def __init__(self, X, Y, C=None, dtype=None):
if dtype is None:
dtype = X.dtype
self.pX = PyMatrix.init_from(X, dtype)
self.pY = PyMatrix.init_from(Y, dtype)
self.pC = PyMatrix.init_from(C, dtype)
Z = None if C is None else smat.csr_matrix(self.Y.dot(self.C))
self.pZ = PyMatrix.init_from(Z, dtype) # Z = Y * C
self.dtype = dtype
def __init__(self, W, C=None, dtype=None):
if C is not None:
if isinstance(C, PyMatrix):
assert C.buf.shape[0] == W.shape[1]
else:
assert C.shape[0] == W.shape[1], 'C:{} W:{}'.format(
C.shape, W.shape)
if dtype is None:
dtype = W.dtype
self.pC = PyMatrix.init_from(C, dtype)
self.pW = PyMatrix.init_from(W, dtype)
def __init__(self, X, Y, C=None, bias=-1.0, dtype=None):
if dtype is None:
dtype = X.dtype
self.bias = bias
if self.bias > 0:
X = smat_util.append_column(X, self.bias)
self.pX = PyMatrix.init_from(X, dtype)
self.pY = PyMatrix.init_from(Y, dtype)
self.pC = PyMatrix.init_from(C, dtype)
Z = None if C is None else smat.csr_matrix(self.Y.dot(self.C))
self.pZ = PyMatrix.init_from(Z, dtype) # Z = Y * C
self.dtype = dtype
def predict_new(
self,
X,
only_topk=None,
csr_codes=None,
cond_prob=None,
normalized=False,
threads=-1,
):
assert X.shape[1] == self.nr_features
if csr_codes is None:
dense = X.dot(self.W).toarray()
if cond_prob:
dense = cond_prob.transform(dense, inplace=True)
coo = smat_util.dense_to_coo(dense)
pred_csr = smat_util.sorted_csr_from_coo(coo.shape,
coo.row,
coo.col,
coo.data,
only_topk=only_topk)
else: # csr_codes is given
assert self.C is not None, "This model does not have C"
assert X.shape[1] == self.nr_features
assert csr_codes.shape[0] == X.shape[0]
assert csr_codes.shape[1] == self.nr_codes
if not csr_codes.has_sorted_indices:
csr_codes = csr_codes.sorted_indices()
if (csr_codes.data == 0).sum() != 0:
# this is a trick to avoid zero entries explicit removal from the smat_dot_smat
offset = sp.absolute(csr_codes.data).max() + 1
csr_codes = smat.csr_matrix(
(csr_codes.data + offset, csr_codes.indices,
csr_codes.indptr),
shape=csr_codes.shape,
)
pZ = PyMatrix.init_from(csr_codes, self.dtype)
csr_labels, pred_csr = clib.multilabel_predict_with_codes(
X, self.pW, self.pC, pZ, threads=threads)
csr_labels.data -= offset
else:
pZ = PyMatrix.init_from(csr_codes.sorted_indices(), self.dtype)
csr_labels, pred_csr = clib.multilabel_predict_with_codes(
X, self.pW, self.pC, pZ, threads=threads)
val = pred_csr.data
if cond_prob:
val = cond_prob.transform(val, inplace=True)
val = cond_prob.combiner(val, csr_labels.data)
pred_csr = smat_util.sorted_csr(pred_csr, only_topk=only_topk)
if normalized:
pred_csr = sk_normalize(pred_csr, axis=1, copy=False, norm="l1")
return pred_csr
def predict_new(self,
X,
only_topk=None,
csr_codes=None,
beam_size=2,
max_depth=None,
cond_prob=True,
normalized=False,
threads=-1):
if max_depth is None:
max_depth = self.depth
if cond_prob is None or cond_prob == False:
cond_prob = PostProcessor(Transform.identity, Combiner.noop)
if cond_prob == True:
cond_prob = PostProcessor(Transform.get_lpsvm(3), Combiner.mul)
assert isinstance(cond_prob, PostProcessor), tpye(cond_prob)
assert X.shape[1] == self.nr_features
if self.bias > 0:
X = smat_util.append_column(X, self.bias)
pX = PyMatrix.init_from(X, dtype=self.model_chain[0].pW.dtype)
max_depth = min(self.depth, max_depth)
pred_csr = csr_codes
for d in range(max_depth):
cur_model = self.model_chain[d]
local_only_topk = only_topk if d == (max_depth - 1) else beam_size
