def setUp(self):
self.v1 = Variable("1")
self.v2 = Variable("2")
self.l1 = Literal(self.v1)
self.l2 = Literal(self.v2)
self.l3 = Literal(self.v1, False)
self.a = Assignment(1, self.v2, True, Clause(self.l2))
self.v2.assign(self.a)
def forward(self, v, kn, ko, s, hidden):
if hidden is None:
h = self.h_initial.view(self.num_layers, self.know_length,
self.seq_hidden_size)
attn_h = self.h_initial
length = Variable(torch.FloatTensor([0.]))
beta = None
else:
h, vs, hs = hidden
# calculate beta weights of seqs using dot product
beta = torch.mm(vs, v.view(-1, 1)).view(-1)
beta, idx = beta.topk(min(len(beta), self.k), sorted=False)
beta = nn.functional.softmax(beta.view(1, -1), dim=-1)
length = Variable(torch.FloatTensor([beta.size()[1]]))
hs = hs.view(-1, self.know_length * self.seq_hidden_size)
attn_h = torch.mm(beta, torch.index_select(hs, 0, idx)).view(-1)
# calculate alpha weights of knowledges using dot product
alpha = torch.mm(self.knowledge_memory, kn.view(-1, 1)).view(-1)
alpha = nn.functional.softmax(alpha.view(1, -1), dim=-1)
hkp = torch.mm(alpha,
attn_h.view(self.know_length,
self.seq_hidden_size)).view(-1)
pred_v = torch.cat([v, hkp]).view(1, -1)
predict_score = self.score_layer(pred_v)
# seq states update
if self.score_mode == 'concat':
x = v
else:
x = torch.cat([
v * (s >= 0.5).type_as(v).expand_as(v),
v * (s < 0.5).type_as(v).expand_as(v)
])
x = torch.cat([x, s])
# print(x.size())
# print(torch.ones(self.know_length,1).size())
# print(x.view(1, -1).size())
# print(x.type())
# xk = torch.mm(torch.ones(self.know_length, 1), x.view(1, -1))
xk = x.view(1, -1).expand(self.know_length, -1)
xk = alpha.view(-1, 1) * xk
# xk = ko.float().view(-1, 1) * xk
# xk = torch.mm(alpha, xk).view(-1)
_, h = self.rnn(xk.unsqueeze(0), h)
return predict_score.view(1), h, beta
def interpret(self, lines):
regx = re.compile(r"(-?)(\d+)")
self.known = dict()
clauses = []
c = []
m = re.match("p (?:cnf )?(.+) (.+)\s*\n", lines)
checksum = m.groups() if m else None
lines = re.sub("[cp].*?\n", "", lines)
for literal in re.split(r"[\s\n]+", lines):
m = re.match(regx, literal)
if m:
varname = m.group(2)
if varname != "0":
var = self.known.get(varname, Variable(varname))
var.count += 1
self.known[varname] = var
l = Literal(var, not bool(m.group(1)))
c.append(l)
else:
addition = Clause(*c)
addition.link()
clauses.append(addition)
c = []
if checksum:
assert (len(self.known) == int(checksum[0]))
assert (len(clauses) == int(checksum[1]))
return clauses
def test_assignment(self):
v = Variable("1")
s = AssignmentStack()
a = Assignment(1, v, True, Clause(Literal(v)))
s.push(a)
p = s.pop()
self.assertEqual(p.var, v)
def forward(self, v, s, hidden):
if hidden is None:
h = self.initial_h.view(self.num_layers, 1, self.seq_hidden_size)
attn_h = self.initial_h
length = Variable(torch.FloatTensor([0.]))
