def omp_n_predict(X, centroids, num_active):
''' omp prediction
Input:
X: the data matrix
centroids: the centroids matrix
num_active: the number of active components
'''
idx = np.empty((X.shape[0], num_active), dtype=np.int)
val = np.zeros((X.shape[0], num_active))
C = mathutil.dot(centroids, centroids.T)
# find the most active k
dots = None
dots_abs = None
for start in range(0, X.shape[0], _MINIBATCH):
end = min(start + _MINIBATCH, X.shape[0])
batchsize = end - start
if dots is None:
dots = mathutil.dot(X[start:end], centroids.T)
dots_abs = np.empty_like(dots)
else:
mathutil.dot(X[start:end], centroids.T, out=dots[:batchsize])
# we only compute the dots once, and keep them for future use
for i in range(num_active):
np.abs(dots, out=dots_abs)
idx[start:end, i] = np.argmax(dots_abs[:batchsize], axis=1)
val[start:end, i] = dots[np.arange(batchsize), idx[start:end, i]]
# remove the effect from dots
dots[:batchsize] -= C[idx[start:end, i]] * \
val[start:end, i][:, np.newaxis]
return idx, val
def testdot_with_out(self):
for A, B in self.test_matrices:
result_ref = np.dot(A, B)
result = np.empty(result_ref.shape, dtype=A.dtype)
mathutil.dot(A, B, out=result)
self.assertTrue(result.flags['C_CONTIGUOUS'])
np.testing.assert_array_almost_equal(result, result_ref)
def omp_n_predict(X, centroids, num_active):
""" omp prediction
Input:
X: the data matrix
centroids: the centroids matrix
num_active: the number of active components
"""
idx = np.empty((X.shape[0], num_active), dtype=np.int)
val = np.zeros((X.shape[0], num_active))
C = mathutil.dot(centroids, centroids.T)
# find the most active k
dots = None
dots_abs = None
for start in range(0, X.shape[0], _MINIBATCH):
end = min(start + _MINIBATCH, X.shape[0])
batchsize = end - start
if dots is None:
dots = mathutil.dot(X[start:end], centroids.T)
dots_abs = np.empty_like(dots)
else:
mathutil.dot(X[start:end], centroids.T, out=dots[:batchsize])
# we only compute the dots once, and keep them for future use
for i in range(num_active):
np.abs(dots, out=dots_abs)
idx[start:end, i] = np.argmax(dots_abs[:batchsize], axis=1)
val[start:end, i] = dots[np.arange(batchsize), idx[start:end, i]]
# remove the effect from dots
dots[:batchsize] -= C[idx[start:end, i]] * val[start:end, i][:, np.newaxis]
return idx, val
def testdot_with_out(self):
for A, B in self.test_matrices:
result_ref = np.dot(A,B)
result = np.empty(result_ref.shape, dtype = A.dtype)
mathutil.dot(A, B, out = result)
self.assertTrue(result.flags['C_CONTIGUOUS'])
np.testing.assert_array_almost_equal(result, result_ref)
def obj(wb, solver):
"""
The objective function used by fmin
"""
# obtain w and b
K = solver._K
dim = solver._dim
w = wb[: K * dim].reshape((dim, K))
b = wb[K * dim :]
# pred is a matrix of size [num_datalocal, K]
pred = mathutil.dot(solver._X, w)
pred += b
# compute the loss function
flocal, gpred = solver.loss(solver._Y, pred, solver._weight, **solver._lossargs)
glocal = np.empty(wb.shape)
glocal[: K * dim] = mathutil.dot(solver._X.T, gpred).flat
glocal[K * dim :] = gpred.sum(axis=0)
# add regularization term, but keep in mind that we have multiple nodes
freg, greg = solver.reg(w, **solver._regargs)
flocal += solver._num_data * solver._gamma * freg / mpi.SIZE
glocal[: K * dim] += solver._num_data * solver._gamma / mpi.SIZE * greg.ravel()
# do mpi reduction
mpi.barrier()
f = mpi.COMM.allreduce(flocal)
g = np.empty(glocal.shape, dtype=glocal.dtype)
mpi.COMM.Allreduce(glocal, g)
return f, g
def process(self, image):
'''Performs llc encoding.
