def run(lam,gam,tau,ITERS,attack_years = [], savetail='', pmitail = ''):
r = 50 # rank
b = nw # batch size
emph = 1 # emphasize the nonzero
savefile = savehead+'L'+str(lam)+'T'+str(tau)+'G'+str(gam)+'A'+str(emph)
if len(savetail) > 0: savetail = '_' + savetail
if len(pmitail) > 0: pmitail = '_' + pmitail
savefile = savefile + savetail
embfilename = 'data/emb_static%s' % savetail
print savefile, embfilename
try:
e = sio.loadmat(embfilename)['emb']
except(IOError):
print 'file not available yet', embfilename
return
print('starting training with following parameters')
print_params(r,lam,tau,gam,emph,ITERS)
print('there are a total of {} words, and {} time points'.format(nw,T))
print('X*X*X*X*X*X*X*X*X')
print('initializing')
Ulist = [copy.deepcopy(e) for x in T]
Vlist = [copy.deepcopy(e) for x in T]
del e
print('getting batch indices')
if b < nw:
b_ind = util.getbatches(nw,b)
else:
b_ind = [range(nw)]
import time
start_time = time.time()
# sequential updates
for iteration in xrange(ITERS):
print_params(r,lam,tau,gam,emph,ITERS)
try:
Ulist = pickle.load(open( "%sngU_iter%d.p" % (savefile,iteration), "rb" ) )
Vlist = pickle.load(open( "%sngV_iter%d.p" % (savefile, iteration), "rb" ) )
print 'iteration %d loaded succesfully' % iteration
continue
except(IOError):
pass
# shuffle times
if iteration == 0: times = T
else: times = np.random.permutation(T)
for t in xrange(len(times)): # select a time
print 'iteration %d, time %d' % (iteration, t)
if T[t] in attack_years:
f = 'data/wordPMI_%d%s.mat' % (T[t],pmitail)
else:
f = 'data/wordPMI_' + str(T[t]) + '.mat'
pmi = sio.loadmat(f)['pmi']
for j in xrange(len(b_ind)): # select a mini batch
print '%d out of %d' % (j,len(b_ind))
ind = b_ind[j]
## UPDATE V
# get data
pmi_seg = pmi[:,ind].todense()
if t==0:
vp = np.zeros((len(ind),r))
up = np.zeros((len(ind),r))
iflag = True
else:
vp = Vlist[t-1][ind,:]
up = Ulist[t-1][ind,:]
iflag = False
if t==len(T)-1:
vn = np.zeros((len(ind),r))
un = np.zeros((len(ind),r))
iflag = True
else:
vn = Vlist[t+1][ind,:]
un = Ulist[t+1][ind,:]
iflag = False
Vlist[t][ind,:] = util.update(Ulist[t],emph*pmi_seg,vp,vn,lam,tau,gam,ind,iflag)
Ulist[t][ind,:] = util.update(Vlist[t],emph*pmi_seg,up,un,lam,tau,gam,ind,iflag)
#### INNER BATCH LOOP END
# save
print 'time elapsed = ', time.time()-start_time
pickle.dump(Ulist, open( "%sngU_iter%d.p" % (savefile,iteration), "wb" ) , pickle.HIGHEST_PROTOCOL)
pickle.dump(Vlist, open( "%sngV_iter%d.p" % (savefile, iteration), "wb" ) , pickle.HIGHEST_PROTOCOL)
def run(lam, gam, tau, ITERS, attack_years=[], savetail="", pmitail=""):
r = 50 # rank
b = nw # batch size
emph = 1 # emphasize the nonzero
savefile = (savehead + "L" + str(lam) + "T" + str(tau) + "G" + str(gam) +
"A" + str(emph))
if len(savetail) > 0:
savetail = "_" + savetail
if len(pmitail) > 0:
pmitail = "_" + pmitail
savefile = savefile + savetail
embfilename = "data/emb_static%s.mat" % savetail
print(savefile, embfilename)
try:
e = sio.loadmat(embfilename)["emb"]
except (IOError):
print("file not available yet", embfilename)
return
print("starting training with following parameters")
