def generate_roadmap_parallel(samples, env, max_dist, leafsize, knn):
"""Parallelized roadmap generator """
n_sample = len(samples)
leafsize = knn
if len(samples) < leafsize: leafsize = len(samples) - 1
import sharedmem
sample_ids = np.arange(n_sample, dtype='i')
roadmap = sharedmem.full((n_sample, knn), 0)
# Start multi processing over samples
with sharedmem.MapReduce() as pool:
if n_sample % sharedmem.cpu_count() == 0:
chunksize = n_sample / sharedmem.cpu_count()
else:
chunksize = n_sample / sharedmem.cpu_count() + 1
def work(i):
skdtree = KDTree(samples, leafsize=leafsize)
sub_sample_ids = sample_ids[slice(i, i + chunksize)]
for j, sub_sample_id in enumerate(sub_sample_ids):
x = samples[sub_sample_id]
try:
inds, dists = skdtree.search(x, k=leafsize)
except:
print "skdtree search failed"
sys.exit()
edge_id = []
append = edge_id.append
for ii, (ind, dist) in enumerate(zip(inds, dists)):
if dist > max_dist: break # undirected
if len(edge_id) >= knn: break # directed?
append(ind)
# to complement fewer number of edges for vectorized valueiteration
if len(edge_id) < knn:
for ii in range(0, len(inds)):
#for ind in edge_id:
# edge_id.append(ind)
# if len(edge_id) >= knn: break
append(inds[0])
if len(edge_id) >= knn: break
assert len(
edge_id
) <= leafsize, "fewer leaves than edges {} (dists={})".format(
len(edge_id), dists[:len(edge_id)])
for k in range(len(edge_id)):
roadmap[sub_sample_id][k] = edge_id[k]
pool.map(work, range(0, n_sample, chunksize)) #, reduce=reduce)
# convert sharedmem array to list
roadmap = np.array(roadmap).astype(int)
skdtree = None #KDTree(samples, leafsize=leafsize)
return roadmap.tolist(), skdtree
def argsort(ar):
min = minimum.reduce(ar)
max = maximum.reduce(ar)
nchunk = sharedmem.cpu_count() * 2
#bins = numpy.linspace(min, max, nchunk, endpoint=True)
step = 1.0 * (max - min) / nchunk
bins = numpy.array(
1.0 * numpy.arange(nchunk + 1) * (max - min) / nchunk + min, min.dtype)
dig = digitize(ar, bins)
binlength = bincount(dig, minlength=len(bins) + 1)
binoffset = numpy.cumsum(binlength)
out = sharedmem.empty(len(ar), dtype='intp')
with sharedmem.MapReduce() as pool:
def work(i):
# we can do this a lot faster
# but already having pretty good speed.
ind = numpy.nonzero(dig == i + 1)[0]
myar = ar[ind]
out[binoffset[i]:binoffset[i + 1]] = ind[myar.argsort()]
pool.map(work, range(nchunk))
return out
def argsort(ar):
min = minimum.reduce(ar)
max = maximum.reduce(ar)
nchunk = sharedmem.cpu_count() * 2
#bins = numpy.linspace(min, max, nchunk, endpoint=True)
step = 1.0 * (max - min) / nchunk
bins = numpy.array(
1.0 * numpy.arange(nchunk + 1) * (max - min) / nchunk + min,
min.dtype)
dig = digitize(ar, bins)
binlength = bincount(dig, minlength=len(bins) + 1)
binoffset = numpy.cumsum(binlength)
out = sharedmem.empty(len(ar), dtype='intp')
with sharedmem.MapReduce() as pool:
def work(i):