'''
K = self.specs.get('k', 5)
reg = self.specs.get('reg', 1e-4)
D = self.dictionary
shape = image.shape[:-1]
X = image.reshape((np.prod(shape), image.shape[-1]))
# D_norm is the precomputed norm of the entries
if 'D_norm' not in self.specs:
self.specs['D_norm'] = (D**2).sum(1) / 2.
D_norm = self.specs['D_norm']
distance = mathutil.dot(X, -D.T)
distance += D_norm
# find the K closest indices
if bn is not None:
# use bottleneck which would be faster
IDX = bn.argpartsort(distance, K, axis=1)[:, :K]
else:
IDX = np.argsort(distance,1)[:, :K]
# do LLC approximate coding
coeff = np.zeros((X.shape[0], D.shape[0]))
ONES = np.ones(K)
Z = np.empty((K, D.shape[1]))
for i in range(X.shape[0]):
# shift to origin
Z[:] = D[IDX[i]]
Z -= X[i]
# local covariance
C = mathutil.dot(Z,Z.T)
# add regularization
C.flat[::K+1] += reg * C.trace()
w = np.linalg.solve(C,ONES)
coeff[i][IDX[i]] = w / w.sum()
return coeff.reshape(shape + (coeff.shape[1],))
def get_predictions_nn(X, weights,arch):
hid = mathutil.dot(X,weights[0])+weights[1]
hid = 1.0/(1+np.exp(-hid)) #sigmoid
pred = mathutil.dot(hid,weights[2])+weights[3]
#prob = pred - pred.max(axis=1)[:,np.newaxis]
#mathutil.exp(prob, out=prob)
#prob /= prob.sum(axis=1)[:, np.newaxis]
prob = 1.0/(1.0+np.exp(-pred))
prob = mpi.COMM.gather(prob)
hid = mpi.COMM.gather(hid)
if mpi.is_root():
return np.vstack(prob),np.vstack(hid)
else:
return np.zeros((0)),np.zeros((0))
def coclassify_dag(self, prob, classifier_weight=1.):
"""Perform co-classification when the hidden concept is organized as
a DAG.
"""
if self.graph == None:
raise ValueError, "No graph given."
# compute log(prob(y|s)) first
prob_ys = mathutil.dot(prob, self.invconfprob.T)
prob_ys /= prob_ys.sum(1)[:, np.newaxis]
prob_ys += np.finfo(np.float64).eps
np.log(prob_ys, out=prob_ys)
# using the tree structure to compute the max_{y\in c} prob_ys for
# each concept.
# basically, we need to follow the inverse topological order
max_prob_ys_in_c = np.zeros(
(prob_ys.shape[0], len(self.graph.nodes())))
for c in self.invtoporder:
cid = self.concept2id[c]
succ = [self.concept2id[s] for s in self.graph.successors(c)]
if len(succ) == 0:
# I am a leaf node
max_prob_ys_in_c[:, cid] = prob_ys[:, self.leaf2id[c]]
else:
max_prob_ys_in_c[:,cid] = np.max(\
max_prob_ys_in_c[:, succ], axis=1)
# now, combine the upstream probability and downstream probability to find
# the argmax
score = self.logprior + max_prob_ys_in_c.sum(0) * classifier_weight \
- self.log_concept_sizes * prob.shape[0]
best_cid = score.argmax()
slice = np.array(list(self.membership_map[best_cid]))
best_labels = slice[prob_ys[:, slice].argmax(1)]
return best_cid, best_labels
def coclassify_dag(self, prob, classifier_weight=1.0):
"""Perform co-classification when the hidden concept is organized as
a DAG.