else:
h, vs, hs = hidden
# print(h)
# print('start')
# print(vs.size())
# print(v.size())
# print(v.view(-1,1).size())
# print(torch.mm(vs,v.view(-1,1)).size())
# print(hs)
# calculate alpha using dot product
alpha = torch.mm(vs, v.view(-1, 1)).view(-1)
# print(alpha.size())
# print('end')
# print(alpha.size())
alpha, idx = alpha.topk(min(len(alpha), self.k), sorted=False)
alpha = nn.functional.softmax(alpha.view(1, -1), dim=-1)
length = Variable(torch.FloatTensor([alpha.size()[1]]))
# flatten each h
hs = hs.view(-1, self.num_layers * self.seq_hidden_size)
attn_h = torch.mm(alpha, torch.index_select(hs, 0, idx)).view(-1)
if self.with_last:
pred_v = torch.cat([v, attn_h, h.view(-1), length]).view(1, -1)
else:
pred_v = torch.cat([v, attn_h]).view(1, -1)
score = self.score(pred_v)
if self.score_mode == 'concat':
x = v
else:
x = torch.cat([
v * (s >= 0.5).type_as(v).expand_as(v),
v * (s < 0.5).type_as(v).expand_as(v)
])
x = torch.cat([x, s])
_, h = self.rnn(x.view(1, 1, -1), h)
return score, h
def PickBranchingVariable(self):
m = Variable('tmp')
for v in self.vars:
if not v.isAssigned() and v.count>=m.count:
m = v
pol = False
logging.info("Branching dl: %s -> %s = %s",self.dl+1,v.name,pol)
return m,pol
def __init__(self):
self.resWins = Variable()
self.spyWins = Variable()
self.votesRes = Variable()
self.votesSpy = Variable()
self.spyVoted = Variable()
self.spySelected = Variable()
self.selections = Variable()
def forward(self, v, kn, ko, s, h, beta=None):
if h is None:
h = self.h_initial.view(self.num_layers, self.know_length,
self.seq_hidden_size)
length = Variable(torch.FloatTensor([0.]))
# calculate alpha weights of knowledges using dot product
# print(self.knowledge_memory.size())
# print(kn.view(-1, 1))
if beta is None:
alpha = torch.mm(self.knowledge_memory, kn.view(-1, 1)).view(-1)
beta = nn.functional.softmax(alpha.view(1, -1), dim=-1)
# print(beta.argmax(1))
# print(alpha.size())
# print(h.view(self.know_length, self.seq_hidden_size).size())
# print(h.type())
# predict score at time t
hkp = torch.mm(beta, h.view(self.know_length,
self.seq_hidden_size)).view(-1)
# print(hkp.size())
pred_v = torch.cat([v, hkp]).view(1, -1)
# print(pred_v.size())
predict_score = self.score_layer(pred_v)
# seq states update
if self.score_mode == 'concat':
x = v
else:
x = torch.cat([
v * (s >= 0.5).type_as(v).expand_as(v),
v * (s < 0.5).type_as(v).expand_as(v)
])
x = torch.cat([x, s])
# print(x.size())
# print(torch.ones(self.know_length,1).size())
# print(x.view(1, -1).size())
# print(x.type())
# xk = torch.mm(torch.ones(self.know_length, 1), x.view(1, -1))
xk = x.view(1, -1).expand(self.know_length, -1)
xk = beta.view(-1, 1) * xk
# xk = ko.float().view(-1, 1) * xk
# print(xk.size())
# print(alpha.size())
# xk = torch.mm(alpha, xk).view(-1)
# thresh, idx = alpha.topk(5)
# alpha = (alpha >= thresh[0, 4]).float()
# xk = alpha.view(-1, 1) * xk
# xk = Variable(torch.zeros_like(x)).expand(self.know_length, -1)
_, h = self.rnn(xk.unsqueeze(0), h)
return predict_score.view(1), h
class TestLiteral(unittest.TestCase):
def setUp(self):
self.v1 = Variable("1")
self.v2 = Variable("2")
self.l1 = Literal(self.v1)
self.l2 = Literal(self.v2)
self.l3 = Literal(self.v1, False)
self.a = Assignment(1, self.v2, True, Clause(self.l2))
self.v2.assign(self.a)
def test_eq(self):
t = Literal(self.v1)
self.assertEqual(t, self.l1)
self.assertNotEqual(self.l1, self.l3)