# we can do this a lot faster
# but already having pretty good speed.
ind = numpy.nonzero(dig == i + 1)[0]
myar = ar[ind]
out[binoffset[i]:binoffset[i+1]] = ind[myar.argsort()]
pool.map(work, range(nchunk))
return out
def __init__(self, filename):
self.filename = filename
self.fileObject = h5py.File(self.filename)
self.dset = self.fileObject['channel_data']
self.isSnapshot = self.fileObject.attrs['ph_flags'][0] & (1 << 6)
self.nsamples, self.nchan = self.dset.shape
self.time_ms = np.arange(self.nsamples) * MS_PER_SEC / SAMPLE_RATE
self.timeMin = np.min(self.time_ms)
self.timeMax = np.max(self.time_ms)
self.boardID = self.fileObject.attrs['board_id'][0]
if self.isSnapshot:
self.cookie = None
else:
self.cookie = self.fileObject.attrs['experiment_cookie'][0]
chipAliveMask = self.fileObject['chip_live'][0]
self.chipList = [
i for i in range(NCHIPS) if (chipAliveMask & (0x1 << i))
]
self.isImported = False
self.sliceImported = False
self.sliceBeenFiltered = False
self.ncpu = sharedmem.cpu_count()
self.slice_nsamples = 0
self.slice_nchans = None
def fstack_mp(img, fmap):
img_stacked = shmem.empty(img.shape[0:2], dtype='uint16')
# This implementation is faster than breaking each image plane up for parallel processing
def do_work(x):
index = ne.evaluate("fmap==x")
img_stacked[index] = img[:, :, x][index]
index = ne.evaluate("(fmap > x) & (fmap < x+1)")
A = fmap[index]
B = img[:, :, x+1][index]
C = img[:, :, x][index]
img_stacked[index] = ne.evaluate("(A-x) * B + (x+1-A) * C")
with shmem.MapReduce(np=img.shape[2]-1) as pool:
pool.map(do_work, range(img.shape[2]-1))
last_ind = img.shape[2]-1
index = ne.evaluate("fmap == last_ind")
num_proc = shmem.cpu_count()
edges = get_edges(img, num_proc)
def mp_assignment(x):
img_stacked[edges[x]:edges[x+1],:][index[edges[x]:edges[x+1],:]] = img[edges[x]:edges[x+1], :, -1]\
[index[edges[x]:edges[x+1], :]]
with shmem.MapReduce(np=num_proc) as pool:
pool.map(mp_assignment, range(num_proc))
return img_stacked
def sharedmem_pool(total_cores, numexpr=True):
# see https://stackoverflow.com/questions/15639779
global AFFINITY_FLAG
if not AFFINITY_FLAG:
AFFINITY_FLAG = True
os.system("taskset -p 0xfff %d" % os.getpid())
if total_cores is None:
total_cores = sm.cpu_count()
return sm.MapReduce(np=good_process_number(total_cores, numexpr))
def regularizing_objective_function(parameter, bundle): # pragma: no cover
# attach the guess for tau to each of the voxels in the bundle
for voxel in bundle:
model = voxel[1]
model.parameter = parameter
# fit each of the voxels
num_cpus = sharedmem.cpu_count()-1
with sharedmem.Pool(np=num_cpus) as pool:
output = pool.map(parallel_fit, bundle)
return output
def regularizing_objective_function(parameter, bundle): # pragma: no cover
# attach the guess for tau to each of the voxels in the bundle
for voxel in bundle:
model = voxel[1]
model.parameter = parameter
# fit each of the voxels
num_cpus = sharedmem.cpu_count()-1
with sharedmem.Pool(np=num_cpus) as pool:
output = pool.map(parallel_fit, bundle)
return output
def get_features_from_states(env, states, feature_fn):
import sharedmem
n_states = len(states)
feat_len = len(feature_fn(env, states[0]))
state_ids = np.arange(n_states, dtype='i')
features = sharedmem.full((n_states, feat_len), 0.)
# Start multi processing over support states
with sharedmem.MapReduce() as pool:
if n_states % sharedmem.cpu_count() == 0:
chunksize = n_states / sharedmem.cpu_count()
else:
chunksize = n_states / sharedmem.cpu_count() + 1
def work(i):
s_ids = state_ids[slice(i, i + chunksize)]
for j, s_id in enumerate(s_ids):
s = states[s_id] # state id in states
features[s_id] = feature_fn(env, s)
pool.map(work, range(0, n_states, chunksize)) #, reduce=reduce)
return np.array(features)
def get_fmap(img):
num_proc = shmem.cpu_count()
log_kernel = get_log_kernel(11, 2)
se = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (17, 17))
def mp_imgproc(x):
bound_in = (edges[x]-(edges[x] > 0)*50, edges[x+1] + (edges[x+1] < img.shape[0]) *50 )
bound_out = (50 if edges[x] > 0 else 0, None if edges[x+1] == img.shape[0] else -50)
part_img = cv2.filter2D(img[bound_in[0]:bound_in[1], :, ii].astype('single'), -1, log_kernel)
part_img = cv2.dilate(part_img, se)
img_filtered[edges[x]:edges[x+1], :] = part_img[bound_out[0]:bound_out[1], :]
def mp_gaussblur(x):
bound_in = (edges[x]-(edges[x] > 0)*50, edges[x+1] + (edges[x+1] < img.shape[0]) *50 )
bound_out = (50 if edges[x] > 0 else 0, None if edges[x+1] == img.shape[0] else -50)
part_img = cv2.GaussianBlur(fmap[bound_in[0]:bound_in[1], :], (31, 31), 6)