print_params(r, lam, tau, gam, emph, ITERS)
print("there are a total of {} words, and {} time points".format(nw, T))
print("X*X*X*X*X*X*X*X*X")
print("initializing")
Ulist = [copy.deepcopy(e) for x in T]
Vlist = [copy.deepcopy(e) for x in T]
del e
print("getting batch indices")
if b < nw:
b_ind = util.getbatches(nw, b)
else:
b_ind = [range(nw)]
import time
start_time = time.time()
# sequential updates
for iteration in range(ITERS):
print_params(r, lam, tau, gam, emph, ITERS)
try:
Ulist = pickle.load(
open("%sngU_iter%d.p" % (savefile, iteration), "rb"))
Vlist = pickle.load(
open("%sngV_iter%d.p" % (savefile, iteration), "rb"))
print("iteration %d loaded succesfully" % iteration)
continue
except (IOError):
pass
# shuffle times
if iteration == 0:
times = T
else:
times = np.random.permutation(T)
for t in range(len(times)): # select a time
print("iteration %d, time %d" % (iteration, t))
if T[t] in attack_years:
# file = "data/wordPMI_%d%s.mat" % (T[t], pmitail)
file = "data/emb_{}{}.mat".format(T[t], pmitail)
else:
# file = "data/wordPMI_" + str(T[t]) + ".mat"
file = "data/emb_{}.mat".format(T[t])
pmi = sio.loadmat(file)["pmi"]
for j in range(len(b_ind)): # select a mini batch
print("%d out of %d" % (j, len(b_ind)))
ind = b_ind[j]
## UPDATE V
# get data
pmi_seg = pmi[:, ind].todense()
if t == 0:
vp = np.zeros((len(ind), r))
up = np.zeros((len(ind), r))
iflag = True
else:
vp = Vlist[t - 1][ind, :]
up = Ulist[t - 1][ind, :]
iflag = False
if t == len(T) - 1:
vn = np.zeros((len(ind), r))
un = np.zeros((len(ind), r))
iflag = True
else:
vn = Vlist[t + 1][ind, :]
un = Ulist[t + 1][ind, :]
iflag = False
Vlist[t][ind, :] = util.update(Ulist[t], emph * pmi_seg, vp,
vn, lam, tau, gam, ind, iflag)
Ulist[t][ind, :] = util.update(Vlist[t], emph * pmi_seg, up,
un, lam, tau, gam, ind, iflag)
#### INNER BATCH LOOP END
# save
print("time elapsed = ", time.time() - start_time)
pickle.dump(
Ulist,
open("%sngU_iter%d.p" % (savefile, iteration), "wb"),
pickle.HIGHEST_PROTOCOL,
)
pickle.dump(
Vlist,
open("%sngV_iter%d.p" % (savefile, iteration), "wb"),
pickle.HIGHEST_PROTOCOL,
)
print('starting training with following parameters')
print_params(r, lam, tau, gam, emph, ITERS)
print('there are a total of {} words, and {} time points'.format(nw, T))
print('X*X*X*X*X*X*X*X*X')
print('initializing')
#Ulist,Vlist = util.initvars(nw,T,r, trainhead)
Ulist, Vlist = util.import_static_init(T)
print(Ulist)
print(Vlist)
#asdf
print('getting batch indices')
if b < nw:
b_ind = util.getbatches(nw, b)
else:
b_ind = [list(range(nw))]
import time
start_time = time.time()
# sequential updates
for iteration in range(ITERS):
print_params(r, lam, tau, gam, emph, ITERS)
try:
Ulist = pickle.load(
open("%sngU_iter%d.p" % (savefile, iteration), "rb"))
Vlist = pickle.load(
open("%sngV_iter%d.p" % (savefile, iteration), "rb"))
print('iteration %d loaded succesfully' % iteration)
continue
def run_dw2v(exper_dir, iters, lam, tau, gam, emph, r):
import DynamicWord2Vec.train_model.util_timeCD as util
data_dir = os.path.join(exper_dir, 'embs')
trainhead = os.path.join(data_dir, 'wordPairPMI_')
voc_path = os.path.join(data_dir, 'wordIDHash.csv')