self.assertNotEqual(self.l1, self.l2)
def test_neg(self):
self.assertEqual(self.l3, -self.l1)
def test_value(self):
self.assertTrue(self.l2.value())
def _estimateProb(self, dic, fullkey):
""" If there is not enough info on the full scenario,
gradually loosen the constraints to get at least a good prior. """
prior_params = {1: (50, 8), 2: (30, 4), 3: (20, 2), 4: (10, 1)}
v = Variable()
v.total = 0.5 + dic[fullkey].total # uncertainty bias
v.samples = 1 + dic[fullkey].samples
for skipunits in range(1, len(fullkey) + 1):
bound, weight = prior_params[skipunits]
if v.samples > bound:
break
for nkey in maskSome(list(fullkey), skipunits):
if tuple(nkey) not in dic:
continue
tmp = dic[tuple(nkey)]
if tmp.samples == 0:
continue
weight = min(weight, tmp.samples / float(skipunits))
v.samples += weight
v.total += weight * tmp.total / float(tmp.samples)
return v.total / float(v.samples)
def _estimateProb(self, dic, fullkey):
""" If there is not enough info on the full scenario,
gradually loosen the constraints to get at least a good prior. """
prior_params = {1: (50, 8),
2: (30, 4),
3: (20, 2),
4: (10, 1)}
v = Variable()
v.total = 0.5 + dic[fullkey].total # uncertainty bias
v.samples = 1 + dic[fullkey].samples
for skipunits in range(1,len(fullkey)+1):
bound, weight = prior_params[skipunits]
if v.samples > bound:
break
for nkey in maskSome(list(fullkey), skipunits):
if tuple(nkey) not in dic:
continue
tmp = dic[tuple(nkey)]
if tmp.samples == 0:
continue
weight = min(weight, tmp.samples/float(skipunits))
v.samples += weight
v.total += weight*tmp.total/float(tmp.samples)
return v.total/float(v.samples)
def predict(model, args):
try:
torch.set_grad_enabled(False)
except AttributeError:
pass
logging.info('model: %s, setup: %s' %
(type(model).__name__, str(model.args)))
logging.info('loading dataset')
if args.snapshot is None:
epoch = load_last_snapshot(model, args.workspace)
else:
epoch = args.snapshot
load_snapshot(model, args.workspace, epoch)
logging.info('loaded model at epoch %s', str(epoch))
to_categorical = Categorical('</s>')
to_categorical.load_dict(model.words)
trans = to_categorical(Words(':', null='</s>'))
while True:
# loop over inputs
try:
line = input()
except EOFError:
logging.info('bye')
break
try:
obj = json.loads(line, encoding='utf-8')
ref_seq = obj['ref']
pred_seq = obj['pred']
except (json.decoder.JSONDecodeError, KeyError):
print('[]')
continue
h = None
for i, item in enumerate(ref_seq):
x = trans.apply(None, item['fea'])
x = Variable(torch.LongTensor(x), volatile=True)
score = Variable(torch.FloatTensor([item['t']]), volatile=True)
t = Variable(torch.FloatTensor([item['s']]), volatile=True)
_, h = model(x, score, t, h)
pred_scores = []
for i, item in enumerate(pred_seq):
x = trans.apply(None, item['fea'])
x = Variable(torch.LongTensor(x), volatile=True)
score = Variable(torch.FloatTensor([0.]), volatile=True)
t = Variable(torch.FloatTensor([item['t']]), volatile=True)
s, _ = model(x, score, t, h)
pred_scores.append(s.cpu().data[0][0])
print(pred_scores)
def __init__(self,
topic_size,
seq_hidden_size,
k,
score_mode,
num_layers=1):
super(AttnSeqTimeDecayModel, self).__init__()
self.topic_size = topic_size
self.seq_hidden_size = seq_hidden_size
self.num_layers = num_layers
self.score_mode = score_mode
if self.score_mode == 'concat':
self.rnn = nn.GRU(topic_size + 1, seq_hidden_size, num_layers)