fmap[edges[x]:edges[x+1], :] = part_img[bound_out[0]:bound_out[1], :]
log_response = shmem.empty(img.shape[0:2], dtype='single')
fmap = shmem.empty(img.shape[0:2], dtype='single')
edges = get_edges(img, num_proc)
def mp_assignment_1(x):
log_response[edges[x]:edges[x+1],:] = img_filtered[edges[x]:edges[x+1],:]
def mp_assignment_2(x):
fmap[index[edges[x]:edges[x+1],:]] = ii
for ii in range(img.shape[2]):
img_filtered = shmem.empty((img.shape[0], img.shape[1]), dtype='single')
with shmem.MapReduce(np=num_proc) as pool:
pool.map(mp_imgproc, range(num_proc))
index = ne.evaluate("img_filtered > log_response")
with shmem.MapReduce(np=num_proc) as pool:
pool.map(mp_assignment_1, range(num_proc))
# log_response[index] = img_filtered[index]
with shmem.MapReduce(np=num_proc) as pool:
pool.map(mp_assignment_2, range(num_proc))
with shmem.MapReduce(np=num_proc) as pool:
pool.map(mp_gaussblur, range(num_proc))
return fmap
def fstack_mp_new(img, fmap):
img_stacked = shmem.empty(img.shape[0:2], dtype='uint16')
indexl = shmem.empty(img.shape[0:2], dtype='bool')
edges = get_edges(img, 16)
# This implementation is faster than breaking each image plane up for parallel processing
def do_work(x):
if x!=img.shape[2]-1:
def mt_assignment(input, y):
return input[index[edges[y]:edges[y+1],:]]
index = ne.evaluate("fmap==x")
img_stacked[index] = img[:, :, x][index]
index = ne.evaluate("(fmap > x) & (fmap < x+1)")
with ThreadPoolExecutor(max_workers=16) as pool:
A = np.concatenate([(pool.submit(mt_assignment, fmap, y)).result() for y in range(16)], axis=0)
B = np.concatenate([(pool.submit(mt_assignment, img[:, :, x+1], y)).result() for y in range(16)], axis=0)
C = np.concatenate([(pool.submit(mt_assignment, img[:, :, x], y)).result() for y in range(16)], axis=0)
print('A Shape is : ', A.shape)
print('A content is: ', A)
img_stacked[index] = ne.evaluate("(A-x) * B + (x+1-A) * C")
else:
last_ind = img.shape[2]-1
indexl = ne.evaluate("fmap == last_ind")
with shmem.MapReduce(np=img.shape[2]) as pool:
pool.map(do_work, range(img.shape[2]))
num_proc = shmem.cpu_count()
edges = get_edges(img, num_proc)
def mp_assignment(x):
img_stacked[edges[x]:edges[x+1],:][indexl[edges[x]:edges[x+1],:]] = img[edges[x]:edges[x+1], :, -1]\
[indexl[edges[x]:edges[x+1], :]]
with shmem.MapReduce(np=num_proc) as pool:
pool.map(mp_assignment, range(num_proc))
return img_stacked
def cache(self, grids, ncpus=1, Ns=None): # pragma: no cover
# get parameter space
if isinstance(grids[0], SliceType):
params = [list(np.arange(g.start, g.stop, g.step)) for g in grids]
else:
params = [list(np.linspace(g[0], g[1], Ns)) for g in grids]
# make combos
combos = [c for c in itertools.product(*params)]
def mini_predictor(combo):
print(combo)
return self.generate_ballpark_prediction(*combo)
# compute predictions
num_cpus = sharedmem.cpu_count() - 1
with sharedmem.Pool(np=num_cpus) as pool:
models = pool.map(self.mini_predictor, combos)
# turn into array
return models
hubble = header.hubble
chunksize = args.chunksize
#ar.ellipsoid.ne.set_num_threads(args.nthreads)
mass = [args.minmass, args.maxmass]
if args.test:
mass = [10, 10.5]
chunksize = 5
binwidth = args.binwidth
print 'Directory: ', fdir
print 'Snapshot: ', snap
print 'Boxsize: ', boxsize
print 'Hubble parameter: ', hubble
print 'Halo mass range for calculation: 10^' + str(mass), 'M_sun'
print 'Number of cores for sharedmem: ', sharedmem.cpu_count()
print 'Number of threads for numexpr: ', ar.ellipsoid.numba.config.NUMBA_NUM_THREADS
print ' '
#nbins=18; a=ar.AxialRatio(fdir,snap,nbins,rmin=10**-2.42857143,useSubhaloes=args.subhaloes)
nbins = 2
a = ar.AxialRatio(fdir,
snap,
0, [1., 2.],
useStellarhalfmassRad=True,
useSubhaloID=True,
binwidth=binwidth)
submass = a.cat.SubhaloMassInRadType[:, 4] / hubble * 1e10
subhalos = np.nonzero((submass > 10**mass[0]) & (submass < 10**mass[1]))[0]
nsubs = a.cat.nsubs
ngroups = a.cat.ngroups
def computePolicies(mdp, goal_states, error=1e-10):
"""Compute Q using multi-process
"""
# initialization of variables
n_goal_states = len(goal_states)
n_actions, n_states = mdp.n_actions, mdp.n_states
roadmap = mdp.roadmap
states = mdp.states
gamma = mdp.gamma
T = mdp.T
rewards = mdp.get_rewards()
#from IPython import embed; embed(); sys.exit()
if rewards is None: rewards=np.zeros(len(mdp.states))
else:
rewards = np.array(rewards)
rewards[np.where(rewards>0)]=0.
goal_state_ids = np.arange(n_goal_states, dtype='i')
goal_policy_mat = sharedmem.full((n_goal_states, mdp.n_states, mdp.n_actions), 0.)
