def mkmitral(gid):
if ismitral(gid):
nrn = genMitral(gid)
m = h.Mitral(gid)
else:
nrn = genMTufted(gid)
m = h.mTufted(gid)
m.createsec(len(nrn.dend), len(nrn.tuft))
for i, d in enumerate(nrn.dend):
if d.parent == nrn.soma[0]:
m.secden[i].connect(m.soma(.5))
else:
index = nrn.dend.index(d.parent)
m.secden[i].connect(m.secden[index](1))
m.geometry() # again to get the hillock stylized shape
fillall(nrn, m)
m.subsets()
m.topol() # need to connect secondary dendrites explicitly
m.segments(
) # again to get the proper number of segments for tuft and secden
m.memb()
m.setup_orns(gid, params.stream_orn_w)
if ismitral(gid):
orn_g_std = params.orn_g_mc_std
orn_g_baseline = params.orn_g_mc_baseline
orn_g_max = params.orn_g_mc_max
elif ismtufted(gid):
orn_g_std = params.orn_g_mt_std
orn_g_baseline = params.orn_g_mt_baseline
orn_g_max = params.orn_g_mt_max
for i in range(int(m.ntuft)):
m.ornsyn.o(i).g_e_baseline = orn_g_baseline
m.ornsyn.o(i).std_e = orn_g_std
m.ornsyn.o(i).g_e_max = orn_g_max
if mgid2glom(gid) in params.vclamp:
m.vcinit()
return m
def _pd(mgid, isec, x):
try:
m = mitrals[mgid]
except KeyError:
if gidfunc.ismitral(mgid):
m = grow.genMitral(mgid)
elif gidfunc.ismtufted(mgid):
m = grow.genMTufted(mgid)
mitrals[mgid] = m
d = x * (len(m.dend[isec].points)-1)
sec = m.dend[isec].parent
while sec != m.soma[0]:
d += len(sec.points)-1
sec = sec.parent
return d*20
def issynapse(gid):
return not (gidfunc.ismitral(gid) or gidfunc.ismtufted(gid)
or gidfunc.isgranule(gid))
def init():
data = {}
for uid in range(nhost):
data.update({ uid:(getmodel().mitrals.keys()) })
data = a2a.all2all(data)
mgids = []
for _mgids in data.values(): mgids += _mgids
mgids = set(mgids)
# initialize source
for mgid in getmodel().mitrals.keys():
mpriden = split.mpriden(mgid)
if not mpriden:
continue
rgj = params.ranstream(mgid, params.stream_gap_junction)
mpriden.push()
secref = h.SectionRef()
h.pop_section()
h.mk_gj_src(pc, mgid, secref)
glomid = mgid2glom(mgid)
sistergids = []
# no longer all to all, only a chain
if not (ismtufted(mgid) and (mgid - nmi) % nmt == (nmt - 1)):
if ismitral(mgid) and mgid % nmxg == (nmxg - 1):
sistergids += [glomid * nmt + nmi]
else:
sistergids += [mgid + 1]
if not (ismitral(mgid) and mgid % nmxg == 0):
if ismtufted(mgid) and (mgid - nmi) % nmt == 0:
sistergids += [(glomid + 1) * nmxg - 1]
else:
sistergids += [mgid - 1]
sistergids = mgids.intersection(range(glomid * nmxg, glomid * nmxg + nmxg) + range(glomid * nmt + nmi, glomid * nmt + nmt + nmi)).difference([ mgid ])
for sistermgid in sistergids:
gap = h.Gap(mpriden(0.99))
if ismitral(mgid) and ismitral(sistermgid):
gap.g = rgj.uniform(gj_min_g1, gj_max_g1)
elif ismtufted(mgid) and ismtufted(sistermgid):
gap.g = rgj.uniform(gj_min_g3, gj_max_g3)
else:
gap.g = rgj.uniform(gj_min_g2, gj_max_g2)
getmodel().gj[(mgid, sistermgid)] = gap
pc.barrier()
# initialize targets
for key, gap in getmodel().gj.items():
mgid, sistermgid = key
pc.target_var(gap, gap._ref_vgap, sistermgid)
util.elapsed('Gap junctions built')
def multisplit_distrib(model):
enter = h.startsw()
cxlist = determine_multisplit_complexity(model)
#start over
destroy_model(model)