else:
self.rnn = nn.GRU(topic_size * 2 + 1, seq_hidden_size, num_layers)
self.score = nn.Linear(topic_size + seq_hidden_size, 1)
self.k = k
self.initial_h = Variable(torch.zeros(self.num_layers *
self.seq_hidden_size),
requires_grad=True)
class TestVariable(unittest.TestCase):
def setUp(self):
self.v = Variable("1")
self.a = Assignment(1, self.v, True, Clause(Literal(self.v)))
def test_repr(self):
v = Variable("somename")
self.assertEqual(str(v), "somename")
def test_value(self):
#v starts unassigned
self.assertEqual(self.v.value(), None)
#associate v
self.v.assign(self.a)
#v should be assigned true
self.assertTrue(self.v.value())
#disassociate v
self.v.unassign()
#v should be unassigned
self.assertFalse(self.v.isAssigned())
def test_repr(self):
v = Variable("somename")
self.assertEqual(str(v), "somename")
def setUp(self):
self.v = Variable("1")
self.a = Assignment(1, self.v, True, Clause(Literal(self.v)))
def __init__(self, topic_size, k):
super(AttnModel, self).__init__()
self.user_emb_size = topic_size
self.k = k
self.initial_guess = Variable(torch.zeros(1), requires_grad=True)
def total(self):
return Variable(self.resWins.total + self.spyWins.total,
self.resWins.samples + self.spyWins.samples)
def default_hidden(self, batch_size):
return Variable(torch.zeros(2, batch_size, self.emb_size)), \
Variable(torch.zeros(self.num_layers - 1,
batch_size, self.emb_size)) \
if self.num_layers > 1 else None
def default_hidden(self):
return Variable(torch.zeros(self.num_layers, 1, self.seq_hidden_size))
class TestClause(unittest.TestCase):
def setUp(self):
self.v1 = Variable("1")
self.v2 = Variable("2")
self.v3 = Variable("3")
self.l1 = Literal(self.v1)
self.l2 = Literal(self.v2)
self.l3 = Literal(self.v2, False)
self.a1 = Assignment(1, self.v2, True, Clause(self.l2))
self.v2.assign(self.a1)
def test_init(self):
c = Clause()
self.assertFalse(hasattr(c, 'revA') or hasattr(c, 'revB'))
c = Clause(Literal(self.v1, True))
self.assertFalse(hasattr(c, 'revA') or hasattr(c, 'revB'))
def test_eq(self):
c1 = Clause(self.l1)
c2 = Clause(self.l2)
c3 = Clause(self.l1)
self.assertEqual(c1, c3)
self.assertNotEqual(c1, c2)
def test_contains(self):
c1 = Clause(self.l1)
c2 = Clause(self.l2)
self.assertTrue(self.l1 in c1)
self.assertFalse(self.l1 in c2)
def test_status(self):
#empty clause
c = Clause()
self.assertEqual(c.status(), "UNSAT")
#single literal
# unassigned
c = Clause(self.l1)
self.assertEqual(c.status(), "UNIT")
# sat
c = Clause(self.l2)
self.assertEqual(c.status(), "SAT")
# unsat
c = Clause(self.l3)
self.assertEqual(c.status(), "UNSAT")
#two-literal
# unit
c = Clause(self.l1, self.l3)
self.assertEqual(c.status(), "UNIT")
# sat
c = Clause(self.l2, self.l3)
self.assertEqual(c.status(), "SAT")
c = Clause(self.l1, self.l2)
self.assertEqual(c.status(), "SAT")
# unsat
c = Clause(self.l3, self.l3)
self.assertEqual(c.status(), "UNSAT")
#multi-literal
# c = Clause(self.l3, self.l1, self.l3)
# self.assertEqual(c.status(),"UNIT")
c = Clause(self.l3, self.l1, self.l1)
self.l3.occurrence_link(c)
self.assertEqual(c.status(), "UNRESOLVED")
c = Clause(self.l3, self.l1, self.l2)
c.link()
self.assertEqual(c.status(), "SAT")
def test_resolution(self):
c1 = Clause(self.l1, self.l2)
c2 = Clause(self.l1, self.l3)
c3 = c1.resolve(c2)
self.assertEqual(c3, Clause(self.l1))
step_score = [[], []]
print('Model initialized, starting training...')