## goal_values = sharedmem.full((n_goal_states, mdp.n_states), 0.)
## goal_validity = sharedmem.full((n_goal_states), True)
# Start multi processing over goal states
with sharedmem.MapReduce() as pool:
if n_goal_states % sharedmem.cpu_count() == 0:
chunksize = n_goal_states / sharedmem.cpu_count()
else:
chunksize = n_goal_states / sharedmem.cpu_count() + 1
def work(i):
state_ids = goal_state_ids[slice (i, i + chunksize)] #0,1,2,3
new_rewards = copy.copy(rewards)
values = np.zeros(n_states)
for j, goal_state_id in enumerate(state_ids):
s = goal_states[goal_state_id] # state id in states
## s_idx = states.index(s)
# vi agent
mdp = vi.valueIterAgent(n_actions, n_states,
roadmap, None, states,
gamma=gamma, T=T)
mdp.set_goal(s)
if new_rewards[s] >= 0.:
new_rewards[s] = 1.
mdp.set_rewards(new_rewards)
# Store q_mat and validity mat per state
policy, _ = mdp.find_policy(error)
goal_policy_mat[goal_state_id] = policy
# find all states that gives f_g in states
for k, gs in enumerate(goal_states):
if s == gs:
goal_policy_mat[k] = policy
# reset
new_rewards[s] = 0.
pool.map(work, range(0, n_goal_states, chunksize))#, reduce=reduce)
# convert sharedmem array to list
policies = []
for i in range(n_goal_states):
policies.append( np.array(goal_policy_mat[i]) )
return policies
def test_parallel_fit_manual_grids():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0, 360, 45)
num_blank_steps = 0
num_bar_steps = 30
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.10
pixels_down = 100
pixels_across = 100
dtype = ctypes.c_int16
voxel_index = (1, 2, 3)
auto_fit = True
verbose = 1
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
# initialize the gaussian model
model = og.GaussianModel(stimulus, utils.double_gamma_hrf)
model.hrf_delay = 0
# generate a random pRF estimate
x = -5.24
y = 2.58
sigma = 1.24
beta = 2.5
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, beta, baseline)
# set search grid
x_grid = slice(-5, 4, 5)
y_grid = slice(-5, 7, 5)
s_grid = slice(1 / stimulus.ppd, 5.25, 5)
b_grid = slice(0.1, 4.0, 5)
# set search bounds
x_bound = (-12.0, 12.0)
y_bound = (-12.0, 12.0)
s_bound = (1 / stimulus.ppd, 12.0)
b_bound = (1e-8, 1e2)
m_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (
x_grid,
y_grid,
s_grid,
)
bounds = (x_bound, y_bound, s_bound, b_bound, m_bound)
# make 3 voxels
all_data = np.array([data, data, data])
num_voxels = data.shape[0]
indices = [(1, 2, 3)] * 3
# bundle the voxels
bundle = utils.multiprocess_bundle(og.GaussianFit, model, all_data, grids,
bounds, indices)
# run analysis
with sharedmem.Pool(np=sharedmem.cpu_count() - 1) as pool:
output = pool.map(utils.parallel_fit, bundle)
# assert equivalence
for fit in output:
npt.assert_almost_equal(fit.x, x, 2)
npt.assert_almost_equal(fit.y, y, 2)
npt.assert_almost_equal(fit.sigma, sigma, 2)
npt.assert_almost_equal(fit.beta, beta, 2)
npt.assert_almost_equal(fit.baseline, baseline, 2)
def computeQ(mdp, support_states, error=1e-10, support_features=None, support_feature_state_dict=None,
cstr_fn=None, add_no_cstr=True, max_cnt=100, **kwargs):
"""Compute Q using multi-process
"""
# initialization of variables
n_support_states = len(support_states)
n_actions, n_states = mdp.n_actions, mdp.n_states
eps = np.finfo(float).eps
roadmap = mdp.roadmap
states = mdp.states
gamma = mdp.gamma
T = mdp.T
rewards = mdp.get_rewards()
#from IPython import embed; embed(); sys.exit()
if rewards is None: rewards=np.zeros(len(mdp.states))
else:
rewards = np.array(rewards)
rewards[np.where(rewards>0)]=0.
