy training set of the same size as the images I input and only one fully-connected layer (full model code is here). This example works with input of size:

X_train = np.random.rand(1, 64, 64, 2)
Y_train = np.random.rand(1, 64, 64)

but doesn't work with input of size

X_train = np.random.rand(1, 128, 128, 2)
Y_train = np.random.rand(1, 128, 128) 

Code:

import numpy as np
import tensorflow as tf


# Dummy training set:
X_train = np.random.rand(1, 128, 128, 2)
Y_train = np.random.rand(1, 128, 128)
print('X_train.shape at input = ', X_train.shape, ", Size = ",
      X_train.shape[0] * X_train.shape[1] * X_train.shape[2]
      * X_train.shape[3])
print('Y_train.shape at input = ', Y_train.shape, ", Size = ",
      Y_train.shape[0] * Y_train.shape[1] * Y_train.shape[2])


def create_placeholders(n_H0, n_W0):

    x = tf.placeholder(tf.float32, shape=[None, n_H0, n_W0, 2], name='x')
    y = tf.placeholder(tf.float32, shape=[None, n_H0, n_W0], name='y')

    return x, y


def forward_propagation(x):

    x_temp = tf.contrib.layers.flatten(x)  # size (n_im, n_H0 * n_W0 * 2)
    n_out = np.int(x.shape[1] * x.shape[2])  # size (n_im, n_H0 * n_W0)

    # FC: input size (n_im, n_H0 * n_W0 * 2), output size (n_im, n_H0 * n_W0)
    FC1 = tf.contrib.layers.fully_connected(
        x_temp,
        n_out,
        activation_fn=tf.tanh,
        normalizer_fn=None,
        normalizer_params=None,
        weights_initializer=tf.contrib.layers.xavier_initializer(),
        weights_regularizer=None,
        biases_initializer=None,
        biases_regularizer=None,
        reuse=True,
        variables_collections=None,
        outputs_collections=None,
        trainable=True,
        scope='fc1')

    # Reshape output from FC layer into array of size (n_im, n_H0, n_W0, 1):
    FC_M = tf.reshape(FC1, [tf.shape(x)[0], tf.shape(x)[1], tf.shape(x)[2], 1])

    return FC_M


def compute_cost(FC_M, Y):

    cost = tf.square(FC_M - Y)

    return cost


def model(X_train, Y_train, learning_rate=0.0001, num_epochs=100):

    (m, n_H0, n_W0, _) = X_train.shape

    # Create Placeholders
    X, Y = create_placeholders(n_H0, n_W0)

    # Build the forward propagation
    DECONV = forward_propagation(X)

    # Add cost function to tf graph
    cost = compute_cost(DECONV, Y)

    # Backpropagation
    optimizer = tf.train.RMSPropOptimizer(learning_rate).minimize(cost)

    # Initialize all the variables globally
    init = tf.global_variables_initializer()

    # Memory config
    config = tf.ConfigProto()
    config.gpu_options.allow_growth = True

    # Start the session to compute the tf graph
    with tf.Session(config = config) as sess:

        # Initialization
        sess.run(init)

        # Training loop
        for epoch in range(num_epochs):

            _, temp_cost = sess.run([optimizer, cost],
                                    feed_dict={X: X_train, Y: Y_train})

            print ('EPOCH = ', epoch, 'COST = ', np.mean(temp_cost))


# Finally run the model
model(X_train, Y_train, learning_rate=0.00002, num_epochs=5)

Traceback:

W tensorflow/core/common_runtime/bfc_allocator.cc:274] ****************************************************************************************************
W tensorflow/core/common_runtime/bfc_allocator.cc:275] Ran out of memory trying to allocate 2.00GiB.  See logs for memory state.
W tensorflow/core/framework/op_kernel.cc:983] Internal: Dst tensor is not initialized.
E tensorflow/core/common_runtime/executor.cc:594] Executor failed to create kernel. Internal: Dst tensor is not initialized.
     [[Node: zeros = Const[dtype=DT_FLOAT, value=Tensor<type: float shape: [32768,16384] values: [0 0 0]...>, _device="/job:localhost/replica:0/task:0/gpu:0"]()]]
Traceback (most recent call last):
  File "myAutomap_MinExample.py", line 99, in <module>
    num_epochs=5)
  File "myAutomap_MinExample.py", line 85, in model
    sess.run(init)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 767, in run
    run_metadata_ptr)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 965, in _run
    feed_dict_string, options, run_metadata)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1015, in _do_run
    target_list, options, run_metadata)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1035, in _do_call
    raise type(e)(node_def, op, message)
tensorflow.python.framework.errors_impl.InternalError: Dst tensor is not initialized.
     [[Node: zeros = Const[dtype=DT_FLOAT, value=Tensor<type: float shape: [32768,16384] values: [0 0 0]...>, _device="/job:localhost/replica:0/task:0/gpu:0"]()]]

Caused by op u'zeros', defined at:
  File "myAutomap_MinExample.py", line 99, in <module>
    num_epochs=5)
  File "myAutomap_MinExample.py", line 72, in model
    optimizer = tf.train.RMSPropOptimizer(learning_rate).minimize(cost)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/training/optimizer.py", line 289, in minimize
    name=name)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/training/optimizer.py", line 403, in apply_gradients
    self._create_slots(var_list)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/training/rmsprop.py", line 103, in _create_slots
    self._zeros_slot(v, "momentum", self._name)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/training/optimizer.py", line 647, in _zeros_slot
    named_slots[var] = slot_creator.create_zeros_slot(var, op_name)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/training/slot_creator.py", line 121, in create_zeros_slot
    val = array_ops.zeros(primary.get_shape().as_list(), dtype=dtype)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/ops/array_ops.py", line 1352, in zeros
    output = constant(zero, shape=shape, dtype=dtype, name=name)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/framework/constant_op.py", line 103, in constant
    attrs={"value": tensor_value, "dtype": dtype_value}, name=name).outputs[0]
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2327, in create_op
    original_op=self._default_original_op, op_def=op_def)
  File "/home/ubuntu/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 1226, in __init__
    self._traceback = _extract_stack()

InternalError (see above for traceback): Dst tensor is not initialized.
     [[Node: zeros = Const[dtype=DT_FLOAT, value=Tensor<type: float shape: [32768,16384] values: [0 0 0]...>, _device="/job:localhost/replica:0/task:0/gpu:0"]()]]

Answer 1

It would be good, if you can upload your code or at least a minimal example in order to see what is going on. Just looking at these numbers, it seems allow_growth is working as it should, that is, it is only allocating the amount of memory that it actually needs (the 2.148 GiB you calculated above).

Also can you provide the full console output of the error you are getting. My guess is, that you are confusing a non fatal warning message from the TF resource allocator for the actual error that's causing your program to fail.

Is this similar to the message that you are seeing? W tensorflow/core/common_runtime/bfc_allocator.cc:217] Allocator (GPU_1_bfc) ran out of memory trying to allocate 2.55GiB. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory is available.

Because this is just a warning that you may ignore unless you want to optimize the runtime performance of your code. It is not something that will cause your program to fail.