for step in range(n_steps):
# sample from agent
seqs, agent_likelihood, entropy = Agent.sample(batch_size)
# Remove duplicates, ie only consider unique seqs
unique_idx = unique(seqs)
seqs = seqs[unique_idx]
agent_likelihood = agent_likelihood[unique_idx]
entropy = entropy[unique_idx]
# Get prior likelihood and score
prior_likelihood, _ = Prior.likelihood(Variable(seqs))
smiles = seq_to_smiles(seqs, voc)
score = scoring_function(smiles)
# Calculate augmented likelihood
augmented_likelihood = prior_likelihood + sigma * Variable(score)
loss = torch.pow((augmented_likelihood - agent_likelihood), 2)
# Experience replay
# First example
if experience_replay and len(experience) > 4:
exp_seqs, exp_score, exp_prior_likelihood = experience.sample(4)
exp_agent_likelihood, exp_entropy = Agent.likelihood(exp_seqs.long())
exp_augmented_likelihood = exp_prior_likelihood + sigma * exp_score
exp_loss = torch.pow((Variable(exp_augmented_likelihood) - exp_agent_likelihood), 2)
loss = torch.cat((loss, exp_loss), 0)
def main(options=None):
num_bars = options.num_bars
cantus_file = options.cantus_file
solution_file = options.solution_file
sa_file = options.sa_file
test_dir = options.test_dir
testing = options.testing
dfs_csv = options.dfs_csv
sa_csv = options.sa_csv
song_list = ['mary', 'ariana', 'shootingstar']
arc_consistency = options.arc_consistency
extra_harmonic = options.extra_harmonic
if options.preset_song in song_list:
print 'options.preset_song is {}'.format(options.preset_song)
elif options.preset_song != '':
print 'preset song not found, proceeding with default'
csp = Csp()
cp = [] # list of counterpoint variables
cf = [] # list of __ variables
binary = [] # list of Constraint objects
if options.preset_song == song_list[0]:
note_list = [64, 62, 60, 62, 64, 64, 64, 62, 62, 62, 64, 67, 67, 64, 62, \
60, 62, 64, 64, 64, 64, 62, 62, 64, 62, 60]
num_bars = len(note_list)
elif options.preset_song == song_list[1]:
num_bars = 8
note_list = [69, 67, 66, 67, 66, 64, 66, 64]
elif options.preset_song == song_list[2]:
num_bars = 10
note_list = [71, 71, 72, 67, 64, 71, 71, 72, 67, 64]
elif testing:
print('Generating a cantus firmus over ' + str(num_bars) + ' bars')
note_list = []
for i in range(num_bars):
note_list.append(random.choice(NOTE_RANGE))
else:
note_list = [57,60,59,57] # default
for i in range(num_bars):
cp.append(Variable('cp' + str(i)))
csp.addToVariables(cp[i])
binary.append(Constraint())
binary[i].setVariable(cp[i])
cf.append(Variable('cf' + str(i)))
csp.addToVariables(cf[i])
for i in range(len(note_list)):
note = Note(note_list[i])
cf[i].addToDomain(note)
print('Cantus firmus sequential note pitches: ')
print(note_list)
for i in range(num_bars):
cp_note_list = range(30, 100)
if i != (num_bars - 2):
map(lambda x: cp[i].addToDomain(Note(x)), cp_note_list)
else:
cp_note_list = range(60, 70)
map(lambda x: cp[i].addToDomain(Note(x)), cp_note_list)
# binary constraints, p. 109 of Ovans
for i in range(1, num_bars):
L = Link()
L.setNode(binary[i-1])
L.setLabel(Note.melodic)
cp[i].addToNeighbors(L)
L = Link()
L.setNode(binary[i])
L.setLabel(Note.melodic)
cp[i-1].addToNeighbors(L)
"""# extra constraint!!!!
