def thunk():
mask_idx = inputs[0][0]
image = inputs[1][0]
batch_size = min(mask_idx.shape[0], image.shape[0])
assert shape_ok(mask_idx.shape)
assert shape_ok(image.shape)
mask_idx = to_gpuarray(mask_idx)
image = to_gpuarray(image)
s = mask_idx.shape[3]
assert mask_idx.shape[2] == mask_idx.shape[3], \
"height and width must be equal"
sdata_shape = (3 * len(MASK), batch_size, 1, s, s)
self._sdata = pycuda_zeros(self._sdata, sdata_shape)
blocks_max = 32
blocks_s = min(blocks_max, s)
grid_s = math.ceil(s / blocks_max)
grid = (batch_size, grid_s, grid_s)
block = (1, blocks_s, blocks_s)
image_mask_split(mask_idx,
image,
np.int32(batch_size),
np.int32(s),
self._sdata,
block=block,
grid=grid)
sdata_as_theano = to_cudandarray(self._sdata)
m = len(MASK)
outputs[0][0] = sdata_as_theano[:m]
outputs[1][0] = sdata_as_theano[m:2 * m]
outputs[2][0] = sdata_as_theano[2 * m:]
def thunk():
mask_idx = inputs[0][0]
image = inputs[1][0]
batch_size = min(mask_idx.shape[0], image.shape[0])
assert shape_ok(mask_idx.shape)
assert shape_ok(image.shape)
mask_idx = to_gpuarray(mask_idx)
image = to_gpuarray(image)
s = mask_idx.shape[3]
assert mask_idx.shape[2] == mask_idx.shape[3], \
"height and width must be equal"
sdata_shape = (3*len(MASK), batch_size, 1, s, s)
self._sdata = pycuda_zeros(self._sdata, sdata_shape)
blocks_max = 32
blocks_s = min(blocks_max, s)
grid_s = math.ceil(s / blocks_max)
grid = (batch_size, grid_s, grid_s)
block = (1, blocks_s, blocks_s)
image_mask_split(mask_idx, image, np.int32(batch_size),
np.int32(s), self._sdata,
block=block, grid=grid)
sdata_as_theano = to_cudandarray(self._sdata)
m = len(MASK)
outputs[0][0] = sdata_as_theano[:m]
outputs[1][0] = sdata_as_theano[m:2*m]
outputs[2][0] = sdata_as_theano[2*m:]
def test_to_cudandarray():
px = pycuda.gpuarray.zeros((3, 4, 5), 'float32')
cx = to_cudandarray(px)
assert isinstance(cx, cuda.CudaNdarray)
assert numpy.allclose(px.get(), numpy.asarray(cx))
assert px.dtype == cx.dtype
assert px.shape == cx.shape
assert all(numpy.asarray(cx._strides) * 4 == px.strides)
try:
px = pycuda.gpuarray.zeros((3, 4, 5), 'float64')
to_cudandarray(px)
assert False
except ValueError:
pass
try:
to_cudandarray(numpy.zeros(4))
assert False
except ValueError:
pass
def test_to_cudandarray():
px = pycuda.gpuarray.zeros((3, 4, 5), 'float32')
cx = to_cudandarray(px)
assert isinstance(cx, cuda.CudaNdarray)
assert numpy.allclose(px.get(),
numpy.asarray(cx))
assert px.dtype == cx.dtype
assert px.shape == cx.shape
assert all(numpy.asarray(cx._strides) * 4 == px.strides)
try:
px = pycuda.gpuarray.zeros((3, 4, 5), 'float64')
to_cudandarray(px)
assert False
except ValueError:
pass
try:
to_cudandarray(numpy.zeros(4))
assert False
except ValueError:
pass
def thunk():
grad = outputs[0][0]
mask_idx = inputs[0][0]
assert shape_ok(mask_idx.shape)
s = mask_idx.shape[3]
block_dim = min(32, s)
grid_dim = math.ceil(s / block_dim)
mask_idx = to_gpuarray(mask_idx, copyif=True)
image = inputs[1][0]
assert shape_ok(image.shape)
image = to_gpuarray(image, copyif=True)
batch_size = min(mask_idx.shape[0], image.shape[0])
grad_shape = (batch_size, 1, s, s)
grad = pycuda_zeros(grad, grad_shape)
grid = (batch_size, grid_dim, grid_dim)
block = (1, block_dim, block_dim)
if "sum" in self.connected and "pow" in self.connected:
og_sum = to_gpuarray(inputs[2][0], copyif=True)
og_pow = to_gpuarray(inputs[3][0], copyif=True)
image_mask_split_grad(mask_idx,
image,
og_sum,
og_pow,
np.int32(batch_size),
np.int32(s),
grad,
block=block,
grid=grid)
elif "sum" in self.connected:
og_sum = to_gpuarray(inputs[2][0], copyif=True)
image_mask_split_grad(mask_idx,
image,
og_sum,
np.int32(batch_size),
np.int32(s),
grad,
block=block,
grid=grid)
elif "pow" in self.connected:
og_pow = to_gpuarray(inputs[2][0], copyif=True)
image_mask_split_grad(mask_idx,
image,
og_pow,
np.int32(batch_size),
np.int32(s),
grad,
block=block,
grid=grid)
outputs[0][0] = to_cudandarray(grad)
def thunk():
grad = outputs[0][0]
mask_idx = inputs[0][0]
assert shape_ok(mask_idx.shape)
s = mask_idx.shape[3]
block_dim = min(32, s)
grid_dim = math.ceil(s / block_dim)
mask_idx = to_gpuarray(mask_idx, copyif=True)
image = inputs[1][0]
assert shape_ok(image.shape)
image = to_gpuarray(image, copyif=True)
batch_size = min(mask_idx.shape[0], image.shape[0])
grad_shape = (batch_size, 1, s, s)
grad = pycuda_zeros(grad, grad_shape)
grid = (batch_size, grid_dim, grid_dim)
block = (1, block_dim, block_dim)
if "sum" in self.connected and "pow" in self.connected:
og_sum = to_gpuarray(inputs[2][0], copyif=True)
og_pow = to_gpuarray(inputs[3][0], copyif=True)
image_mask_split_grad(
mask_idx, image, og_sum, og_pow,
np.int32(batch_size), np.int32(s), grad,
block=block, grid=grid)
elif "sum" in self.connected:
og_sum = to_gpuarray(inputs[2][0], copyif=True)
image_mask_split_grad(
mask_idx, image, og_sum,
np.int32(batch_size), np.int32(s), grad,
block=block, grid=grid)
elif "pow" in self.connected:
og_pow = to_gpuarray(inputs[2][0], copyif=True)
image_mask_split_grad(
mask_idx, image, og_pow,
np.int32(batch_size), np.int32(s), grad,
block=block, grid=grid)
outputs[0][0] = to_cudandarray(grad)