"""
if self.graph == None:
raise ValueError, "No graph given."
# compute log(prob(y|s)) first
prob_ys = mathutil.dot(prob, self.invconfprob.T)
prob_ys /= prob_ys.sum(1)[:, np.newaxis]
prob_ys += np.finfo(np.float64).eps
np.log(prob_ys, out=prob_ys)
# using the tree structure to compute the max_{y\in c} prob_ys for
# each concept.
# basically, we need to follow the inverse topological order
max_prob_ys_in_c = np.zeros((prob_ys.shape[0], len(self.graph.nodes())))
for c in self.invtoporder:
cid = self.concept2id[c]
succ = [self.concept2id[s] for s in self.graph.successors(c)]
if len(succ) == 0:
# I am a leaf node
max_prob_ys_in_c[:, cid] = prob_ys[:, self.leaf2id[c]]
else:
max_prob_ys_in_c[:, cid] = np.max(max_prob_ys_in_c[:, succ], axis=1)
# now, combine the upstream probability and downstream probability to find
# the argmax
score = self.logprior + max_prob_ys_in_c.sum(0) * classifier_weight - self.log_concept_sizes * prob.shape[0]
best_cid = score.argmax()
slice = np.array(list(self.membership_map[best_cid]))
best_labels = slice[prob_ys[:, slice].argmax(1)]
return best_cid, best_labels
def obj(wb,solver):
'''
The objective function used by fmin
'''
# obtain w and b
K = solver._K
dim = solver._dim
w = wb[:K*dim].reshape((dim, K))
b = wb[K*dim:]
# pred is a matrix of size [num_datalocal, K]
mathutil.dot(solver._X, w, out = solver._pred)
solver._pred += b
# compute the loss function
if solver.gpredcache:
flocal,gpred = solver.loss(solver._Y, solver._pred, solver._weight,
solver._gpred, solver._gpredcache,
**solver._lossargs)
else:
flocal,gpred = solver.loss(solver._Y, solver._pred, solver._weight,
**solver._lossargs)
mathutil.dot(solver._X.T, gpred,
out = solver._glocal[:K*dim].reshape(dim, K))
solver._glocal[K*dim:] = gpred.sum(axis=0)
# we should normalize them with the number of data
flocal /= solver._num_data
solver._glocal /= solver._num_data
# add regularization term, but keep in mind that we have multiple nodes
# so we only carry it out on root to make sure we only added one
# regularization term
if mpi.is_root():
freg, greg = solver.reg(w, **solver._regargs)
flocal += solver._gamma * freg
solver._glocal[:K*dim] += solver._gamma * greg.ravel()
# do mpi reduction
mpi.barrier()
f = mpi.COMM.allreduce(flocal)
mpi.COMM.Allreduce(solver._glocal, solver._g)
######### DEBUG PART ##############
if np.isnan(f):
# check all the components to see what went wrong.
print 'rank %s: isnan X: %d' % (mpi.RANK,np.any(np.isnan(solver._X)))
print 'rank %s: isnan Y: %d' % (mpi.RANK,np.any(np.isnan(solver._Y)))
print 'rank %s: isnan flocal: %d' % (mpi.RANK,np.any(np.isnan(flocal)))
print 'rank %s: isnan pred: %d' % (mpi.RANK,np.any(np.isnan(solver._pred)))
print 'rank %s: isnan w: %d' % (mpi.RANK,np.any(np.isnan(w)))
print 'rank %s: isnan b: %d' % (mpi.RANK,np.any(np.isnan(b)))
return f, solver._g
def soap(self, X, out = None):
"""Performs second-order average pooling on a n*k matrix X. Returns a
k*(k+1)/2 vector that is the upper triangular part of the average pooled
region.
