python - Tensorflow runs out of memory while computing: how to find memory leaks?

Question

I'm iteratively deepdreaming images in a directory using the Google's TensorFlow DeepDream implementation (https://github.com/tensorflow/tensorflow/blob/master/tensorflow/examples/tutorials/deepdream/deepdream.ipynb).

My code is as follows:

model_fn = tensorflow_inception_graph.pb

# creating TensorFlow session and loading the model
graph = tf.Graph()
sess = tf.InteractiveSession(graph=graph)
with tf.gfile.FastGFile(model_fn, 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
t_input = tf.placeholder(np.float32, name='input') # define the input tensor
imagenet_mean = 117.0
t_preprocessed = tf.expand_dims(t_input-imagenet_mean, 0)
tf.import_graph_def(graph_def, {'input':t_preprocessed})



def render_deepdream(t_obj, img0=img_noise,
                     iter_n=10, step=1.5, octave_n=4, octave_scale=1.4):
    t_score = tf.reduce_mean(t_obj) # defining the optimization objective
    t_grad = tf.gradients(t_score, t_input)[0] # behold the power of automatic differentiation!

    # split the image into a number of octaves
    img = img0
    octaves = []
    for i in range(octave_n-1):
        hw = img.shape[:2]
        lo = resize(img, np.int32(np.float32(hw)/octave_scale))
        hi = img-resize(lo, hw)
        img = lo
        octaves.append(hi)

    # generate details octave by octave
    for octave in range(octave_n):
        if octave>0:
            hi = octaves[-octave]
            img = resize(img, hi.shape[:2])+hi
        for i in range(iter_n):
            g = calc_grad_tiled(img, t_grad)
            img += g*(step / (np.abs(g).mean()+1e-7))
            #print('.',end = ' ')
        #clear_output()
    #showarray(img/255.0)
    return img/255.0


def morphPicture(filename1,filename2,blend,width):
    img1 = PIL.Image.open(filename1)
    img2 = PIL.Image.open(filename2)
    if width is not 0:
        img2 = resizePicture(filename2,width)
    finalImage= PIL.Image.blend(img1, img2, blend)
    del img1
    del img2
    return finalImage

def save_array(arr, name,direc, ext="png"):
    img = np.uint8(np.clip(arr, 0, 1)*255)
    img =cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
    cv2.imwrite("{d}/{n}.{e}".format(d=direc, n=name, e=ext), img)
    del img

framesDir = "my directory"
os.chdir(framesDir)

outputDir ="my directory"
for file in os.listdir(framesDir):
    img0 = PIL.Image.open(file)
    img0 = np.float32(img0)
    dreamedImage = render_deepdream(tf.square(T('mixed4c')),img0,iter_n=3,octave_n=6)
    save_array(dreamedImage,1,outputDir,'jpg')
    break

i=1
j=0
with tf.device('/gpu:0'):
    for file in os.listdir(framesDir):
        if j<=1: #already processed first image so we skip it here
            j+=1
            continue
        else:
            dreamedImage = "my directory"+str(i)+'.jpg' # get the previous deep dreamed frame

            img1 = file # get the next undreamed frame

            morphedImage = morphPicture(dreamedImage,img1,0.5,0) #blend the images
            morphedImage=np.float32(morphedImage)
            dreamedImage = render_deepdream(tf.square(T('mixed4c')),morphedImage,iter_n=3,octave_n=6) #deep dream a 
                                                                                                    #blend of the two frames
            i+=1
            save_array(dreamedImage,i,outputDir,'jpg') #save the dreamed image
            del dreamedImage
            del img1
            del morphedImage


            time.sleep(0.5)

Whenever I run the code for more than an hour, the script stops with a MemoryError. I'm assuming there must be a memory leak somewhere, but I'm unable to find it. I thought that by including multiple del statements, I would get rid of the objects that were clogging up the RAM/CPU, but it doesn't seem to be working.

Is there an obvious build up of objects that I am missing within my code? Or is the build up somewhere beneath my code, i.e. within tensorflow?

Any help/suggestions would be much appreciated. Thanks.

FYI there are 901 images in the directory. I am using Windows 7 with NVIDIA GeForce GTX 980 Ti.

Answer 1

99% of the time, when using tensorflow, "memory leaks" are actually due to operations that are continuously added to the graph while iterating — instead of building the graph first, then using it in a loop.

The fact that you specify a device (with tf.device('/gpu:0) for your loop is a hint that it is the case: you typically specify a device for new nodes as this does not affect nodes that are already defined.

Fortunately, tensorflow has a convenient tool to spot those errors: tf.Graph.finalize. When called, this function prevents further nodes to be added to your graph. It is good practice to call this function before iterating.

So in your case I would call tf.get_default_graph().finalize() before your loop and look for any error it may throw.