pred_csr = cur_model.predict_new(pX,
only_topk=local_only_topk,
csr_codes=pred_csr,
cond_prob=cond_prob,
threads=threads)
if normalized:
pred_csr = sk_normalize(pred_csr, axis=1, copy=False, norm='l1')
return pred_csr
def __init__(self, X, Y, C=None, bias=-1.0, dtype=None, Z_pred=None):
if dtype is None:
dtype = X.dtype
self.bias = bias
if self.bias > 0:
X = smat_util.append_column(X, self.bias)
self.pX = PyMatrix.init_from(X, dtype)
self.pY = PyMatrix.init_from(Y, dtype)
self.pC = PyMatrix.init_from(C, dtype)
Z = None if C is None else smat.csr_matrix(self.Y.dot(self.C))
if Z_pred is not None and Z is not None:
print("Z", Z.shape)
print("Z_pred", Z_pred.shape)
Z = Z + Z_pred
Z = Z.tocsr()
self.pZ = PyMatrix.init_from(Z, dtype) # Z = Y * C
self.dtype = dtype
def balaced_ordinal_gen(self, kdim, depth, seed, threads=-1):
assert int(2**sp.log2(kdim)) == kdim
sp.random.seed(seed)
random_matrix = sp.randn(self.feat_mat.shape[1], depth)
X = PyMatrix(self.feat_mat.dot(random_matrix))
codes = sp.zeros(X.rows, dtype=sp.uint32)
new_depth = depth * int(sp.log2(kdim))
clib.get_codes(X, new_depth, Indexer.KDTREE_CYCLIC, seed, codes, threads=threads)
return codes
def predict_values(self, X, inst_idx, label_idx, out=None, threads=-1):
assert X.shape[1] == self.nr_features
if out is None:
out = sp.zeros(inst_idx.shape, dtype=self.pW.dtype)
pX = PyMatrix.init_from(X, dtype=self.pW.dtype)
out = clib.sparse_inner_products(pX,
self.pW,
inst_idx.astype(sp.uint32),
label_idx.astype(sp.uint32),
out,
threads=threads)
return out
def __init__(self,
X,
Y,
C=None,
dtype=None,
Z_pred=None,
negative_sampling_scheme=None):
if dtype is None:
dtype = X.dtype
self.pX = PyMatrix.init_from(X, dtype)
self.pY = PyMatrix.init_from(Y, dtype)
self.pC = PyMatrix.init_from(C, dtype)
Z = None if C is None else smat.csr_matrix(self.Y.dot(self.C))
if negative_sampling_scheme is None or negative_sampling_scheme == 1:
Z = Z
elif negative_sampling_scheme is not None:
if negative_sampling_scheme == 0:
Z = (Z + Z_pred).tocsr()
elif negative_sampling_scheme == 1:
Z = Z
elif negative_sampling_scheme == 2 and Z_pred is not None:
Z = Z_pred
self.pZ = PyMatrix.init_from(Z, dtype) # Z = Y * C
self.dtype = dtype
def predict_new(self,
X,
only_topk=None,
csr_codes=None,
beam_size=2,
max_depth=None,
cond_prob=True,
normalized=False,
threads=-1):
if max_depth is None:
max_depth = self.depth
if cond_prob is None or cond_prob == False:
cond_prob = PostProcessor(Transform.identity, Combiner.noop)
if cond_prob == True:
cond_prob = PostProcessor(Transform.get_lpsvm(3), Combiner.mul)
assert isinstance(cond_prob, PostProcessor), tpye(cond_prob)
pX = PyMatrix.init_from(X, dtype=self.model_chain[0].pW.dtype)
max_depth = min(self.depth, max_depth)
transform = cond_prob.transform if cond_prob else Transform.identity
pred_csr = csr_codes
#timer = WallTimer()
for d in range(max_depth):
'''
print('predict at depth {}'.format(d))
sys.stdout.flush()
timer.tic()
'''
cur_model = self.model_chain[d]
local_only_topk = only_topk if d == (max_depth - 1) else beam_size
pred_csr = cur_model.predict_new(pX,
only_topk=local_only_topk,
csr_codes=pred_csr,
transform=transform,
cond_prob=cond_prob,
threads=threads)
'''
print('>>> {}ms'.format(timer.toc()))
sys.stdout.flush()
'''
#if cond_prob and normalized: # perform normalization to avoid numerical issue
# pred_csr = sk_normalize(pred_csr, axis=1, copy=False, norm='l1')
#print('d = {} codes:{} nnz:{}'.format(d, pred_csr.shape[1], pred_csr.nnz))
