cuda memory leak

[fang-2020-github]

**when this operator is called repeatedly during iterations, the GPU memory usage keeps increasing, finally the error "cuda 3D Allocation error:out of memory "occurs. And this phenomenon only occurs in ubuntu system, not in window system **

class projector_t(nn.Module): def init(self): super(projector_t, self).init()

def forward(self, sino, index, batchSize, d_angle=0):
    device_name = "cuda:1"
    device = torch.device(device_name)
    proj = Projector(forward_project=False, use_static=True, use_gpu=True, gpu_device=device, batch_size=batchSize)
    # proj = FBP(forward_FBP=True, use_static=True, use_gpu=True, gpu_device=device, batch_size=batchSize)
    numCols = 1024
    numTurns = 1024 / 1024
    numAngles = int(1024 * numTurns)
    pixelHeight = 1.12
    pixelWidth = 1.12
    # Set the number of detector rows
    numRows = 32
    phis = proj.leapct.setAngleArray(numAngles, 360.0 * numTurns)
    # Set the scanner geometry
    proj.leapct.set_conebeam(numAngles, numRows, numCols, pixelHeight, pixelWidth, 0.5 * (numRows - 1),
                             0.5 * (numCols - 1) - 4.4912, phis+d_angle, 570, 1005, tau=0.0)
    proj.leapct.set_curvedDetector()
    # Set the helical pitch.
    proj.leapct.set_normalizedHelicalPitch(1.6)
    pixels = 512
    slices = 5
    voxelWidth = 0.6
    voxelHeight = 1.0
    proj.leapct.set_volume(pixels, pixels, slices, voxelWidth, voxelHeight)
    proj.allocate_batch_data()
    g = sino
    f = proj.leapct.allocateVolume()
    f = torch.from_numpy(f).to(device)
    # proj.leapct.set_FORBILD(f, True)
    f[:] = 0.0
    # Reconstruct the data
    proj.leapct.FBP(g, f)
    # f = proj(g)
    del sino, g, proj
    return f

[kylechampley]

Weird that it only happens on Ubuntu. Try explicitly calling the python garbage collector. At the top of your script add this:

import gc

And then right after the del statements add this:

gc.collect()