def test_transpose_merge(): # LOAD MPS fname = './mps.h5' mps, qnum = mps_io.loadMPS(fname) lmps = [mps, qnum] bdim = [len(x) for x in qnum] print(' bdim = ', bdim) nsite = len(mps) qphys = mpo_dmrg_qphys.initSpatialOrb(nsite, 2) for isite in range(nsite): ql = qnum[isite] qn = qphys[isite] qr = qnum[isite + 1] cl = qtensor_util.classification(ql) cn = qtensor_util.classification(qn) cr = qtensor_util.classification(qr) site = mps[isite] tsite0 = site.transpose(2, 0, 1) print('isite=', isite) tmps = qtensor.qtensor([False, False, True]) tmps.fromDenseTensor(site, [ql, qn, qr]) tmps = tmps.transpose(2, 0, 1) tsite = tmps.toDenseTensor() diff1 = numpy.linalg.norm(tsite0 - tsite) print(' diff1=', diff1) shape = tsite0.shape tmp = tsite0.reshape((shape[0], shape[1] * shape[2])) tmps = tmps.merge([[0], [1, 2]]) tmat = tmps.toDenseTensor() diff2 = numpy.linalg.norm(tmp - tmat) print(' diff2=', diff2) return 0
def test_io(): # LOAD MPS fname = './mps.h5' mps, qnum = mps_io.loadMPS(fname) lmps = [mps, qnum] bdim = [len(x) for x in qnum] print(' bdim = ', bdim) for item in mps: print(item.shape) for item in qnum: print(item) nsite = len(mps) print(nsite) qphys = mpo_dmrg_qphys.initSpatialOrb(nsite, 2) print(qphys) print(len(qnum)) f1 = h5py.File("mpsQt.h5", "w") for isite in range(nsite): ql = qnum[isite] qn = qphys[isite] qr = qnum[isite + 1] site = mps[isite] print() print('isite=', isite) tmps = qtensor.qtensor([False, False, True]) tmps.fromDenseTensor(site, [ql, qn, qr]) tmps.dump(f1, 'site' + str(isite)) tmps2 = qtensor.qtensor() tmps2.load(f1, 'site' + str(isite)) tmps2.prt() f1.close() return 0
def test_creann(): # LOAD MPS fname = './mps.h5' mps, qnum = mps_io.loadMPS(fname) lmps = [mps, qnum] bdim = [len(x) for x in qnum] print(' bdim = ', bdim) nsite = len(mps) qphys = mpo_dmrg_qphys.initSpatialOrb(nsite, 2) ta = 0. tb = 0. for isite in range(nsite): ql = qnum[isite] qn = qphys[isite] qr = qnum[isite + 1] cl = qtensor_util.classification(ql) cn = qtensor_util.classification(qn) cr = qtensor_util.classification(qr) print('isite/nsite=', isite, nsite) site = mps[isite] tmps = qtensor.qtensor([False, False, True]) tmps.fromDenseTensor(site, [ql, qn, qr]) tmps.prt() for iop in [1, 0]: for p in range(2 * nsite): t0 = time.time() op = mpo_dmrg_opers.genElemSpatialMat(p, isite, iop) #csite = numpy.einsum('ij,ajb->aib',op,site) csite = numpy.tensordot(op, site, axes=([1], [1])) # iab csite = csite.transpose(1, 0, 2) t1 = time.time() qop = qtensor_opers.genElemSpatialQt(p, isite, iop) # ijab,xby-> ijaxy -> ix,a,jy tmps2 = qtensor.tensordot(qop, tmps, axes=([3], [1])) tmps2 = tmps2.transpose(0, 3, 2, 1, 4) tmps2 = tmps2.merge([[0, 1], [2], [3, 4]]) tsite = tmps2.toDenseTensor() t2 = time.time() assert csite.shape == tsite.shape diff = numpy.linalg.norm(tsite - csite) print('iop,p,diff=', iop, p, csite.shape, diff, ' t0=', t1 - t0, ' t1=', t2 - t1) assert diff < 1.e-10 ta += t1 - t0 tb += t2 - t1 # In case of large bond dimension, e.g., # D=2000, t0/t1~0.21/0.09 due to sparsity! print() print('ta=', ta) # ta= 20.7766697407 print('tb=', tb) # tb= 18.2862818241 print() return 0
