def merge_adjpair(self, i0, i1):
assert i0 + 1 == i1
p0 = self.ptys[i0]
p1 = self.ptys[i1]
d0 = self.dims[i0]
d1 = self.dims[i1]
p01 = []
d01 = []
ns0 = len(p0)
ns1 = len(p1)
for j0 in range(ns0):
for j1 in range(ns1):
p01.append(p0[j0] ^ p1[j1])
d01.append(d0[j0] * d1[j1])
ptys = copy.deepcopy(self.ptys)
ptys.pop(i0)
ptys.pop(i0)
ptys.insert(i0, p01)
dims = copy.deepcopy(self.dims)
dims.pop(i0)
dims.pop(i0)
dims.insert(i0, d01)
new = PArray(ptys, dims)
for ix in np.ndindex(self.shape):
if not self.ifpc(ix): continue
nix = np.ravel_multi_index([ix[i0], ix[i1]], (ns0, ns1))
nix = ix[:i0] + tuple([nix]) + ix[i1 + 1:]
shape = list(self[ix].shape)
nshape = shape[i0] * shape[i1]
shape.pop(i0)
shape.pop(i0)
shape.insert(i0, nshape)
new[nix] = self[ix].reshape(shape)
new = new.merge_axis(i0)
return new
def merge_axis(self, i):
ptys = copy.deepcopy(self.ptys)
dims = copy.deepcopy(self.dims)
dic = {}
for idx, ip in enumerate(ptys[i]):
if ip not in dic:
dic[ip] = [idx]
else:
dic[ip].append(idx)
qsec = sorted(dic.keys())
nq = len(qsec)
dsec = [
sum([dims[i][iq] for iq in dic[qsec[idx]]]) for idx in range(nq)
]
ptys.pop(i)
ptys.insert(i, qsec)
dims.pop(i)
dims.insert(i, dsec)
new = PArray(ptys, dims)
for ix in np.ndindex(new.shape):
if not new.ifpc(ix): continue
tensors = []
for j in dic[qsec[ix[i]]]:
px = list(ix)
px[i] = j
tensors.append(self[tuple(px)])
new[ix] = np.concatenate(tensors, axis=i)
return new
def fill(self, vec):
ptr = 0
for ix in np.ndindex(self.shape):
if self.ifpc(ix):
shape = self[ix].shape
nelem = np.product(shape)
self[ix] = vec[ptr:ptr + nelem].reshape(shape)
ptr += nelem
return vec
def transpose(self, axes):
ptys_new = [self.ptys[i] for i in axes]
dims_new = [self.dims[i] for i in axes]
new = PArray(ptys_new, dims_new)
# OLD_INDEX
for ix in np.ndindex(self.shape):
if self.ifpc(ix):
ix_new = tuple([ix[i] for i in axes])
new[ix_new] = self[ix].transpose(axes)
return new
def prt(self):
print 'PArray.prt'
print ' ptys=', self.ptys
print ' dims=', self.dims
nrm2 = 0.0
for ix in np.ndindex(self.shape):
line = None
if self.ifpc(ix):
nrm2i = np.linalg.norm(self[ix])**2
nrm2 += nrm2i
line = ' norm2=' + str(nrm2i)
print ' iblk=', ix, self.ifpc(ix), self[ix].shape, line
print ' norm2=', nrm2, '\n'
return 0
def __new__(cls, ptys, dims):
shape = map(lambda x: len(x), ptys)
shape1 = map(lambda x: len(x), dims)
assert shape == shape1
obj = np.ndarray.__new__(cls, shape, dtype=np.object)
obj[:] = null.Null()
obj.ptys = copy.deepcopy(ptys)
obj.dims = copy.deepcopy(dims)
# Initialize
for ix in np.ndindex(obj.shape): # np.ndindex(3, 2, 1)
if not obj.ifpc(ix): continue
shp = obj.get_shp(ix)
obj[ix] = np.zeros(shp, dtype=dataType)
return obj
def ravel(self):
vec = np.empty((0))
for ix in np.ndindex(self.shape):
if self.ifpc(ix):
vec = np.append(vec, np.ravel(self[ix]))
return vec
def size(self):
size = 0
for ix in np.ndindex(self.shape):
if self.ifpc(ix): size += np.product(self[ix].shape)
return size
def random(self, fac=1.0):
for ix in np.ndindex(self.shape):
if self.ifpc(ix):
self[ix] = fac * np.random.uniform(-1, 1, self[ix].shape)
return 0
def scale(self, fac=1.0):
for ix in np.ndindex(self.shape):
if self.ifpc(ix):
self[ix] = fac * self[ix]
return 0