support_state_ids = np.arange(n_support_states, dtype='i')
if support_features is not None:
support_feature_ids, support_feature_values = support_features
computed_f_id = sharedmem.full(len(support_feature_values), False, dtype='b')
else:
return NotImplementedError
if cstr_fn is None:
support_q_mat = sharedmem.full((n_support_states, mdp.n_states, mdp.n_actions), 0.)
support_values = sharedmem.full((n_support_states, mdp.n_states), 0.)
support_validity = sharedmem.full((n_support_states), True)
else:
if add_no_cstr: n_cstr_fn = len(cstr_fn) + 1
else: n_cstr_fn = len(cstr_fn)
support_q_mat = sharedmem.full((n_support_states, n_cstr_fn, mdp.n_states,
mdp.n_actions), 0.)
support_values = sharedmem.full((n_support_states, n_cstr_fn, mdp.n_states), 0.)
support_validity = sharedmem.full((n_support_states, n_cstr_fn), True)
if len(cstr_fn)>0:
feat_map = kwargs['feat_map']
roadmap = kwargs['roadmap']
states = mdp.states
cstr_T = []
for i in range(len(cstr_fn)):
validity_map = cstr_fn[i](None, f=feat_map)[roadmap]
validity_map[:,0] = True
Tc = mdp.T*validity_map[:,np.newaxis,:]
Tc[:,:,0] = eps
sum_T = np.sum(Tc, axis=-1)
Tc /= sum_T[:,:,np.newaxis]
cstr_T.append(Tc)
# Start multi processing over support states
with sharedmem.MapReduce() as pool:
if n_support_states % sharedmem.cpu_count() == 0:
chunksize = n_support_states / sharedmem.cpu_count()
else:
chunksize = n_support_states / sharedmem.cpu_count() + 1
def work(i):
state_ids = support_state_ids[slice (i, i + chunksize)]
new_rewards = copy.copy(rewards)
values = np.zeros(n_states)
for j, state_id in enumerate(state_ids):
s = support_states[state_id] # state id in states
# vi agent
mdp = vi.valueIterAgent(n_actions, n_states,
roadmap, None, states,
gamma=gamma, T=T)
mdp.set_goal(s)
if support_feature_ids is None:
if new_rewards[s] >= 0.:
new_rewards[s] = 1.
else:
# find all states that gives f_g in states
f_id = support_feature_ids[state_id]
goal_state_ids = support_feature_state_dict[f_id]
if computed_f_id[f_id]:
continue
else:
computed_f_id[f_id] = True
new_rewards[goal_state_ids] = 1.
mdp.set_rewards(new_rewards)
# Store q_mat and validity mat per state
if cstr_fn is not None:
for k in range(len(cstr_fn)):
# check if the goal is isolated
if np.sum(cstr_T[k][goal_state_ids])>0.:
values, param_dict = mdp.solve_mdp(error, init_values=values,
T=cstr_T[k], max_cnt=max_cnt,
goal=s,
return_params=True)
support_q_mat[state_id][k] = param_dict['q']
support_validity[state_id][k] = cstr_fn[k](s)
support_values[state_id][k] = values
if add_no_cstr:
values, param_dict = mdp.solve_mdp(error, init_values=values,
T=T, max_cnt=max_cnt,
## goal=s,
return_params=True)
support_q_mat[state_id][-1] = param_dict['q']
support_validity[state_id][-1] = True
support_values[state_id][-1] = values
else:
values, param_dict = mdp.solve_mdp(error, init_values=values,
max_cnt=max_cnt,
return_params=True)
support_q_mat[state_id] = param_dict['q']
support_values[state_id] = values
# find all states that gives f_g in states
for gs in goal_state_ids:
k = support_states.index(gs)
if k!=state_id:
support_q_mat[k] = support_q_mat[state_id]
# reset
## new_rewards = copy.copy(rewards)
if support_feature_ids is None:
new_rewards[s] = 0.
else:
new_rewards[goal_state_ids] = 0.
pool.map(work, range(0, n_support_states, chunksize))#, reduce=reduce)
# convert sharedmem array to dict
support_q_mat_dict = {}
support_values_dict = {}
support_validity_dict = {}
for i, s in enumerate(support_states):
support_q_mat_dict[s] = np.array(support_q_mat[i])
support_values_dict[s] = np.array(support_values[i])
if cstr_fn is not None:
support_validity_dict[s] = np.array(support_validity[i])
if cstr_fn is not None:
return support_q_mat_dict, support_values_dict, support_validity_dict
else:
return support_q_mat_dict, support_values_dict
def main(A):