if extra_harmonic and i != num_bars - 1:
L = Link()
L.setNode(binary[i+1])
L.setLabel(Note.harmonic)
cp[i].addToNeighbors(L)"""
L = Link()
L.setNode(binary[0])
L.setLabel(Note.perfectCfHarmonic)
cf[0].addToNeighbors(L)
# harmonic constraints for each cantus firmus note
for i in range(1, num_bars-2):
L = Link()
L.setNode(binary[i])
L.setLabel(Note.harmonic)
cf[i].addToNeighbors(L)
# perfect harmonic constraints in last two bars
for i in range(num_bars-2, num_bars):
L = Link()
L.setNode(binary[i])
L.setLabel(Note.perfectHarmonic)
cf[i].addToNeighbors(L)
test_csp = copy.deepcopy(csp)
if arc_consistency:
arc_start = timeit.default_timer()
if csp.AC3():
arc_stop = timeit.default_timer()
print 'Made initial arcs consistent in {} seconds.'.format(arc_stop - arc_start)
print 'Arc consistent - looking for a solution with DFS'
sol_start = timeit.default_timer()
csp.backtracking_search()
sol_stop = timeit.default_timer()
print('Trying simulated annealing...')
sim_start = timeit.default_timer()
test_csp.simAnnealing()
sim_stop = timeit.default_timer()
print 'Completed simulated annealing after {} seconds.'.format(sim_stop - sim_start)
print 'Simulated annealing returns with cost {}, after {} iterations.'.format(test_csp.getCost(test_csp.vars), test_csp.iters)
else:
arc_stop = timeit.default_timer()
print 'Failed to make initial arcs consistent after {} seconds.'.format(arc_stop - arc_start)
print('Not consistent')
return None
else:
sol_start = timeit.default_timer()
csp.backtracking_search()
sol_stop = timeit.default_timer()
print('Attempt to find a DFS solution finished after {} seconds.'.format(sol_stop - sol_start))
print('Trying simulated annealing...')
sim_start = timeit.default_timer()
test_csp.simAnnealing()
sim_stop = timeit.default_timer()
print 'Completed simulated annealing after {} seconds.'.format(sim_stop - sim_start)
print 'Simulated annealing returns with cost {}, after {} iterations.'.format(test_csp.getCost(test_csp.vars), test_csp.iters)
print('Looking for a solution with DFS')
if csp.one_sol is not None:
print('Found a solution with arc consistency! Expanded {} nodes with {} backtracks'.format(csp.getNodes(), csp.getBts()))
# Log stats in csv file for testing
if testing:
dfs_trial_info = '{},{},{},{}\n'.format(num_bars, csp.getNodes(), csp.getBts(), sol_stop - sol_start)
with open(dfs_csv, 'a+') as f:
f.write(dfs_trial_info)
f.closed
else:
print('No solution found')
return None
write_solution(csp.one_sol, num_bars=num_bars, solution_file=test_dir + '/' + solution_file, random_length=options.random)
if test_csp.getCost(test_csp.vars) == 0:
write_solution(test_csp.vars, num_bars=num_bars,solution_file=test_dir + '/' + sa_file, random_length=options.random)
# Log stats in csv file for testing
if testing:
sim_trial_info = '{},{},{},{}\n'.format(num_bars, test_csp.getCost(test_csp.vars), test_csp.iters, sim_stop - sim_start)
with open(sa_csv, 'a+') as f:
f.write(sim_trial_info)
f.closed
else:
print('Simulated annealing failed, not writing to output')
if testing:
return csp
def test(model, args):
try:
torch.set_grad_enabled(False)
except AttributeError:
pass
logging.info('model: %s, setup: %s' %
(type(model).__name__, str(model.args)))
logging.info('loading dataset')
data = get_dataset(args.dataset)