"""
X = np.ascontiguousarray(X, dtype = np.float64)
self._update_cache(X)
if self._use_cov:
self._cache_gram_matrix= np.cov(X, rowvar=0)
else:
mathutil.dot(X.T, X, out = self._cache_gram_matrix)
self._cache_gram_matrix /= X.shape[0]
if out is None:
out = np.empty(X.shape[1] * (X.shape[1] + 1) / 2)
out[:] = self._cache_gram_matrix[self._cache_triu_id[0],\
self._cache_triu_id[1]]
return out
def log_soap(self, X, out = None):
"""Performs matrix_log(second order average pooling)
"""
X = np.ascontiguousarray(X, dtype = np.float64)
self._update_cache(X)
if self._use_cov:
self._cache_gram_matrix= np.cov(X, rowvar=0)
else:
mathutil.dot(X.T, X, out = self._cache_gram_matrix)
self._cache_gram_matrix /= X.shape[0]
self._cache_gram_matrix.flat[::X.shape[1]+1] += self._reg
# compute the matrix log
logmat = scipy.linalg.logm(self._cache_gram_matrix)
if out is None:
out = np.empty(X.shape[1] * (X.shape[1] + 1) / 2)
out[:] = logmat[self._cache_triu_id[0], self._cache_triu_id[1]]
return out
def get_predictions_logreg_perclass(X, weights):
pred = mathutil.dot(X,weights[0])+weights[1]
prob = 1.0/(1.0+np.exp(-pred))
prob = mpi.COMM.gather(prob)
if mpi.is_root():
return np.vstack(prob)
else:
return np.zeros((0))
def get_predictions_logreg(X, weights):
pred = mathutil.dot(X,weights[0])+weights[1]
prob = pred - pred.max(axis=1)[:,np.newaxis]
mathutil.exp(prob, out=prob)
prob /= prob.sum(axis=1)[:, np.newaxis]
prob = mpi.COMM.gather(prob)
if mpi.is_root():
return np.vstack(prob)
else:
return np.zeros((0))
def coclassify_oracle(self, prob, cid):
"""Perform coclassification when the hidden concept is just given in cid
"""
prob_ys = mathutil.dot(prob, self.invconfprob.T)
prob_ys /= prob_ys.sum(1)[:, np.newaxis]
prob_ys += np.finfo(np.float64).eps
np.log(prob_ys, out=prob_ys)
slice = np.array(list(self.membership_map[cid]))
best_labels = slice[prob_ys[:, slice].argmax(1)]
return cid, best_labels
def coclassify_oracle(self, prob, cid):
"""Perform coclassification when the hidden concept is just given in cid
"""
prob_ys = mathutil.dot(prob, self.invconfprob.T)
prob_ys /= prob_ys.sum(1)[:, np.newaxis]
prob_ys += np.finfo(np.float64).eps
np.log(prob_ys, out=prob_ys)
slice = np.array(list(self.membership_map[cid]))
best_labels = slice[prob_ys[:, slice].argmax(1)]
return cid, best_labels
def obj(wb,solver):
'''
The objective function used by fmin
'''
# obtain w and b
Khidden = solver._Khidden
dim = solver._dim
whidden = wb[:Khidden*dim].reshape((dim, Khidden))
tree = solver._regargs['tree']
w = mathutil.dot(whidden, tree)
b = wb[Khidden*dim:]
# pred is a matrix of size [num_datalocal, K]
mathutil.dot(solver._X, w, out = solver._pred)
solver._pred += b
# compute the loss function
flocal,gpred = solver.loss(solver._Y, solver._pred, solver._weight,
**solver._lossargs)
mathutil.dot(mathutil.dot(solver._X.T, gpred), tree.T,
out = solver._glocal[:Khidden*dim].reshape(dim, Khidden))
solver._glocal[Khidden*dim:] = gpred.sum(axis=0)
# add regularization term, but keep in mind that we have multiple nodes
freg, greg = solver.reg(whidden, **solver._regargs)
flocal += solver._num_data * solver._gamma * freg / mpi.SIZE
solver._glocal[:Khidden*dim] += solver._num_data * solver._gamma \
* greg.ravel() / mpi.SIZE
# do mpi reduction
mpi.barrier()
f = mpi.COMM.allreduce(flocal)
mpi.COMM.Allreduce(solver._glocal, solver._g)
return f, solver._g
def process(self, image, out=None):
'''Performs llc encoding.