# fake cell to solve problem with gap junctions
# init_fake_cell()
# but we still have model.mconnections and model.rank_gconnections
# from the round-robin distribution perspective.
# we will need to tell the ranks where to distribute that information
cxlist = load_bal(
cxlist,
nhost) #cxlist is the list of (cx,(gid,piece)) we want on this process
# the new distribution of mitrals and granules
model.gids = set([item[1][0] for item in cxlist])
model.mitral_gids = set(
[gid for gid in model.gids if ismitral(gid) or ismtufted(gid)])
model.granule_gids = set([gid for gid in model.gids if isgranule(gid)])
model.blanes_gids = set([gid for gid in model.gids if isblanes(gid)])
# for splitting need gid:[pieceindices]
gid2pieces = {}
for item in cxlist:
gid = item[1][0]
piece = item[1][1]
if not gid2pieces.has_key(gid):
gid2pieces[gid] = []
gid2pieces[gid].append(piece)
# get the correct mconnections and gconnections
# Round-robin ranks that presently have the connection info for the gids
rr = {}
for gid in model.gids:
r = gid % nhost
if not rr.has_key(r):
rr[r] = []
rr[r].append(gid)
rr = all2all(rr)
# rr is now the ranks for where to send the synapse information
# all the gids in rr were 'owned' by the round-robin distribution
# may wish to revisit so that only synapse info for relevant pieces is
# scattered.
mc = model.mconnections
gc = model.rank_gconnections
# construct a ggid2connection dict
ggid2connection = {}
for r in gc:
for ci in gc[r]:
ggid = ci[3]
if not ggid2connection.has_key(ggid):
ggid2connection[ggid] = []
ggid2connection[ggid].append(ci)
for r in rr:
gids = rr[r]
mgci = []
rr[r] = mgci
for gid in gids:
if mc.has_key(gid):
mgci.append(mc[gid])
elif gidfunc.isgranule(gid):
mgci.append(ggid2connection[gid])
mgci = all2all(rr)
# mgci contains all the connection info needed by the balanced distribution
# create mitrals and granules, split and register and create synapses
nmc.dc.mk_mitrals(model) # whole cells
nmc.build_granules(model)
nmc.build_blanes(model)
for gid in gid2pieces:
if ismitral(gid) or ismtufted(gid):
split.splitmitral(gid, model.mitrals[gid], gid2pieces[gid])
pc.multisplit()
nmc.register_mitrals(model)
nmc.register_granules(model)
nmc.register_blanes(model)
# build_synapses() ... use mgci to build explicitly
model.mgrss = {}
for r in mgci:
for cil in mgci[r]:
for ci in cil:
if not model.mgrss.has_key(mgrs.mgrs_gid(ci[0], ci[3], ci[6])):
rsyn = mgrs.mk_mgrs(*ci[0:7])
if rsyn:
model.mgrss[rsyn.md_gid] = rsyn
nmultiple = int(pc.allreduce(mgrs.multiple_cnt(), 1))
# it is faster if generated again
blanes.mk_gl2b_connections()
# excitatory
for mgid, blanes_gid, w in model.mt2blanes_connections:
syn = blanes.mt2blanes(mgid, blanes_gid, w)
model.mt2blanes[syn.gid] = syn
blanes.mk_b2g_connections()
# inhibitory synapses from blanes to gc
for ggid, blanes_gid, w in model.blanes2gc_connections:
syn = blanes.blanes2granule(blanes_gid, ggid, w)
