def evaluate(data, net):
margins = [0.01 * j for j in xrange(51)]
lossmatr = [[[0 for _ in xrange(3)] for _ in xrange(3)] for _ in margins]
for q in data:
C = ranknet.cost(q, net, 2.0, True)
labels = C[1]
pairs = C[3]
for pair in pairs:
i = pair[0]
j = pair[1]
p = pair[2]
s = 1
if labels[i] > labels[j]: s = 2
if labels[i] < labels[j]: s = 0
for jm in xrange(len(margins)):
z = 1
if p > 0.5 + margins[jm]: z = 2
if p < 0.5 - margins[jm]: z = 0
lossmatr[jm][z][s] += 1
for jm in xrange(len(margins)):
print margins[jm], lossmatr[jm]
def evaluate(data, net):
margins = [0.01*j for j in xrange(51)]
lossmatr = [[[0 for _ in xrange(3)] for _ in xrange(3)] for _ in margins]
for q in data:
C = ranknet.cost(q, net, 2.0, True)
labels = C[1]
pairs = C[3]
for pair in pairs:
i = pair[0]
j = pair[1]
p = pair[2]
s = 1
if labels[i] > labels[j]: s = 2
if labels[i] < labels[j]: s = 0
for jm in xrange(len(margins)):
z = 1
if p > 0.5 + margins[jm]: z = 2
if p < 0.5 - margins[jm]: z = 0
lossmatr[jm][z][s] += 1
for jm in xrange(len(margins)):
print margins[jm], lossmatr[jm]
def xtest_training_2(self):
# trainsg on several queries
data = []
d = range(10)
for j in d:
data.append( [ [j, random.choice([0, 1])] + [random.random() for _ in xrange(self.ninp)] for _ in xrange(self.nq) ] )
print data
nepoch = 10000 # number of training epochs
rate = 0.1 # learning rate
nprint = 1000 # print frequency
# compute current cost and estimations
for je in xrange(nepoch):
# select training sample at random
jq = random.choice(d)
if je % nprint == 0:
# compute cost of a first sample
C = ranknet.cost(data[0], self.model, self.sigma)
print je, C[0], C[1], C[2]
print "w:", self.model.getw()
# compute gradients
g = ranknet.gradient(data[jq], self.model, self.sigma)
# update weights
w = la.vsum( self.model.getw(), la.sax(-rate, g) )
self.model.setw(w)
# final report
for query in data:
print "Query: ", query[0][0]
C = ranknet.cost(query, self.model, self.sigma)
for j in xrange(len(query)):
print query[j][1], C[1][j]
def test_gradient(self):
# analytical gradient
dbpr = ranknet.gradient(self.query, self.model, self.sigma)
# numerical gradient
dw = 1.e-5
jw = 7
w_u = self.w[:]
w_u[jw] = w_u[jw] + dw
self.model.setw(w_u)
C_u = ranknet.cost(self.query, self.model, self.sigma)
w_d = self.w[:]
w_d[jw] = w_d[jw] - dw
self.model.setw(w_d)
C_d = ranknet.cost(self.query, self.model, self.sigma)
dnum = (C_u[0] - C_d[0]) / dw / 2.0
print dbpr[jw], dnum
def test_gradient_2(self):
# run tests
ntest = 100
ninp = 10
nq = 20
nhid = self.nhid
nw = (ninp + 1) * nhid + (nhid + 1) * 1 # total number of weights
dw = 1.e-6
for j in xrange(ntest):
# generate query data
query = [ [0, random.choice([0, 1])] + [random.random() for _ in xrange(ninp)] for _ in xrange(nq) ]
# get analytical gradient
grad = ranknet.gradient(query, self.model, self.sigma)
# select weight at random
jw = random.choice(xrange(nw))
# numerical derivative
w_u = self.w[:]
w_u[jw] = w_u[jw] + dw
self.model.setw(w_u)
C_u = ranknet.cost(query, self.model, self.sigma)
w_d = self.w[:]
w_d[jw] = w_d[jw] - dw
self.model.setw(w_d)
C_d = ranknet.cost(query, self.model, self.sigma)
dnum = (C_u[0] - C_d[0]) / dw / 2.0
# compare results
#self.assertAlmostEqual(grad[jw], dnum, 5, "Run %d: %e %e " % (j, grad[jw], dnum))
print j, jw, grad[jw], dnum
def xtest_training_1(self):
# trainsg on a single query
nepoch = 10000 # number of training epochs
rate = 0.1 # learning rate
nprint = 1000 # print frequency
for je in xrange(nepoch):
# compute current cost and estimations
C = ranknet.cost(self.query, self.model, self.sigma)
if je % nprint == 0:
print je, C[0], C[1], C[2]
print "w:", self.model.getw()
# compute gradients
g = ranknet.gradient(self.query, self.model, self.sigma)
# update weights
w = la.vsum( self.model.getw(), la.sax(-rate, g) )
self.model.setw(w)
def test_cost(self):
C = ranknet.cost(self.query, self.model, self.sigma)
print C
def train(data, opts, net, writefcn):
"""
Batch training of ranknet model using RProp
"""
# random permutation of data
perm = range(len(data))
random.shuffle(perm)
jvalid = perm[0:opts.nvalid] # validation data index
jtrain = perm[opts.nvalid:] # training data index
nfail = 0 # current number of validation fails
mincost = 1.e+100 # current known minimal validation cost
wbest = net.getw() # weights for minimal validation error
# write out options and initial network
writefcn(str(opts) + "\n")
writefcn(str(net) + "\n")
# initialize RProp working memory
rpropmem = ([1.e-5] * len(net.getw()), [1] * len(net.getw()))
print "Start batch training, number of queries: ", len(data)
print str(opts)
print str(net)