# gaussian are used for each subbox.
# slaves are processes so they won't damage these variables
# from the master.
global sightlines
global deltafield
global objectidfield
global velfield
sightlines = Sightlines(A)
powerspec = PowerSpectrum(A)
# fine1 takes some time, so we do it async
# while initing lya and estimating box layout.
delta0, var0, disp0 = initcoarse(A, powerspec)
varlya = initlya(A)
den1 = density2.Density(A.NmeshFine,
# Kmax=A.Kmax,
Kmin=A.KSplit, power=powerspec,
BoxSize=A.BoxSize / A.Nrep)
layout = A.layout(len(sightlines), chunksize=1024)
Nsamples = sightlines.Nsamples.sum()
print 'total number of pixels', Nsamples
deltafield = sharedmem.empty(shape=Nsamples, dtype='f4')
velfield = sharedmem.empty(shape=Nsamples, dtype='f4')
objectidfield = sharedmem.empty(shape=Nsamples, dtype='i4')
processors = [
(AddDelta, delta0),
(AddDisp(0), disp0[0]),
(AddDisp(1), disp0[1]),
(AddDisp(2), disp0[2]),
]
for proc, d0 in processors:
proc.prepare(A, d0)
MemoryBytes = numpy.max([proc.MemoryBytes for proc, d0 in processors])
np = int((sharedmem.total_memory() - 1024 ** 3) // MemoryBytes)
np = numpy.min([sharedmem.cpu_count(), np])
print 'spawn and work, with ', np, 'slaves', \
'each use', MemoryBytes /1024.**2, 'MB'
var1list = []
with sharedmem.Pool(np=np) as pool:
def work(i, j, k):
box = layout[i, j, k]
var = None
for cls, d0 in processors:
proc = cls(box, den1, varlya)
N = 0
for chunk in box:
N += proc.visit(chunk)
if cls is AddDelta:
var1 = proc.var1
# free memory
del proc
if N == 0:
# no pixels
# No need to work on other processors
break
print 'done', i, j, k, N, var1
return var1
def reduce(v1):
if v1 is not None:
var1list.append(v1)
pool.map(work, A.yieldwork(), reduce=reduce, star=True)
deltafield.tofile(A.DeltaField)
velfield.tofile(A.VelField)
objectidfield.tofile(A.ObjectIDField)
D2 = A.cosmology.Dplus(1 / 3.0) / A.cosmology.Dplus(1.0)
D3 = A.cosmology.Dplus(1 / 4.0) / A.cosmology.Dplus(1.0)
var1 = numpy.nanmean(var1list)
var = var0 + var1 + varlya
print 'gaussian-variance is', var
numpy.savetxt(A.datadir + '/gaussian-variance.txt', [var])
print 'lya field', 'var', var
print 'growth factor at z=2.0, 3.0', D2, D3
print 'lya variance adjusted to z=2.0, z=3.0', D2 ** 2 * var, D3 **2 * var
def test_parallel_fit_manual_grids():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0,360,45)
num_blank_steps = 0
num_bar_steps = 30
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.10
pixels_down = 100
pixels_across = 100
dtype = ctypes.c_int16
voxel_index = (1,2,3)
auto_fit = True
verbose = 1
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps, num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width, scale_factor, tr_length, dtype)
# initialize the gaussian model
model = og.GaussianModel(stimulus, utils.double_gamma_hrf)
model.hrf_delay = 0
# generate a random pRF estimate
x = -5.24
y = 2.58
sigma = 1.24
beta = 2.5
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, beta, baseline)
# set search grid
x_grid = slice(-5,4,5)
y_grid = slice(-5,7,5)
s_grid = slice(1/stimulus.ppd,5.25,5)
b_grid = slice(0.1,4.0,5)
# set search bounds
x_bound = (-12.0,12.0)
y_bound = (-12.0,12.0)
s_bound = (1/stimulus.ppd,12.0)
b_bound = (1e-8,1e2)
m_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (x_grid, y_grid, s_grid,)
bounds = (x_bound, y_bound, s_bound, b_bound, m_bound)
# make 3 voxels
all_data = np.array([data,data,data])
num_voxels = data.shape[0]
indices = [(1,2,3)]*3
# bundle the voxels
bundle = utils.multiprocess_bundle(og.GaussianFit, model, all_data, grids, bounds, indices)
# run analysis
with sharedmem.Pool(np=sharedmem.cpu_count()-1) as pool:
output = pool.map(utils.parallel_fit, bundle)
# assert equivalence
for fit in output:
npt.assert_almost_equal(fit.x, x, 2)
npt.assert_almost_equal(fit.y, y, 2)
npt.assert_almost_equal(fit.sigma, sigma, 2)
npt.assert_almost_equal(fit.beta, beta, 2)
npt.assert_almost_equal(fit.baseline, baseline, 2)
def paint(pos, color, luminosity, sml, camera, CCD, tree=None,
return_tree_and_sml=False, normalize=True, np=None, direct_write=False,
cumulative=True):
""" pos = (x, y, z)
color can be None or 1 or array
luminosity can be None or 1 or array
sml can be None or 1, or array
camera is Camera
CCD is array of camera.shape, 2.
