def make_bsplines_parallel(self):
#print (self.temps_gpu.shape, len(self.template_inds), self.template_inds[0].shape)
self.temp_temp_cpp = deconv.BatchedTemplates([deconv.Template(nzData, nzInd) for nzData, nzInd in zip(self.temps_gpu, self.template_inds)])
#print (" making template bsplines")
#fname = os.path.join(data_dir,'voltage_bsplines_'+
# str((self.chunk_id+1)*self.CONFIG.resources.n_sec_chunk_gpu) + '.npy')
#if os.path.exists(fname)==False:
if True:
# Cat; TODO: don't need to pass tensor/cuda templates to parallel function
# - can just pass the raw cpu templates
# multi-core bsplines
if self.CONFIG.resources.multi_processing:
templates_cpu = []
for template in self.temp_temp_cpp:
templates_cpu.append(template.data.cpu().numpy())
coefficients = parmap.map(transform_template_parallel, templates_cpu,
processes=self.CONFIG.resources.n_processors,
pm_pbar=False)
# single core
else:
coefficients = []
for template in self.temp_temp_cpp:
template_cpu = template.data.cpu().numpy()
coefficients.append(transform_template_parallel(template_cpu))
#np.save(fname, coefficients)
# else:
# print (" ... loading coefficients from disk")
# coefficients = np.load(fname)
#print (" ... moving coefficients to cuda objects")
coefficients_cuda = []
for p in range(len(coefficients)):
coefficients_cuda.append(deconv.Template(torch.from_numpy(coefficients[p]).cuda(), self.temp_temp_cpp[p].indices))
# print ('self.temp_temp_cpp[p].indices: ', self.temp_temp_cpp[p].indices)
# print ("self.vis_units: ", self.vis_units[p])
# coefficients_cuda.append(deconv.Template(torch.from_numpy(coefficients[p]).cuda(), self.vis_units[p]))
self.coefficients = deconv.BatchedTemplates(coefficients_cuda)
del self.template_inds
del self.temps_gpu
del self.temp_temp_cpp
del coefficients_cuda
del coefficients
torch.cuda.empty_cache()
def data_to_gpu(self):
self.peak_pts = torch.arange(-1, +2).cuda()
#norm
norm = np.sum(np.square(self.temps), (0, 1))
self.norms = torch.from_numpy(norm).float().cuda()
# spatial and temporal component of svd
self.spat_comp = torch.from_numpy(self.spat_comp).float().cuda()
self.temp_comp = torch.from_numpy(self.temp_comp).float().cuda()
print(" ... moving coefficients to cuda objects")
# load vis units
fname_vis_units = os.path.join(self.init_dir, 'vis_units.npy')
vis_units = np.load(fname_vis_units, allow_pickle=True)
#
coefficients_cuda = []
for p in range(len(self.coefficients)):
coefficients_cuda.append(
deconv.Template(
torch.from_numpy(self.coefficients[p]).float().cuda(),
torch.from_numpy(vis_units[p]).long().cuda()))
self.coefficients = deconv.BatchedTemplates(coefficients_cuda)
del coefficients_cuda
torch.cuda.empty_cache()
if self.fit_height:
self.large_units = torch.from_numpy(self.large_units).cuda()
def get_bspline_coeffs(self):
n_data, n_times, n_channels = self.templates_aligned.shape
channels = torch.arange(n_channels).cuda()
temps_torch = torch.from_numpy(-(self.templates_aligned.transpose(0, 2, 1))).cuda()
temp_cpp = deconv.BatchedTemplates([deconv.Template(temp, channels) for temp in temps_torch])
self.coeffs = deconv.BatchedTemplates([transform_template(template) for template in temp_cpp])
def transform_template(template, knots=None, prepad=7, postpad=3, order=3):
if knots is None:
knots = np.arange(len(template.data[0]) + prepad + postpad)
splines = [
fit_spline(curve, knots=knots, prepad=prepad, postpad=postpad, order=order)
for curve in template.data.cpu().numpy()
]
coefficients = np.array([spline[1][prepad-1:-1*(postpad+1)] for spline in splines], dtype='float32')
return deconv.Template(torch.from_numpy(coefficients).cuda(), template.indices)
def templates_to_bsplines(self):
print (" making template bsplines")
fname = os.path.join(self.dir_bsplines,'bsplines_'+
str((self.chunk_id+1)*self.CONFIG.deconvolution.template_update_time) + '.npy')
print ("Bspline file: ", fname)
if os.path.exists(fname)==False:
#if True:
# Cat; TODO: don't need to pass tensor/cuda templates to parallel function
# - can just pass the raw cpu templates
# multi-core bsplines
if self.CONFIG.resources.multi_processing:
templates_cpu = []
for template in self.temp_temp_cpp:
templates_cpu.append(template.data.cpu().numpy())
import parmap
coefficients = parmap.map(transform_template_parallel, templates_cpu,
processes=self.CONFIG.resources.n_processors,
pm_pbar=False)
# single core
else:
coefficients = []
for template in self.temp_temp_cpp:
template_cpu = template.data.cpu().numpy()
coefficients.append(transform_template_parallel(template_cpu))
#np.save(fname, coefficients)
else:
print (" ... loading coefficients from disk")
coefficients = np.load(fname, allow_pickle=True)
# #
print (" recomputed coefficients: ", coefficients[0].shape)
print (" recomputed coefficients: ", coefficients[0])
coefficients = np.load(fname)
print (" loaded coefficients: ", coefficients[0].shape)
print (" loaded coefficients: ", coefficients[0])
print (" ... moving coefficients to cuda objects")
coefficients_cuda = []
for p in range(len(coefficients)):
coefficients_cuda.append(deconv.Template(torch.from_numpy(coefficients[p]).cuda(), self.temp_temp_cpp[p].indices))
self.coefficients = deconv.BatchedTemplates(coefficients_cuda)
del self.temp_temp
del self.temp_temp_cpp
del coefficients_cuda
del coefficients
torch.cuda.empty_cache()
def initialize_cpp(self):
# make a list of pairwise batched temp_temp and their vis_units
# Cat: TODO: this isn't really required any longer;
# - the only thing required from this in parallel bsplines function is
# self.temp_temp_cpp.indices - is self.vis_units
#
self.temp_temp_cpp = deconv.BatchedTemplates([
deconv.Template(nzData, nzInd)
for nzData, nzInd in zip(self.temp_temp, self.vis_units)
])
def make_bsplines(self):
#print (" making bsplines...")
# make template objects
self.templates = deconv.BatchedTemplates([
deconv.Template(vals, inds)
for vals, inds in zip(self.template_vals, self.template_inds)
])
# make bspline objects
def fit_spline(curve, knots=None, prepad=0, postpad=0, order=3):
if knots is None:
knots = np.arange(len(curve) + prepad + postpad)
return splrep(knots,
np.pad(curve, (prepad, postpad), mode='symmetric'),
k=order)
def transform_template(template,
knots=None,
prepad=7,
postpad=3,
order=3):
if knots is None:
knots = np.arange(len(template.data[0]) + prepad + postpad)
splines = [
fit_spline(curve,
knots=knots,
prepad=prepad,
postpad=postpad,
order=order)
for curve in template.data.cpu().numpy()
]
coefficients = np.array([
spline[1][prepad - 1:-1 * (postpad + 1)] for spline in splines
],
dtype='float32')
return deconv.Template(
torch.from_numpy(coefficients).cuda(), template.indices)
# make bspline coefficients
self.coefficients = deconv.BatchedTemplates(
[transform_template(template) for template in self.templates])