# training iterations
for je in xrange(opts.maxepoch):
# validation cost
vcost = 0.0
for j in jvalid:
c = ranknet.cost(data[j], net, opts.sigma)
vcost += c[0]
# update best estimates
if vcost < mincost:
mincost = vcost
wbest = net.getw()
else:
nfail += 1
# check stopping criteria
if opts.maxfail > 0 and nfail >= opts.maxfail:
break
# reset accumulators
tcost = 0.0 # training cost
G = [0] * len(net.getw()) # accumulated gradient
# batch training
for jt in jtrain:
# take next training query
query = data[jt]
# compute cost
c = ranknet.cost(query, net, opts.sigma)
tcost += c[0]
# compute gradient
g = ranknet.gradient(query, net, opts.sigma)
# update batch gradient
G = la.vsum(G, g)
# print to screen
print je, nfail, tcost, vcost, net.getw()
# write out
sw = str(net.getw()).replace("[", "").replace("]", "").replace(",", "")
writefcn("%d %d %e %e %s\n" % (je, nfail, tcost, vcost, sw))
# RProp update steps
rpropmem = opt.rprop(G, rpropmem)
steps = rpropmem[0]
signs = rpropmem[1]
# update network weights
w = la.vsum(net.getw(), la.vmul(steps, la.sax(-1, signs)))
net.setw(w)
# training complete
print "Training stopped"
print "Final cost: ", mincost
print "Final weights: ", wbest
# return updated model
net.setw(wbest)
return net
def train(data, opts, net, writefcn):
"""
Batch training of ranknet model using RProp
"""
# random permutation of data
perm = range(len(data))
random.shuffle(perm)
jvalid = perm[0:opts.nvalid] # validation data index
jtrain = perm[opts.nvalid:] # training data index
nfail = 0 # current number of validation fails
mincost = 1.e+100 # current known minimal validation cost
wbest = net.getw() # weights for minimal validation error
# write out options and initial network
writefcn(str(opts) + "\n")
writefcn(str(net) + "\n")
# initialize RProp working memory
rpropmem = ( [1.e-5] * len(net.getw()), [ 1 ] * len(net.getw()) )
print "Start batch training, number of queries: ", len(data)
print str(opts)
print str(net)
# training iterations
for je in xrange(opts.maxepoch):
# validation cost
vcost = 0.0
for j in jvalid:
c = ranknet.cost(data[j], net, opts.sigma)
vcost += c[0]
# update best estimates
if vcost < mincost:
mincost = vcost
wbest = net.getw()
else:
nfail += 1
# check stopping criteria
if opts.maxfail > 0 and nfail >= opts.maxfail:
break
# reset accumulators
tcost = 0.0 # training cost
G = [0] * len(net.getw()) # accumulated gradient
# batch training
for jt in jtrain:
# take next training query
query = data[jt]
# compute cost
c = ranknet.cost(query, net, opts.sigma)
tcost += c[0]
# compute gradient
g = ranknet.gradient(query, net, opts.sigma)
# update batch gradient
G = la.vsum(G, g)
# print to screen
print je, nfail, tcost, vcost, net.getw()
# write out
sw = str(net.getw()).replace("[", "").replace("]", "").replace(",", "")
writefcn("%d %d %e %e %s\n" % (je, nfail, tcost, vcost, sw))
# RProp update steps
rpropmem = opt.rprop(G, rpropmem)
steps = rpropmem[0]
signs = rpropmem[1]
# update network weights
w = la.vsum(net.getw(), la.vmul(steps, la.sax(-1, signs)))
net.setw(w)
# training complete
print "Training stopped"
print "Final cost: ", mincost
print "Final weights: ", wbest
# return updated model
net.setw(wbest)
return net
def train(data, opts, net, writefcn):
"""
Stochastic gradient training of ranknet model
"""
# random permutation of data
perm = range(len(data))
random.shuffle(perm)
jvalid = perm[0:opts.nvalid] # validation data index
jtrain = perm[opts.nvalid:] # training data index
nfail = 0 # current number of validation fails
mincost = 1.e+100 # current known minimal validation cost
wbest = net.getw() # weights for minimal validation error
print "Start stochastic gradient descent training, number of queries: ", len(
data)
print str(opts)
print str(net)
# stochastic gradient training
for je in xrange(opts.maxepoch):
# validation
if je % opts.nepval == 0:
# compute validation cost
C = 0.0
for j in jvalid:
c = ranknet.cost(data[j], net, opts.sigma)
C += c[0]
# update best estimates
if C < mincost:
mincost = C
wbest = net.getw()
else:
nfail += 1
# check stopping criteria
if opts.maxfail > 0 and nfail >= opts.maxfail:
break
# print
print je, nfail, C, net.getw()
# write to report file
sw = str(net.getw()).replace("[", "").replace("]",
"").replace(",", "")
writefcn("%d %d %e %s\n" % (je, nfail, C, sw))
# next training query
jq = je % len(jtrain)
query = data[jtrain[jq]]
# compute gradients
g = ranknet.gradient(query, net, opts.sigma)
# update weights
w = la.vsum(net.getw(), la.sax(-opts.rate, g))
net.setw(w)
print "Training stopped"
print "Final cost: ", mincost
print "Final weights: ", wbest
# return updated model
net.setw(wbest)
return net