'''
K = self.specs.get('k', 5)
reg = self.specs.get('reg', 1e-4)
D = self.dictionary
shape = image.shape[:-1]
X = image.reshape((np.prod(shape), image.shape[-1]))
# D_norm is the precomputed norm of the entries
if 'D_norm' not in self.specs:
self.specs['D_norm'] = (D**2).sum(1) / 2.
D_norm = self.specs['D_norm']
distance = mathutil.dot(X, -D.T)
distance += D_norm
# find the K closest indices
if bn is not None:
# use bottleneck which would be faster
IDX = bn.argpartsort(distance, K, axis=1)[:, :K]
else:
IDX = np.argsort(distance,1)[:, :K]
# do LLC approximate coding
if out is None:
out = np.zeros((X.shape[0], D.shape[0]))
else:
out.resize((X.shape[0], D.shape[0]))
out[:] = 0
ONES = np.ones(K)
Z = np.empty((K, D.shape[1]))
for i in range(X.shape[0]):
# shift to origin
Z[:] = D[IDX[i]]
Z -= X[i]
# local covariance
C = mathutil.dot(Z,Z.T)
# add regularization
C.flat[::K+1] += reg * C.trace()
w = np.linalg.solve(C,ONES)
out[i][IDX[i]] = w / w.sum()
out.resize(shape + (out.shape[1],))
return out
def omp_n_maximize(X, labels, val, k):
"""Maximization of omp_n, with the given labels and vals.
Note that X is the local data hosted in each MPI node.
"""
dim = X.shape[1]
# G is the gram matrix of the activations
AtA_local = np.zeros((k, k))
AtX_local = np.zeros((k, dim))
A = None
for start in range(0, X.shape[0], _MINIBATCH):
end = min(start + _MINIBATCH, X.shape[0])
batchsize = end - start
if A is None:
A = np.zeros((batchsize, k))
else:
A[:] = 0
for i in range(batchsize):
A[i, labels[start + i]] = val[start + i]
AtA_local += mathutil.dot(A.T, A)
AtX_local += mathutil.dot(A[:batchsize].T, X[start:end])
AtA = np.empty_like(AtA_local)
AtX = np.empty_like(AtX_local)
mpi.COMM.Allreduce(AtA_local, AtA)
mpi.COMM.Allreduce(AtX_local, AtX)
# add a regularization term
isempty = np.diag(AtA) == 0
AtA.flat[:: k + 1] += 1e-8
centroids = np.ascontiguousarray(np.linalg.solve(AtA, AtX))
# let's deal with inactive guys
for i in range(k):
if isempty[i]:
# randomly restart one
centroids[i] = X[np.random.randint(X.shape[0])]
mpi.COMM.Bcast(centroids[i], root=mpi.elect())
scale = np.sqrt((centroids ** 2).sum(1)) + np.finfo(np.float64).eps
centroids /= scale[:, np.newaxis]
return centroids
def omp_n_maximize(X, labels, val, k):
'''Maximization of omp_n, with the given labels and vals.
Note that X is the local data hosted in each MPI node.