model.blanes2gc[syn.gid] = syn
if rank == 0:
print 'nmultiple = ', nmultiple
detectors = h.List("AmpaNmda")
util.elapsed('%d ampanmda for reciprocalsynapses constructed' %
int(pc.allreduce(detectors.count(), 1)))
detectors = h.List("FastInhib")
util.elapsed('%d fi for reciprocalsynapses constructed' %
int(pc.allreduce(detectors.count(), 1)))
detectors = h.List("ThreshDetect")
util.elapsed('%d ThreshDetect for reciprocalsynapses constructed' %
int(pc.allreduce(detectors.count(), 1)))
util.elapsed('%d mt to bc' % int(pc.allreduce(len(model.mt2blanes), 1)))
util.elapsed('%d bc to gc' % int(pc.allreduce(len(model.blanes2gc), 1)))
if rank == 0: print 'multisplit_distrib time ', h.startsw() - enter
def chk_gran_conn():
model = modeldata.getmodel()
ggids = model.granule_gids
mgids = model.mitral_gids
mgrss = model.mgrss
# for each granule, list of all mgid it connects to and vice versa
g2m = {}
m2g = {}
for mgrs in mgrss.values():
ggid = mgrs.ggid
mgid = mgrs.mgid
if mgrs.gd: # granule exists on this rank
if ggid not in g2m:
g2m[ggid] = []
g2m[ggid].append(mgid)
if mgrs.md: # mitral or mtufted exists on this rank
if mgid not in m2g:
m2g[mgid] = []
m2g[mgid].append(ggid)
a = 'Number of granules that go to mitral and mtufted (regardless of glomerulus)'
na = 0
for ms in g2m.values():
nmit = 0
nmtuft = 0
for mgid in ms:
nmit += 1 if gidfunc.ismitral(mgid) else 0
nmtuft += 1 if gidfunc.ismtufted(mgid) else 0
na += 1 if nmit > 0 and nmtuft > 0 else 0
na = int(pc.allreduce(na, 1))
if pc.id() == 0:
print("%d %s" % (na, a))
a = '[total,imin,imax,iavg,xmin,xmax,xavg] granule connections to mitral'
b = '[total,imin,imax,iavg,xmin,xmax,xavg] granule connections to mtuft'
nab = [[[], [], [], []], [[], [], [], []]]
for ms in g2m.values():
# assume if first is mitral then all are mitrals otherwise mtuft
dat = nab[0] if gidfunc.ismitral(ms[0]) else nab[1]
# want to distinguish intra mitral, distinct mitral, and any mitral
# generate map of distinct:count for that distinct instance
distinct = {}
for d in set(ms):
distinct[d] = 0
for m in ms:
distinct[m] += 1
n_any = len(ms)
n_distinct_m = len(distinct)
n_intra_m_min = min(distinct.values())
n_intra_m_max = max(distinct.values())
dat[0].append(n_any)
dat[1].append(n_distinct_m)
dat[2].append(n_intra_m_min)
dat[3].append(n_intra_m_max)
def nhost_len(v):
return int(pc.allreduce(len(v), 1))
def nhost_sum(v):
return int(pc.allreduce(sum(v), 1))
def nhost_max(v):
return int(pc.allreduce(max(v), 2))
def nhost_min(v):
return int(pc.allreduce(min(v), 3))
def pr0(s):
if pc.id() == 0:
print(s)
def pr(title, dat):
i = 0
pr0(title)
totg = nhost_len(dat[i])
pr0("%d granules of this type" % totg)
pr0("%d total connections" % nhost_sum(dat[i]))
if totg:
pr0("%d min ; %d max ; %g avg" % (nhost_min(
dat[i]), nhost_max(dat[i]), float(nhost_sum(dat[i])) / totg))
pr(a, nab[0])
pr(b, nab[1])
def __init__(self, mgid, isec, xm, ggid, ipri, xg, slot):