CCD[..., 0] is the color channel (sum of color * luminosity)
CCD[..., 1] is the luminosity channel (sum of luminosity)
if color is None, CCD.shape == camera.shape
if color is not None,
CCD.shape == camera.shape, 2
CCD[..., 0] is color
CCD[..., 1] is brightness
if normalize is False, do not do CCD[..., 0] will be
the weighted sum of color.
if normalize is True, CCD[..., 0] will be the weighted average of color
if direct_write is true, each process will directly write to CCD (CCD
must be on sharedmem)
if cumulative is False, original content in CCD will be disregarded.
if cumulative is True, original content in CCD will be preserved (+=)
"""
CCDlimit = 20 * 1024 * 1024 # 20M pixel per small CCD
camera.shape = (CCD.shape[0], CCD.shape[1])
nCCD = int((CCD.shape[0] * CCD.shape[1] / CCDlimit) ** 0.5)
if np is None: np = sharedmem.cpu_count()
if nCCD <= np ** 0.5:
nCCD = int(np ** 0.5 + 1)
cams = camera.divide(nCCD, nCCD)
cams = cams.reshape(-1, 3)
if tree is None:
scale = fc.scale([x.min() for x in pos], [x.ptp() for x in pos])
zkey = sharedmem.empty(len(pos[0]), dtype=fc.fckeytype)
with sharedmem.MapReduce(np=np) as pool:
chunksize = 1024 * 1024
def work(i):
sl = slice(i, i+chunksize)
x, y, z = pos
fc.encode(x[sl], y[sl], z[sl], scale=scale, out=zkey[i:i+chunksize])
pool.map(work, range(0, len(zkey), chunksize))
arg = sharedmem.argsort(zkey)
tree = zt.Tree(zkey=zkey, scale=scale, arg=arg, minthresh=8, maxthresh=20)
if sml is None:
sml = sharedmem.empty(len(zkey), 'f4')
with sharedmem.MapReduce(np=np) as pool:
chunksize = 1024 * 64
def work(i):
setupsml(tree, [x[i:i+chunksize] for x in pos],
out=sml[i:i+chunksize])
pool.map(work, range(0, len(zkey), chunksize))
def writeCCD(i, sparse, myCCD):
cam, ox, oy = cams[i]
#print i, sparse, len(cams)
if sparse:
index, C, L = myCCD
x = index[0] + ox
y = index[1] + oy
p = CCD.flat
if color is not None:
ind = numpy.ravel_multi_index((x, y, 0), CCD.shape)
if cumulative:
p[ind] += C
else:
p[ind] = C
ind = numpy.ravel_multi_index((x, y, 1), CCD.shape)
if cumulative:
p[ind] += L
else:
p[ind] = L
else:
ind = numpy.ravel_multi_index((x, y), CCD.shape)
if cumulative:
p[ind] += L
else:
p[ind] = L
else:
if color is not None:
if cumulative:
CCD[ox:ox + cam.shape[0], oy:oy+cam.shape[1], :] += myCCD
else:
CCD[ox:ox + cam.shape[0], oy:oy+cam.shape[1], :] = myCCD
else:
if cumulative:
CCD[ox:ox + cam.shape[0], oy:oy+cam.shape[1]] += myCCD[..., 1]
else:
CCD[ox:ox + cam.shape[0], oy:oy+cam.shape[1]] = myCCD[..., 1]
with sharedmem.MapReduce(np=np) as pool:
def work(i):
cam, ox, oy = cams[i]
myCCD = numpy.zeros(cam.shape, dtype=('f8', 2))
cam.paint(pos[0], pos[1], pos[2],
sml, color, luminosity, out=myCCD, tree=tree)
mask = (myCCD[..., 1] != 0)
if mask.sum() < 0.1 * myCCD[..., 1].size:
index = mask.nonzero()
C = myCCD[..., 0][mask]
L = myCCD[..., 1][mask]
sparse, myCCD = True, (index, C, L)
else:
sparse, myCCD = False, myCCD
if not direct_write:
return i, sparse, myCCD
else:
writeCCD(i, sparse, myCCD)
return 0, 0, 0
def reduce(i, sparse, myCCD):
if not direct_write:
writeCCD(i, sparse, myCCD)
pool.map(work, range(len(cams)), reduce=reduce)
if color is not None and normalize:
CCD[..., 0] /= CCD[..., 1]
if return_tree_and_sml:
return CCD, tree, sml
else:
tree = None
sml = None
return CCD
def paint(pos,
color,
luminosity,
sml,
camera,
CCD,
tree=None,
return_tree_and_sml=False,
normalize=True,
np=None,
direct_write=False,
cumulative=True):
""" pos = (x, y, z)
color can be None or 1 or array
luminosity can be None or 1 or array
sml can be None or 1, or array
camera is Camera
CCD is array of camera.shape, 2.