'''
dim = X.shape[1]
# G is the gram matrix of the activations
AtA_local = np.zeros((k, k))
AtX_local = np.zeros((k, dim))
A = None
for start in range(0, X.shape[0], _MINIBATCH):
end = min(start + _MINIBATCH, X.shape[0])
batchsize = end - start
if A is None:
A = np.zeros((batchsize, k))
else:
A[:] = 0
for i in range(batchsize):
A[i, labels[start + i]] = val[start + i]
AtA_local += mathutil.dot(A.T, A)
AtX_local += mathutil.dot(A[:batchsize].T, X[start:end])
AtA = np.empty_like(AtA_local)
AtX = np.empty_like(AtX_local)
mpi.COMM.Allreduce(AtA_local, AtA)
mpi.COMM.Allreduce(AtX_local, AtX)
# add a regularization term
isempty = (np.diag(AtA) == 0)
AtA.flat[::k + 1] += 1e-8
centroids = np.ascontiguousarray(np.linalg.solve(AtA, AtX))
# let's deal with inactive guys
for i in range(k):
if isempty[i]:
# randomly restart one
centroids[i] = X[np.random.randint(X.shape[0])]
mpi.COMM.Bcast(centroids[i], root=mpi.elect())
scale = np.sqrt((centroids**2).sum(1)) + np.finfo(np.float64).eps
centroids /= scale[:, np.newaxis]
return centroids
def omp1_predict(X, centroids):
''' omp1 prediction
This function does one-dimensional orthogonal matching pursuit.
the returned values are simply going to be the indices and
inner products.
'''
idx = np.empty(X.shape[0], dtype=np.int)
val = np.empty(X.shape[0])
# in case we are going to deal with a large matrix, we buffer dots to avoid
# multiple memory new / deletes.
dots = np.empty((min(_MINIBATCH, X.shape[0]), centroids.shape[0]),
dtype=X.dtype)
for start in range(0, X.shape[0], _MINIBATCH):
end = min(start + _MINIBATCH, X.shape[0])
batchsize = end - start
mathutil.dot(X[start:end], centroids.T, out=dots[:batchsize])
np.abs(dots, out=dots)
idx[start:end] = np.argmax(dots[:batchsize], axis=1)
val[start:end] = dots[range(batchsize), idx[start:end]]
mpi.barrier()
return idx, val
def omp1_predict(X, centroids):
''' omp1 prediction
This function does one-dimensional orthogonal matching pursuit.
the returned values are simply going to be the indices and
inner products.
'''
idx = np.empty(X.shape[0], dtype=np.int)
val = np.empty(X.shape[0])
# in case we are going to deal with a large matrix, we buffer dots to avoid
# multiple memory new / deletes.
dots = np.empty((min(_MINIBATCH, X.shape[0]), centroids.shape[0]),
dtype = X.dtype)
for start in range(0, X.shape[0], _MINIBATCH):
end = min(start+_MINIBATCH, X.shape[0])
batchsize = end-start
mathutil.dot(X[start:end], centroids.T, out = dots[:batchsize])
np.abs(dots, out=dots)
idx[start:end] = np.argmax(dots[:batchsize], axis=1)
val[start:end] = dots[range(batchsize), idx[start:end]]
mpi.barrier()
return idx, val
def train(self, incoming_patches):
size = mpi.COMM.allreduce(incoming_patches.shape[0])
b = - mpi.COMM.allreduce(np.sum(incoming_patches,axis=0)) / size
# remove the mean from data
patches = incoming_patches + b
covmat = mpi.COMM.allreduce(mathutil.dot(patches.T, patches)) / size
if mpi.RANK == 0:
eigval, eigvec = np.linalg.eigh(covmat)
reg = self.specs.get('reg', np.finfo(np.float64).eps)
W = eigvec * 1.0 / (np.sqrt(np.maximum(eigval, 0.0)) + reg)
else:
eigval, eigvec, W = None, None, None
W = mpi.COMM.bcast(W)
eigval = mpi.COMM.bcast(eigval)
eigvec = mpi.COMM.bcast(eigvec)
return (W, b), (eigval, eigvec)
confmats = None