embmat = os.path.join(data_dir, 'emb_static.mat')
b = nw = get_vocabulary_size(voc_path)
ITERS = iters
T = get_points(data_dir)
savehead = os.path.join(exper_dir, 'results')
savefile = 'L{}T{}G{}A{}'.format(lam, tau, gam, emph)
savefile = os.path.join(savehead, savefile)
print(f"b = {b}, nw = {nw}, T = {T}")
print(iters, lam, tau, gam, emph, r)
print('starting training with following parameters')
print_params(r, lam, tau, gam, emph, ITERS)
print('there are a total of {} words, and {} time points'.format(nw, T))
print('X*X*X*X*X*X*X*X*X')
print('initializing')
#Ulist,Vlist = util.initvars(nw,T,r, trainhead)
Ulist, Vlist = util.import_static_init(T, embmat)
print(Ulist)
print(Vlist)
print('getting batch indices')
if b < nw:
b_ind = util.getbatches(nw, b)
else:
b_ind = [list(range(nw))]
import time
start_time = time.time()
# sequential updates
for iteration in range(ITERS):
print_params(r, lam, tau, gam, emph, ITERS)
try:
Ulist = pickle.load(
open("%sngU_iter%d.p" % (savefile, iteration), "rb"))
Vlist = pickle.load(
open("%sngV_iter%d.p" % (savefile, iteration), "rb"))
print('iteration %d loaded succesfully' % iteration)
continue
except (IOError):
pass
loss = 0
# shuffle times
if iteration == 0: times = T
else: times = np.random.permutation(T)
for t in range(len(times)): # select a time
print('iteration %d, time %d' % (iteration, t))
f = trainhead + str(T[t]) + '.csv'
print(f)
"""
try:
Ulist = pickle.load( open( "%sngU_iter%d_time%d_tmp.p" % (savefile,iteration,t), "rb" ) )
Vlist = pickle.load( open( "%sngV_iter%d_time%d_tmp.p" % (savefile, iteration,t), "rb" ) )
times = pickle.load( open( "%sngtimes_iter%d_time%d_tmp.p" % (savefile, iteration,t), "rb" ) )
print 'iteration %d time %d loaded succesfully' % (iteration, t)
continue
except(IOError):
pass
"""
pmi = util.getmat(f, nw, False)
for j in range(len(b_ind)): # select a mini batch
print('%d out of %d' % (j, len(b_ind)))
ind = b_ind[j]
## UPDATE V
# get data
pmi_seg = pmi[:, ind].todense()
if t == 0:
vp = np.zeros((len(ind), r))
up = np.zeros((len(ind), r))
iflag = True
else:
vp = Vlist[t - 1][ind, :]
up = Ulist[t - 1][ind, :]
iflag = False
if t == len(T) - 1:
vn = np.zeros((len(ind), r))
un = np.zeros((len(ind), r))
iflag = True
else:
vn = Vlist[t + 1][ind, :]
un = Ulist[t + 1][ind, :]
iflag = False
Vlist[t][ind, :] = util.update(Ulist[t], emph * pmi_seg, vp,
vn, lam, tau, gam, ind, iflag)
Ulist[t][ind, :] = util.update(Vlist[t], emph * pmi_seg, up,
un, lam, tau, gam, ind, iflag)
#pickle.dump(Ulist, open( "%sngU_iter%d_time%d_tmp.p" % (savefile,iteration,t), "wb" ) , pickle.HIGHEST_PROTOCOL)
#pickle.dump(Vlist, open( "%sngV_iter%d_time%d_tmp.p" % (savefile, iteration,t), "wb" ) , pickle.HIGHEST_PROTOCOL)
#pickle.dump(times, open( "%sngtimes_iter%d_time%d_tmp.p" % (savefile, iteration,t), "wb" ) , pickle.HIGHEST_PROTOCOL)
#### INNER BATCH LOOP END
# save
print('time elapsed = ', time.time() - start_time)
pickle.dump(Ulist, open("%sngU_iter%d.p" % (savefile, iteration),
"wb"), pickle.HIGHEST_PROTOCOL)
pickle.dump(Vlist, open("%sngV_iter%d.p" % (savefile, iteration),
"wb"), pickle.HIGHEST_PROTOCOL)