CCD[..., 0] is the color channel (sum of color * luminosity)
CCD[..., 1] is the luminosity channel (sum of luminosity)
if color is None, CCD.shape == camera.shape
if color is not None,
CCD.shape == camera.shape, 2
CCD[..., 0] is color
CCD[..., 1] is brightness
if normalize is False, do not do CCD[..., 0] will be
the weighted sum of color.
if normalize is True, CCD[..., 0] will be the weighted average of color
if direct_write is true, each process will directly write to CCD (CCD
must be on sharedmem)
if cumulative is False, original content in CCD will be disregarded.
if cumulative is True, original content in CCD will be preserved (+=)
"""
CCDlimit = 20 * 1024 * 1024 # 20M pixel per small CCD
camera.shape = (CCD.shape[0], CCD.shape[1])
nCCD = int((CCD.shape[0] * CCD.shape[1] / CCDlimit)**0.5)
if np is None: np = sharedmem.cpu_count()
if nCCD <= np**0.5:
nCCD = int(np**0.5 + 1)
cams = camera.divide(nCCD, nCCD)
cams = cams.reshape(-1, 3)
if tree is None:
scale = fc.scale([x.min() for x in pos], [x.ptp() for x in pos])
zkey = sharedmem.empty(len(pos[0]), dtype=fc.fckeytype)
with sharedmem.MapReduce(np=np) as pool:
chunksize = 1024 * 1024
def work(i):
sl = slice(i, i + chunksize)
x, y, z = pos
fc.encode(x[sl],
y[sl],
z[sl],
scale=scale,
out=zkey[i:i + chunksize])
pool.map(work, range(0, len(zkey), chunksize))
arg = numpy.argsort(zkey)
tree = zt.Tree(zkey=zkey,
scale=scale,
arg=arg,
minthresh=8,
maxthresh=20)
if sml is None:
sml = sharedmem.empty(len(zkey), 'f4')
with sharedmem.MapReduce(np=np) as pool:
chunksize = 1024 * 64
def work(i):
setupsml(tree, [x[i:i + chunksize] for x in pos],
out=sml[i:i + chunksize])
pool.map(work, range(0, len(zkey), chunksize))
def writeCCD(i, sparse, myCCD):
cam, ox, oy = cams[i]
#print i, sparse, len(cams)
if sparse:
index, C, L = myCCD
x = index[0] + ox
y = index[1] + oy
p = CCD.flat
if color is not None:
ind = numpy.ravel_multi_index((x, y, 0), CCD.shape)
if cumulative:
p[ind] += C
else:
p[ind] = C
ind = numpy.ravel_multi_index((x, y, 1), CCD.shape)
if cumulative:
p[ind] += L
else:
p[ind] = L
else:
ind = numpy.ravel_multi_index((x, y), CCD.shape)
if cumulative:
p[ind] += L
else:
p[ind] = L
else:
if color is not None:
if cumulative:
CCD[ox:ox + cam.shape[0], oy:oy + cam.shape[1], :] += myCCD
else:
CCD[ox:ox + cam.shape[0], oy:oy + cam.shape[1], :] = myCCD
else:
if cumulative:
CCD[ox:ox + cam.shape[0],
oy:oy + cam.shape[1]] += myCCD[..., 1]
else:
CCD[ox:ox + cam.shape[0],
oy:oy + cam.shape[1]] = myCCD[..., 1]
with sharedmem.MapReduce(np=np) as pool:
def work(i):
cam, ox, oy = cams[i]
myCCD = numpy.zeros(cam.shape, dtype=('f8', 2))
cam.paint(pos[0],
pos[1],
pos[2],
sml,
color,
luminosity,
out=myCCD,
tree=tree)
mask = (myCCD[..., 1] != 0)
if mask.sum() < 0.1 * myCCD[..., 1].size:
index = mask.nonzero()
C = myCCD[..., 0][mask]
L = myCCD[..., 1][mask]
sparse, myCCD = True, (index, C, L)
else:
sparse, myCCD = False, myCCD
if not direct_write:
return i, sparse, myCCD
else:
writeCCD(i, sparse, myCCD)
return 0, 0, 0
def reduce(i, sparse, myCCD):
if not direct_write:
writeCCD(i, sparse, myCCD)
pool.map(work, range(len(cams)), reduce=reduce)
if color is not None and normalize:
CCD[..., 0] /= CCD[..., 1]
if return_tree_and_sml:
return CCD, tree, sml
else:
tree = None
sml = None
return CCD