for i in range(mpi.RANK, len(files), mpi.SIZE):
# for the i-th file the gt label is i
file = files[i]
predfile = preds[i]
fid = h5py.File(file, 'r')
features = np.array(fid['features'])
fid.close()
fid = h5py.File(predfile, 'r')
pred = np.array(fid['pred'])
fid.close()
# compute new prediction
diff = mathutil.softmax(pred)
diff[:, i] -= 1
features **= 2
weighted_direction = mathutil.dot(features, Hwinv) + Hbinv
dpred = diff * weighted_direction
if use_validation:
iter = 0
accu = sum((val_pred == i) & (val_label == i)) / float(
sum(val_label == i))
s_low = 0
s_high = scales[0]
while iter < 10:
s = (s_low + s_high) / 2
newprob = pred + dpred * s
mathutil.softmax(newprob, out=newprob)
newpred = newprob.argmax(1)
my_accu = sum(newpred == i) / float(pred.shape[0])
if my_accu > accu:
s_low = s
def forward(X, outputs):
output = mathutil.dot(X, w)
output += b
def testdot(self):
for A, B in self.test_matrices:
result = mathutil.dot(A, B)
result_ref = np.dot(A, B)
self.assertTrue(result.flags['C_CONTIGUOUS'])
np.testing.assert_array_almost_equal(result, result_ref)
files = glob.glob(os.path.join(FLAGS.folder, '*.mat'))
files.sort()
count = 0
for i in range(mpi.RANK, len(files), mpi.SIZE):
file = files[i]
print '%d / %d: %s' % (i, len(files), file)
fid = h5py.File(file, 'r')
features = np.array(fid['features'])
count += features.shape[0]
pred = np.dot(features, w)
pred += b
prob = mathutil.softmax(pred)
dpdp = prob * (1 - prob)
# the gradient for b is simple
hb += np.sum(dpdp, 0)
features **= 2
hw += mathutil.dot(features.T, dpdp)
fid.close()
del features
count = mpi.COMM.allreduce(count)
# we no longer need w and b, so we use them as mpi buffer
mpi.COMM.Allreduce(hw, w)
mpi.COMM.Allreduce(hb, b)
if mpi.is_root():
hw /= count
hb /= count
# add regularization term
hw += 2 * FLAGS.reg
np.savez(FLAGS.output, count=count, hw=w, hb=b)
confmats = None
for i in range(mpi.RANK, len(files), mpi.SIZE):
# for the i-th file the gt label is i
file = files[i]
predfile = preds[i]
fid = h5py.File(file, 'r')
features = np.array(fid['features'])
fid.close()
fid = h5py.File(predfile, 'r')
pred = np.array(fid['pred'])
fid.close()
# compute new prediction
diff = mathutil.softmax(pred)
diff[:,i] -= 1
features **= 2
weighted_direction = mathutil.dot(features, Hwinv) + Hbinv
dpred = diff * weighted_direction
if use_validation:
iter = 0
accu = sum((val_pred == i) & (val_label == i)) / float(sum(val_label == i))
s_low = 0
s_high = scales[0]
while iter < 10:
s = (s_low + s_high) / 2
newprob = pred + dpred * s
mathutil.softmax(newprob, out=newprob)
newpred = newprob.argmax(1)
my_accu = sum(newpred==i) / float(pred.shape[0])
if my_accu > accu:
s_low = s
else:
def loss_multiclass_logreg(Y, X, weights):
pred = mathutil.dot(X,weights[0])+weights[1]
local_likelihood = classifier.Loss.loss_multiclass_logistic(classifier.to_one_of_k_coding(Y, 0), pred, None)[0]
likelihood = mpi.COMM.allreduce(local_likelihood)
num_data = mpi.COMM.allreduce(len(Y))
return float(likelihood) / num_data
def testdot(self):
for A, B in self.test_matrices:
result = mathutil.dot(A, B)
result_ref = np.dot(A,B)
self.assertTrue(result.flags['C_CONTIGUOUS'])
np.testing.assert_array_almost_equal(result, result_ref)
def forward(X, outputs):
output = mathutil.dot(X, w)
output += b
