Superpoint excessive memory usage

[Eunchan24]

Hello,
As the title indicates, all memory seems to be occupied when loading the superpoint model.

The image below is the gpu state when only the superpoint model is loaded.

superpoint_extract.py
Line 40 : tf1.saved_model.loader.load(self._sess, [tf1.saved_model.tag_constants.SERVING
It doesn't load as much as the size of the model on this line and seems to be using all of the GPUs.

[kke19]

HI,I have the same problem, do you have any good solution now? @Eunchan24

[kke19]

Hi,
When testing, I only loaded the original og model, and my all memory was full without loading other models. This seems to be a problem when loading the og model.
This situation also occurs in graphics cards with different display memory.
In my scenario, both 3090 (24G) and L20 (48G) basically filled up the memory.

[Eunchan24]

Hi, my error case user up all of my GPUs When loading superpoint models.
A temporary fix is that I can use set_memory_growth() before loading Omniglue to make it flexible without using up all of my GPUs.
omniglue_extract.py
Add the line below 36 lines

    physical_devices = tf.config.list_physical_devices('GPU')

    for gpu in physical_devices:
        try:
            tf.config.experimental.set_memory_growth(gpu, True)
            print(f"Memory growth set for {gpu}")
        except RuntimeError as e:
            print(e)

I'm not sure if this is the correct solution, but I'm looking for a more complete solution.