def _impl_test_argminmax(self, dtype):
x = np.random.normal(scale=5.0, size=(3, 5))
x = x.astype(dtype=dtype, order='C')
x_gpu = gpuarray.to_gpu(x)
assert_allclose(misc.argmax(x_gpu, axis=0).get(), x.argmax(axis=0),
rtol=dtype_to_rtol[dtype],
atol=dtype_to_atol[dtype])
assert_allclose(misc.argmax(x_gpu, axis=1).get(), x.argmax(axis=1),
rtol=dtype_to_rtol[dtype],
atol=dtype_to_atol[dtype])
assert_allclose(misc.argmin(x_gpu, axis=0).get(), x.argmin(axis=0),
rtol=dtype_to_rtol[dtype],
atol=dtype_to_atol[dtype])
assert_allclose(misc.argmin(x_gpu, axis=1).get(), x.argmin(axis=1),
rtol=dtype_to_rtol[dtype],
atol=dtype_to_atol[dtype])
x = x.astype(dtype=dtype, order='F')
x_gpu = gpuarray.to_gpu(x)
assert_allclose(misc.argmax(x_gpu, axis=0).get(), x.argmax(axis=0),
rtol=dtype_to_rtol[dtype],
atol=dtype_to_atol[dtype])
assert_allclose(misc.argmax(x_gpu, axis=1).get(), x.argmax(axis=1),
rtol=dtype_to_rtol[dtype],
atol=dtype_to_atol[dtype])
assert_allclose(misc.argmin(x_gpu, axis=0).get(), x.argmin(axis=0),
rtol=dtype_to_rtol[dtype],
atol=dtype_to_atol[dtype])
assert_allclose(misc.argmin(x_gpu, axis=1).get(), x.argmin(axis=1),
rtol=dtype_to_rtol[dtype],
atol=dtype_to_atol[dtype])
def _impl_test_argminmax(self, dtype):
x = np.random.normal(scale=5.0, size=(3, 5))
x = x.astype(dtype=dtype, order='C')
x_gpu = gpuarray.to_gpu(x)
assert np.allclose(misc.argmax(x_gpu, axis=0).get(), x.argmax(axis=0))
assert np.allclose(misc.argmax(x_gpu, axis=1).get(), x.argmax(axis=1))
assert np.allclose(misc.argmin(x_gpu, axis=0).get(), x.argmin(axis=0))
assert np.allclose(misc.argmin(x_gpu, axis=1).get(), x.argmin(axis=1))
x = x.astype(dtype=dtype, order='F')
x_gpu = gpuarray.to_gpu(x)
assert np.allclose(misc.argmax(x_gpu, axis=0).get(), x.argmax(axis=0))
assert np.allclose(misc.argmax(x_gpu, axis=1).get(), x.argmax(axis=1))
assert np.allclose(misc.argmin(x_gpu, axis=0).get(), x.argmin(axis=0))
assert np.allclose(misc.argmin(x_gpu, axis=1).get(), x.argmin(axis=1))
def argmax(self, dim):
if self.device == 'cuda':
arg_max = misc.argmax(self.data, axis=dim, keepdims=False)
else:
arg_max = np.argmax(self.data, axis=dim)
return Tensor(
data=arg_max,
device=self.device,
d_type=np.int32,
)
def fitSlcGPU(slc, srcFatT2, t2, b1, ff):
global ROWSTEP
print("Fitting slice", slc)
yValues = dicomStack[:, :, slc, :].squeeze()
slcShape = yValues.shape
nrows = slcShape[0]
ncols = slcShape[1]
sigLen = slcShape[2]
success = False
ffParams_gpu = None
ffValues_gpu = None
if np.any(ff[:,:,slc] > 0):
useFF = True
ffParams_gpu = findmax_ff.prepareAndLoadParams(parameterCombinations)
else:
useFF = False
while not success:
try:
for r in range(0,nrows,ROWSTEP):
rowMax = min(r+ROWSTEP, nrows)
slcLin = yValues[r:rowMax,:,:].reshape(ncols*(rowMax-r), sigLen).astype(np.float32)
slcGPU = None
slcGPU = pycuda.gpuarray.to_gpu(slcLin)
slcGPU = sklinalg.multiply(slcGPU, slcGPU)
corrMatrixGPU = sklinalg.mdot(slcGPU, signalsGPU) # correlation
tryFree(slcGPU)
if useFF:
ffValues_gpu = findmax_ff.prepareAndLoadFF(ff[r:rowMax, :, slc])
corrMax = findmax_ff.findmax_gpu(corrMatrixGPU, ffValues_gpu, ffParams_gpu)
else:
corrMaxGPU = skmisc.argmax(corrMatrixGPU, 1)
corrMax = corrMaxGPU.get()
tryFree(corrMaxGPU)
tryFree(corrMatrixGPU)
tryFree(ffValues_gpu)
for row in range(r, rowMax):
for c in range(ncols):
ind = (row-r)*ncols + c
t2[row,c,slc] = parameterCombinations[corrMax[ind]][0]
b1[row,c,slc] = parameterCombinations[corrMax[ind]][1]
ff[row,c,slc] = parameterCombinations[corrMax[ind]][2]
if DOPLOT >= 1:
plotImages()
success = True
except pycuda._driver.MemoryError:
ROWSTEP -= 1
tryFree(slcGPU)
tryFree(corrMatrixGPU)
tryFree(ffValues_gpu)
gc.collect()
print("Not enough GPU Mem: decreasing ROWSTEP to", ROWSTEP)