self.mgid = mgid
self.ggid = ggid
self.slot = slot
self.xm = xm
self.xg = xg
self.isec = isec
self.ipri = ipri
self.msecden = split.msecden(mgid, isec)
self.gpriden = split.gpriden(ggid, ipri)
self.md_gid = mgrs_gid(mgid, ggid)
self.gd_gid = mgrs_gid(ggid, mgid)
self.md = None #ThreshDetect on mitral
self.gd = None #ThreshDetect on granule
if params.use_fi_stdp:
self.fi = None #FastInhibSTDP on mitral
self.postspike2fi = None # negative weight NetCon from md to fi
else:
self.fi = None #FastInhib on mitral
self.ampanmda = None #AmpaNmda on granule
self.gd2fi = None #NetCon to fi
self.md2ampanmda = None #NetCon to ampanmda
if pc.gid_exists(self.md_gid) > 0. or pc.gid_exists(self.gd_gid) > 0.:
print "md_gid=%d and/or gd_gid=%d already registered" % (self.md_gid, self.gd_gid)
raise RuntimeError
if self.msecden:
self.md = h.ThreshDetect(self.msecden(xm))
if params.use_fi_stdp:
self.fi = h.FastInhibSTDP(self.msecden(xm))
nc = h.NetCon(self.md, self.fi)
self.postspike2fi = nc
nc.weight[0] = -1
nc.delay = 1
else:
self.fi = h.FastInhib(self.msecden(xm))
try:
if params.training_inh:
self.fi.training = 1
else:
self.fi.training = 0
except:
# print 'error'
self.fi.training = 1
if ismitral(mgid):
self.fi.gmax = params.mc_inh_gmax
self.fi.tau1 = params.mc_fi_tau1
self.fi.tau2 = params.mc_fi_tau2
elif ismtufted(mgid):
self.fi.gmax = params.mt_inh_gmax
self.fi.tau1 = params.mt_fi_tau1
self.fi.tau2 = params.mt_fi_tau2
pc.set_gid2node(self.md_gid, pc.id())
pc.cell(self.md_gid, h.NetCon(self.md, None), 1)
if self.gpriden:
self.spine = h.GranuleSpine()
if gcissup(self.ggid):
self.spine.sup_deep_flag(1)
else:
self.spine.sup_deep_flag(0)
self.spine.neck.connect(self.gpriden(xg))
#self.dsac = dSAC(self.ggid, self.spine)
self.gd = h.ThreshDetect(self.spine.head(0.5))
self.gd.vthresh = -50
self.ampanmda = h.AmpaNmda(self.spine.head(0.5))
if ismitral(mgid):
self.ampanmda.gmax = params.mc_exc_gmax
elif ismtufted(mgid):
self.ampanmda.gmax = params.mt_exc_gmax
try:
if params.training_exc:
self.ampanmda.training = 1
else:
self.ampanmda.training = 0
except:
# print 'error'
self.ampanmda.training = 1
pc.set_gid2node(self.gd_gid, pc.id())
pc.cell(self.gd_gid, h.NetCon(self.gd, None), 1)
# Cannot be done above because output ports must exist prior to using
# an output gid as an input port on the same process.
if self.fi:
self.gd2fi = pc.gid_connect(self.gd_gid, self.fi)
if params.use_fi_stdp:
self.gd2fi.weight[0] = 1
else:
self.gd2fi.weight[0] = 1 # normalized
try:
init_inh_weight = params.init_inh_weight
except:
init_inh_weight = 0
self.gd2fi.weight[1] = init_inh_weight
self.gd2fi.delay = 1
if self.ampanmda:
self.md2ampanmda = pc.gid_connect(self.md_gid, self.ampanmda)
self.md2ampanmda.weight[0] = 1 #normalized
try:
init_exc_weight = params.init_exc_weight
except:
init_exc_weight = 0
self.md2ampanmda.weight[1] = init_exc_weight
self.md2ampanmda.delay = 1