#pred_csr.data[:] = sp.exp(pred_csr.data[:])
if normalized:
pred_csr = sk_normalize(pred_csr, axis=1, copy=False, norm='l1')
return pred_csr
def predict_new(self,
X,
only_topk=None,
transform=None,
csr_codes=None,
cond_prob=None,
normalized=False,
threads=-1):
assert X.shape[1] == self.nr_features
if csr_codes is None:
dense = X.dot(self.W).toarray()
if transform:
dense = transform(dense, inplace=True)
coo = smat_util.dense_to_coo(dense)
pred_csr = smat_util.sorted_csr_from_coo(coo.shape,
coo.row,
coo.col,
coo.data,
only_topk=only_topk)
else: # csr_codes is given
assert self.C is not None, "This model does not have C"
assert X.shape[1] == self.nr_features
assert csr_codes.shape[0] == X.shape[0]
assert csr_codes.shape[1] == self.nr_codes
pZ = PyMatrix.init_from(csr_codes, self.dtype)
csr_labels, pred_csr = clib.multilabel_predict_with_codes(
X, self.pW, self.pC, pZ, threads=threads)
val = pred_csr.data
if transform:
val = transform(val, inplace=True)
if cond_prob:
val[:] = cond_prob.combiner(val, csr_labels.data)
pred_csr = smat_util.sorted_csr(pred_csr, only_topk=only_topk)
if normalized:
pred_csr = sk_normalize(pred_csr, axis=1, copy=False, norm='l1')
return pred_csr
def __init__(self, feat_mat):
self.py_feat_mat = PyMatrix.init_from(feat_mat)
def test_svm(datafolder='dataset/Eurlex-4K', depth=3):
data = Data.load(datafolder, depth=depth)
X = PyMatrix(data.X, dtype=data.X.dtype)
#X = data.X
Y = data.Y
C = data.C
only_topk = 20
topk = 10
Cp = 1
Cn = 1
threshold = 0.01
#solver_type = L2R_LR_DUAL
solver_type = L2R_L2LOSS_SVC_DUAL
# test multi-label with codes
prob = MLProblem(X, Y, C)
m = MLModel.train(prob,
threshold=threshold,
solver_type=solver_type,
Cp=Cp,
Cn=Cn)
pred_Y = m.predict(X, only_topk=only_topk)
print('sparse W with top {}'.format(topk))
metric = Metrics.generate(Y, pred_Y, topk)
print(metric)
'''
print('|W|^2 = {}'.format((m.W.toarray() * m.W.toarray()).sum()))
coo = smat_util.dense_to_coo(sp.ones(pred_Y.shape))
YY = smat_util.sorted_csr(smat.csr_matrix(m.predict_values(X, coo.row, coo.col).reshape(pred_Y.shape)))
metric = Metrics.generate(Y, YY, topk)
print(metric)
YY = smat_util.sorted_csr(smat.csr_matrix(X.dot(m.W)))
metric = Metrics.generate(Y, YY, topk)
print(metric)
'''
# test hierarchical multi-label
print('Hierarchical-Multilabel')
beam_size = 4
min_labels = 2
nr_splits = 2
m = ml_train(prob,
hierarchical=True,
min_labels=min_labels,
threshold=threshold,
solver_type=solver_type,
Cp=Cp,
Cn=Cn)
print('m.depth = {}'.format(m.depth))
pred_Y = m.predict(X, beam_size=beam_size, only_topk=only_topk)
print(pred_Y.shape)
print('sparse W with top {}'.format(topk))
metric = Metrics.generate(Y, pred_Y, topk)
print(metric)
'''
max_depth = 2
print('Predict up to depth = {}'.format(max_depth))
pred_Y = m.predict(X, only_topk=only_topk, max_depth=max_depth)
trueY = Y.copy()
for d in range(m.depth - 1, max_depth - 1, -1):
trueY = trueY.dot(m.model_chain[d].C)
metric = Metrics.generate(trueY, pred_Y, topk)
print(metric)
#print('|W|^2 = {}'.format((m.W.toarray() * m.W.toarray()).sum()))
'''
# test pure multi-label
print('pure one-vs-rest Multi-label')
prob = MLProblem(X, Y)
m = MLModel.train(prob,
threshold=threshold,
solver_type=solver_type,
Cp=Cp,
Cn=Cn)
pred_Y = m.predict(X, only_topk=only_topk)
metric = Metrics.generate(Y, pred_Y, topk)
print(metric)
print('|W|^2 = {}'.format((m.W.toarray() * m.W.toarray()).sum()))