Memory rising for each execution of algorithm

[neisserBOT]

First, it's a great job.

I have this issue each time I process an image. The memory is rising till filling out, I tried to use session.close() or tf.keras.backend.clear_session but I don't have results. Is there any way to stop the rising o refresh the memory each time process the algorithm?

Thank you in advance

[alexanderkroner]

Hey, I did not encounter this issue before but does this behavior occur when you execute the command python main.py test -d DATA -p PATH, where PATH represents a folder with images? If not, how exactly do you initialize the model and process the images?

[neisserBOT]

This is the portion of code. Every time the function executes an image, the memory is rising. I don't know if is the version of 1.13.1 of TensorFlow or another library.

def saliency_model(img):
    #model_name = "model_%s_%s.pb" % (dataset, device)
    model_name = "model_salicon_cpu.pb"

    with tf.gfile.Open(model_name, "rb") as file:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(file.read())

    ## define the placeholder for image
    input_img = tf.placeholder(tf.float32, (None, None, None, 3))

    ## define the input and output of the graph
    [predicted_maps] = tf.import_graph_def(graph_def,
                                        input_map={"input": input_img},
                                        return_elements=["output:0"])

    ## read any image
    img_input = cv2.resize(img, (320, 240))
    img_input = cv2.cvtColor(img_input, cv2.COLOR_BGR2RGB)
    img_input = img_input[np.newaxis, :, :, :] ## reshape image (1, 240, 320, 3)

    ## run session   
    with tf.Session() as sess:
        ## send image to the graph
        saliency = sess.run(predicted_maps,
                            feed_dict={input_img: img_input})

        ## reshape image (240, 320, 3)
        saliency = cv2.cvtColor(saliency.squeeze(),
                                cv2.COLOR_GRAY2BGR)
        saliency = np.uint8(saliency * 255)

    saliency = cv2.resize(saliency, (400, 300))

    tf.keras.backend.clear_session()
    gc.collect()

    return saliency

This is the output of the memory.

• First image:

              total        used        free      shared  buff/cache   available
Mem:           3933        1030         804           2        2098        2620
Swap:             0           0           0

• Second image:

              total        used        free      shared  buff/cache   available
Mem:           3933        1183         651           2        2098        2468
Swap:             0           0           0

• Third image:

              total        used        free      shared  buff/cache   available
Mem:           3933        1204         630           2        2098        2446
Swap:             0           0           0

What should I do? Thank you in advance

[alexanderkroner]

It looks like you call the function saliency_model() every time you want to process an image, which will repeatedly load the model, create a session, etc. I can imagine that this could lead to the memory leak you observe.

Try to instead load the model once, create a single session, and within that session iterate over your images. For example:

graph_def = tf.GraphDef()

with tf.gfile.Open("model_salicon_gpu.pb", "rb") as file:
    graph_def.ParseFromString(file.read())

input_plhd = tf.placeholder(tf.float32, (None, None, None, 3))

[predicted_maps] = tf.import_graph_def(graph_def,
                                       input_map={"input": input_plhd},
                                       return_elements=["output:0"])

with tf.Session() as sess:
    for img in img_list:
        img = cv2.resize(img, (320, 240))
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        img = img[np.newaxis, :, :, :]

        saliency = sess.run(predicted_maps,
                            feed_dict={input_plhd: img})

        saliency = cv2.cvtColor(saliency.squeeze(),
                                cv2.COLOR_GRAY2BGR)

        saliency = np.uint8(saliency * 255)
        saliency = cv2.resize(saliency, (400, 300))

[neisserBOT]

Great Thank you very much