from zmpo_dmrg.source import mpo_dmrg_class from zmpo_dmrg.source import mpo_dmrg_schedule from zmpo_dmrg.source.mpsmpo import mps_io #--------------------------------- # COPY MPS.h5 to each directory #--------------------------------- import os cmd = 'cp mps.h5 '+mol.path os.system(cmd) #--------------------------------- # LOAD MPS fname = mol.path+'/mps.h5' mps,qnum = mps_io.loadMPS(fname) lmps = [mps,qnum] bdim = [len(x) for x in qnum] if mol.comm.rank == 0: print(' bdim = ',bdim) # One site algorithm sval = 0.0 sz = 0.0 Dmax = max(bdim) dmrg = mpo_dmrg_class.mpo_dmrg() dmrg.nsite = mol.sbas/2 dmrg.nhops = mol.sbas dmrg.isym = 2 dmrg.build() dmrg.comm = mol.comm
def test_HRfac(): # LOAD MPS fname = './mps.h5' mps, qnum = mps_io.loadMPS(fname) lmps = [mps, qnum] bdim = [len(x) for x in qnum] print(' bdim = ', bdim) nsite = len(mps) qphys = mpo_dmrg_qphys.initSpatialOrb(nsite, 2) ta = 0. tb = 0. nbas = 2 * nsite # random hmo = numpy.random.uniform(-1, 1, (nbas, nbas)) hmo[::2, 1::2] = hmo[1::2, ::2] = 0. # [ij|kl] eri = numpy.random.uniform(-1, 1, (nbas, nbas, nbas, nbas)) eri[::2, 1::2] = eri[1::2, ::2] = eri[:, :, ::2, 1::2] = eri[:, :, 1::2, ::2] = 0. # The spin symmetry is essential. # <ij|kl>=[ik|jl] eri = eri.transpose(0, 2, 1, 3) maxn = 3 for p in [6, 5]: isz = p % 2 for isite in range(min(nsite, 5)): ql = qnum[isite] qn = qphys[isite] qr = qnum[isite + 1] cl = qtensor_util.classification(ql) cn = qtensor_util.classification(qn) cr = qtensor_util.classification(qr) print() print('isite/nsite=', isite, nsite) site = mps[isite] tmps = qtensor.qtensor([False, False, True]) tmps.fromDenseTensor(site, [ql, qn, qr]) print('mps site info:') tmps.prt() if isite < maxn: # Reference value: t0 = time.time() pindx = (p, 0) qpts = numpy.array([0.3]) op = mpo_dmrg_opers.genHRfacSpatial(pindx, nbas, isite, hmo, eri, qpts) print(' wop=', op.shape) #csite = numpy.einsum('lrij,ajb->lairb',op,site) csite = numpy.tensordot(op, site, axes=([3], [1])) # lriab csite = csite.transpose(0, 3, 2, 1, 4) # lriab->lairb s = csite.shape csite = csite.reshape((s[0] * s[1], s[2], s[3] * s[4])) t1 = time.time() print(' t1=', t1 - t0) # Lowering the symmetry of MPS? qop = qtensor_opers.genHRfacSpatialQt(pindx, nbas, isite, hmo, eri, qpts) # We need to change qop construction allowing given qsyms ! tmps2 = tmps.reduceQsymsToN() #tmps2 = tmps2.projectionNMs(tmps.qsyms) #diff = numpy.linalg.norm(tmps2.value-tmps.value) #print ' diff=',diff tmps2 = qtensor.tensordot(qop, tmps2, axes=([3], [1])) tmps2 = tmps2.transpose(0, 3, 2, 1, 4) tmps2 = tmps2.merge([[0, 1], [2], [3, 4]]) tsite = tmps2.toDenseTensor() t2 = time.time() print(' t2=', t2 - t1) assert csite.shape == tsite.shape diff = numpy.linalg.norm(tsite - csite) print('isite,diff=', isite, csite.shape, diff, ' t0=', t1 - t0, ' t1=', t2 - t1) assert diff < 1.e-10 ta += t1 - t0 tb += t2 - t1 else: # Reference value: t0 = time.time() pindx = (p, 0) qpts = numpy.array([0.3]) t1 = time.time() qop = qtensor_opers.genHRfacSpatialQt(pindx, nbas, isite, hmo, eri, qpts) tmps2 = tmps.reduceQsymsToN() tmps2 = qtensor.tensordot(qop, tmps2, axes=([3], [1])) print('before transposing:') tmps2.prt() tmps2 = tmps2.transpose(0, 3, 2, 1, 4) print('after transposing:') tmps2.prt() tmps2 = tmps2.merge([[0, 1], [2], [3, 4]]) print('after merging:') tmps2.prt() t2 = time.time() print('isite=', isite, ' t2=', t2 - t1) print() print('ta=', ta) print('tb=', tb) print() return 0
def test_Hfac(): # #------------------------------------------------------------------------ # The large memory cost of Wop*|Psi> [O(K2D2)] requires, for large Dmps # and Dwop, a sequential compression should be implemented as a sweep, # such that the [Wsite*MPSsite] can be avoided partially. #------------------------------------------------------------------------ # # isite,diff= 0 (1, 4, 232) 0.0 t0= 0.0527150630951 t1= 0.0431280136108 # isite,diff= 1 (232, 4, 928) 0.0 t0= 0.0108029842377 t1= 0.050961971283 # isite,diff= 2 (928, 4, 3712) 0.0 t0= 0.146265029907 t1= 0.182390928268 # isite,diff= 3 (3712, 4, 14848) 0.0 t0= 2.27565908432 t1= 1.25626206398 # isite= 4 [14848 4 58870] t1= 3.11030101776 # isite= 5 [ 58870 4 105676] t1= 30.1744351387 # isite= 6 [105676 4 162806] t1= 101.686741114 # isite= 7 [162806 4 76966] t1= 75.6775298119 # isite= 8 [76966 4 29638] t1= 8.44200801849 # isite= 9 [29638 4 9976] t1= 1.69688081741 # isite= 10 [9976 4 3074] t1= 0.343852996826 # isite= 11 [3074 4 870] t1= 0.101953983307 # isite= 12 [870 4 232] t1= 0.044182062149 # isite= 13 [232 4 1] t1= 0.00591492652893 # # LOAD MPS fname = './mps.h5' mps, qnum = mps_io.loadMPS(fname) lmps = [mps, qnum] bdim = [len(x) for x in qnum] print(' bdim = ', bdim) nsite = len(mps) qphys = mpo_dmrg_qphys.initSpatialOrb(nsite, 2) ta = 0. tb = 0. nbas = 2 * nsite # random hmo = numpy.random.uniform(-1, 1, (nbas, nbas)) hmo[::2, 1::2] = hmo[1::2, ::2] = 0. # [ij|kl] eri = numpy.random.uniform(-1, 1, (nbas, nbas, nbas, nbas)) eri[::2, 1::2] = eri[1::2, ::2] = eri[:, :, ::2, 1::2] = eri[:, :, 1::2, ::2] = 0. # The spin symmetry is essential. # <ij|kl>=[ik|jl] eri = eri.transpose(0, 2, 1, 3) maxn = 3 for p in [6, 5]: isz = p % 2 hq = hmo[isz] vqrs = eri[isz] for isite in range(min(nsite, 5)): ql = qnum[isite] qn = qphys[isite] qr = qnum[isite + 1] cl = qtensor_util.classification(ql) cn = qtensor_util.classification(qn) cr = qtensor_util.classification(qr) print() print('isite/nsite=', isite, nsite) site = mps[isite] tmps = qtensor.qtensor([False, False, True]) tmps.fromDenseTensor(site, [ql, qn, qr]) print('mps site info:') tmps.prt() if isite < maxn: t0 = time.time() #op = mpo_dmrg_opers.genWfacSpatial(nbas,isite,hq,vqrs) op = mpo_dmrg_opers.genHfacSpatial(p, nbas, isite, hq, vqrs) print(' wop=', op.shape) #csite = numpy.einsum('lrij,ajb->lairb',op,site) csite = numpy.tensordot(op, site, axes=([3], [1])) # lriab csite = csite.transpose(0, 3, 2, 1, 4) # lriab->lairb s = csite.shape csite = csite.reshape((s[0] * s[1], s[2], s[3] * s[4])) t1 = time.time() print(' t1=', t1 - t0) #qop = qtensor_opers.genWfacSpatialQt(nbas,isite,hq,vqrs,isz) qop = qtensor_opers.genHfacSpatialQt(p, nbas, isite, hq, vqrs) tmps2 = qtensor.tensordot(qop, tmps, axes=([3], [1])) tmps2 = tmps2.transpose(0, 3, 2, 1, 4) tmps2 = tmps2.merge([[0, 1], [2], [3, 4]]) tsite = tmps2.toDenseTensor() t2 = time.time() print(' t2=', t2 - t1) assert csite.shape == tsite.shape diff = numpy.linalg.norm(tsite - csite) print('isite,diff=', isite, csite.shape, diff, ' t0=', t1 - t0, ' t1=', t2 - t1) assert diff < 1.e-10 ta += t1 - t0 tb += t2 - t1 else: if isite == maxn: print('>>> Check internal consistency <<<') t0 = time.time() #qop = qtensor_opers.genWfacSpatialQt(nbas,isite,hq,vqrs,isz) qop = qtensor_opers.genHfacSpatialQt(p, nbas, isite, hq, vqrs) tmps2 = qtensor.tensordot(qop, tmps, axes=([3], [1])) tmps2 = tmps2.transpose(0, 3, 2, 1, 4) tmps2 = tmps2.merge([[0, 1], [2], [3, 4]]) tmps2.prt() t1 = time.time() sum1 = numpy.sum(tmps2.value) print('isite=', isite, tmps.shape, ' t0=', t1 - t0, ' sum=', sum1) tmps2 = None qop = qtensor_opers.genHfacSpatialQt0(p, nbas, isite, hq, vqrs) tmps2 = qtensor.tensordot(qop, tmps, axes=([3], [1])) tmps2 = tmps2.transpose(0, 3, 2, 1, 4) tmps2 = tmps2.merge([[0, 1], [2], [3, 4]]) tmps2.prt() t2 = time.time() sum2 = numpy.sum(tmps2.value) print('isite=', isite, tmps2.shape, ' t1=', t2 - t1, ' sum=', sum2) tmps2 = None diff = abs(sum1 - sum2) print('diff =', diff) assert diff < 1.e-10 print() print('ta=', ta) print('tb=', tb) print() return 0
def test_Wfac(): # # Wop*Site = [256*30,4,30*1015] = [935424000] - 7G # # isite/nsite= 4 14 # Basic information: # rank = 3 shape= [ 256 4 1015] nsyms= [25 4 36] # nblks_allowed = 100 nblks = 3600 # size_allowed = 62804 size = 1039360 savings= 0.0604256465517 # wop= (30, 30, 4, 4) # t1= 13.3494501114 # t2= 2.59876251221e-05 # # isite/nsite= 5 14 --- 50G for storage. # Basic information: # rank = 3 shape= [1015 4 1822] nsyms= [36 4 44] # nblks_allowed = 136 nblks = 6336 # size_allowed = 376882 size = 7397320 savings= 0.0509484516014 # wop= (30, 30, 4, 4) # #>>> Sparse op becomes better for D~500 for computational time. # #isite/nsite= 3 14 #Basic information: # rank = 3 shape= [ 64 4 256] nsyms= [16 4 25] # nblks_allowed = 64 nblks = 1600 # size_allowed = 4900 size = 65536 savings= 0.0747680664062 # wop= (30, 30, 4, 4) # t1= 0.615185976028 # t2= 0.715934991837 #isite,diff= 3 (1920, 4, 7680) 0.0 t0= 0.615185976028 t1= 0.715934991837 # #isite/nsite= 4 14 #Basic information: # rank = 3 shape= [ 256 4 1015] nsyms= [25 4 36] # nblks_allowed = 100 nblks = 3600 # size_allowed = 62804 size = 1039360 savings= 0.0604256465517 # wop= (30, 30, 4, 4) # t1= 10.8923699856 # t2= 3.5177989006 #isite,diff= 4 (7680, 4, 30450) 0.0 t0= 10.8923699856 t1= 3.5177989006 # #ta= 11.5469501019 #tb= 4.54184865952 # # LOAD MPS fname = './mps.h5' mps, qnum = mps_io.loadMPS(fname) lmps = [mps, qnum] bdim = [len(x) for x in qnum] print(' bdim = ', bdim) nsite = len(mps) qphys = mpo_dmrg_qphys.initSpatialOrb(nsite, 2) ta = 0. tb = 0. isz = 0 nbas = 2 * nsite # random hmo = numpy.random.uniform(-1, 1, (nbas, nbas)) hmo[::2, 1::2] = hmo[1::2, ::2] = 0. # [ij|kl] eri = numpy.random.uniform(-1, 1, (nbas, nbas, nbas, nbas)) eri[::2, 1::2] = eri[1::2, ::2] = eri[:, :, ::2, 1::2] = eri[:, :, 1::2, ::2] = 0. # The spin symmetry is essential. # <ij|kl>=[ik|jl] eri = eri.transpose(0, 2, 1, 3) hq = hmo[isz] vqrs = eri[isz] maxn = 3 for isite in range(min(nsite, maxn)): ql = qnum[isite] qn = qphys[isite] qr = qnum[isite + 1] cl = qtensor_util.classification(ql) cn = qtensor_util.classification(qn) cr = qtensor_util.classification(qr) print() print('isite/nsite=', isite, nsite) site = mps[isite] tmps = qtensor.qtensor([False, False, True]) tmps.fromDenseTensor(site, [ql, qn, qr]) tmps.prt() t0 = time.time() op = mpo_dmrg_opers.genWfacSpatial(nbas, isite, hq, vqrs) print(' wop=', op.shape) #csite = numpy.einsum('lrij,ajb->lairb',op,site) csite = numpy.tensordot(op, site, axes=([3], [1])) # lriab csite = csite.transpose(0, 3, 2, 1, 4) # lriab->lairb s = csite.shape csite = csite.reshape((s[0] * s[1], s[2], s[3] * s[4])) t1 = time.time() print(' t1=', t1 - t0) qop = qtensor_opers.genWfacSpatialQt(nbas, isite, hq, vqrs, isz) tmps2 = qtensor.tensordot(qop, tmps, axes=([3], [1])) tmps2 = tmps2.transpose(0, 3, 2, 1, 4) tmps2 = tmps2.merge([[0, 1], [2], [3, 4]]) tsite = tmps2.toDenseTensor() t2 = time.time() print(' t2=', t2 - t1) assert csite.shape == tsite.shape diff = numpy.linalg.norm(tsite - csite) print('isite,diff=', isite, csite.shape, diff, ' t0=', t1 - t0, ' t1=', t2 - t1) assert diff < 1.e-10 ta += t1 - t0 tb += t2 - t1 print() print('ta=', ta) print('tb=', tb) print() return 0
def test_tensordot(): # LOAD MPS fname = './mps.h5' mps, qnum = mps_io.loadMPS(fname) lmps = [mps, qnum] bdim = [len(x) for x in qnum] print(' bdim = ', bdim) for item in mps: print(item.shape) for item in qnum: print(item) nsite = len(mps) print(nsite) qphys = mpo_dmrg_qphys.initSpatialOrb(nsite, 2) print(qphys) print(len(qnum)) ta = 0. tb = 0. for isite in range(nsite): ql = qnum[isite] qn = qphys[isite] qr = qnum[isite + 1] cl = qtensor_util.classification(ql) cn = qtensor_util.classification(qn) cr = qtensor_util.classification(qr) site = mps[isite] print() print('isite=', isite) #print len(ql),len(qn),len(qr) #print 'cl',cl #print 'cn',cn #print 'cr',cr tmps = qtensor.qtensor([False, False, True]) tmps.fromDenseTensor(site, [ql, qn, qr]) tsite = tmps.toDenseTensor() diffDense = numpy.linalg.norm(tsite - site) print(' diffDense=', diffDense) assert diffDense < 1.e-12 ## ## test-1 ## #print site.shape #t0 = time.time() #tmp = numpy.tensordot(site,site,axes=([0,1],[0,1])) ##tmp1 = numpy.einsum('ijk,lmn',site,site) ##print 'outer=',numpy.linalg.norm(tmp-tmp1) #t1 = time.time() #print 'norm=',numpy.linalg.norm(tmp),'t1-t0=',t1-t0 #t1 = time.time() #tmp2 = qtensor.tensordot(tmps,tmps,axes=([0,1],[0,1]),debug=False) #t2 = time.time() #print 'norm=',numpy.linalg.norm(tmp2.value),'t2-t1=',t2-t1 #print 'ratio=',(t2-t1)/(t1-t0) #ta += t1-t0 #tb += t2-t1 ## compare #tmp3 = tmp2.toDenseTensor() #diffDense2 = numpy.linalg.norm(tmp-tmp3) #print ' diffDense2=',diffDense2 #assert diffDense2<1.e-12 # # test-2: full contraction # print('full contraction:', site.shape) t0 = time.time() tmp = numpy.tensordot(site, site, axes=([0, 1, 2], [0, 1, 2])) t1 = time.time() print('norm=', numpy.linalg.norm(tmp), 't1-t0=', t1 - t0) t1 = time.time() tmp2 = qtensor.tensordot(tmps, tmps, axes=([0, 1, 2], [0, 1, 2]), debug=False) t2 = time.time() print('norm=', numpy.linalg.norm(tmp2), 't2-t1=', t2 - t1) print('ratio=', (t2 - t1) / (t1 - t0)) ta += t1 - t0 tb += t2 - t1 # compare diff = numpy.linalg.norm(tmp - tmp2) print(' diff=', diff) assert diff < 1.e-10 print() print('ta=', ta) print('tb=', tb) print() return 0