data.random_level = args.random_level
if not args.dataset.endswith('test'):
if args.split_method == 'user':
_, data = data.split_user(args.frac)
testsets = [('user_split', data, {})]
elif args.split_method == 'future':
_, data = data.split_future(args.frac)
testsets = [('future_split', data, {})]
elif args.split_method == 'old':
trainset, _, _, _ = data.split()
data = trainset.get_seq()
train, user, exam, new = data.split()
train = train.get_seq()
user = user.get_seq()
exam = exam.get_seq()
new = new.get_seq()
testsets = zip(['user', 'exam', 'new'], [user, exam, new],
[{}, train, user])
else:
if args.ref_set:
ref = get_dataset(args.ref_set)
ref.random_level = args.random_level
testsets = [(args.dataset.split('/')[-1], data.get_seq(),
ref.get_seq())]
else:
testsets = [('student', data.get_seq(), {})]
else:
testsets = [('school', data.get_seq(), {})]
if type(model).__name__.startswith('DK'):
topic_dic = {}
kcat = Categorical(one_hot=True)
kcat.load_dict(open('data/know_list.txt').read().split('\n'))
for line in open('data/id_know.txt'):
uuid, know = line.strip().split(' ')
know = know.split(',')
topic_dic[uuid] = \
torch.LongTensor(kcat.apply(None, know)) \
.max(0)[0] \
.type(torch.LongTensor)
zero = [0] * len(kcat.apply(None, '<NULL>'))
else:
topics = get_topics(args.dataset, model.words)
if args.snapshot is None:
epoch = load_last_snapshot(model, args.workspace)
else:
epoch = args.snapshot
load_snapshot(model, args.workspace, epoch)
logging.info('loaded model at epoch %s', str(epoch))
if use_cuda:
model.cuda()
for testset, data, ref_data in testsets:
logging.info('testing on: %s', testset)
f = open_result(args.workspace, testset, epoch)
then = time.time()
total_mse = 0
total_mae = 0
total_acc = 0
total_seq_cnt = 0
users = list(data)
random.shuffle(users)
seq_cnt = len(users)
MSE = torch.nn.MSELoss()
MAE = torch.nn.L1Loss()
for user in users[:5000]:
seq = data[user]
if user in ref_data:
ref_seq = ref_data[user]
else:
ref_seq = []
seq2 = []
seen = set()
for item in ref_seq:
if item.topic in seen:
continue
seen.add(item.topic)
seq2.append(item)
ref_seq = seq2
seq2 = []
for item in seq:
if item.topic in seen:
continue
seen.add(item.topic)
seq2.append(item)
seq = seq2
ref_len = len(ref_seq)
seq = ref_seq + seq
length = len(seq)
if ref_len < args.ref_len:
length = length + ref_len - args.ref_len
ref_len = args.ref_len
if length < 1:
continue
total_seq_cnt += 1
mse = 0
mae = 0
acc = 0
pred_scores = Variable(torch.zeros(len(seq)))
s = None
h = None
for i, item in enumerate(seq):
if args.test_on_last:
x = topics.get(seq[-1].topic).content
x = Variable(torch.LongTensor(x), volatile=True)
score = Variable(torch.FloatTensor([round(seq[-1].score)]),
volatile=True)
t = Variable(torch.FloatTensor([seq[-1].time]),
volatile=True)
s, _ = model(x, score, t, h)
s_last = torch.clamp(s, 0, 1)
if type(model).__name__.startswith('DK'):
if item.topic in topic_dic:
x = topic_dic[item.topic]
else:
x = zero
else:
x = topics.get(item.topic).content
x = Variable(torch.LongTensor(x))
score = Variable(torch.FloatTensor([round(item.score)]),
volatile=True)
t = Variable(torch.FloatTensor([item.time]), volatile=True)
if args.test_as_seq and i > ref_len and ref_len > 0:
s, h = model(x, s.view(1), t, h)
else:
if ref_len > 0 and i > ref_len and not args.test_on_one:
s, _ = model(x, score, t, h)
else:
s, h = model(x, score, t, h)
if args.loss == 'cross_entropy':
s = F.sigmoid(s)
else:
s = torch.clamp(s, 0, 1)
if args.test_on_last:
pred_scores[i] = s_last
else:
pred_scores[i] = s
if i < ref_len:
continue
mse += MSE(s, score)
m = MAE(s, score).data[0]
mae += m
acc += m < 0.5
print_seq(seq,
pred_scores.data.cpu().numpy(), ref_len, f,
args.test_on_last)
mse /= length
mae /= length
acc /= length
total_mse += mse.data[0]
total_mae += mae
total_acc += acc
if total_seq_cnt % args.print_every != 0 and \
total_seq_cnt != seq_cnt:
continue
now = time.time()
duration = (now - then) / 60
logging.info(
'[%d/%d] (%.2f seqs/min) '
'rmse %.6f, mae %.6f, acc %.6f' %
(total_seq_cnt, seq_cnt,
((total_seq_cnt - 1) % args.print_every + 1) / duration,
math.sqrt(total_mse / total_seq_cnt),
total_mae / total_seq_cnt, total_acc / total_seq_cnt))
then = now
f.close()
def testseq(model, args):
try:
torch.set_grad_enabled(False)
except AttributeError:
pass
logging.info('model: %s, setup: %s' %
(type(model).__name__, str(model.args)))
logging.info('loading dataset')
data = get_dataset(args.dataset)
data.random_level = args.random_level
if not args.dataset.endswith('test'):
if args.split_method == 'user':
_, data = data.split_user(args.frac)
testsets = [('user_split', data, {})]
elif args.split_method == 'future':
_, data = data.split_future(args.frac)
testsets = [('future_split', data, {})]
elif args.split_method == 'old':
trainset, _, _, _ = data.split()
data = trainset.get_seq()
train, user, exam, new = data.split()
train = train.get_seq()
user = user.get_seq()
exam = exam.get_seq()
new = new.get_seq()
testsets = zip(['user', 'exam', 'new'], [user, exam, new],
[{}, train, user])
else:
if args.ref_set:
ref = get_dataset(args.ref_set)
ref.random_level = args.random_level
testsets = [(args.dataset.split('/')[-1], data.get_seq(),
ref.get_seq())]
else:
testsets = [('student', data.get_seq(), {})]
else:
testsets = [('school', data.get_seq(), {})]
if args.input_knowledge:
logging.info('loading knowledge concepts')
topic_dic = {}
kcat = Categorical(one_hot=True)
kcat.load_dict(open(model.args['knows']).read().split('\n'))
know = 'data/id_firstknow.txt' if 'first' in model.args['knows'] \
else 'data/id_know.txt'
for line in open(know):
uuid, know = line.strip().split(' ')
know = know.split(',')
topic_dic[uuid] = torch.LongTensor(kcat.apply(None,
know)).max(0)[0]
zero = [0] * len(kcat.apply(None, '<NULL>'))
if args.input_text:
logging.info('loading exercise texts')
topics = get_topics(args.dataset, model.words)
if args.snapshot is None:
epoch = load_last_snapshot(model, args.workspace)
else:
epoch = args.snapshot
load_snapshot(model, args.workspace, epoch)
logging.info('loaded model at epoch %s', str(epoch))
if use_cuda:
model.cuda()
for testset, data, ref_data in testsets:
logging.info('testing on: %s', testset)
f = open_result(args.workspace, testset, epoch)
then = time.time()
total_mse = 0
total_mae = 0
total_acc = 0
total_seq_cnt = 0
users = list(data)
random.shuffle(users)
seq_cnt = len(users)
MSE = torch.nn.MSELoss()
MAE = torch.nn.L1Loss()
for user in users[:5000]:
total_seq_cnt += 1
seq = data[user]
if user in ref_data:
ref_seq = ref_data[user]
else:
ref_seq = []
length = len(seq)
ref_len = len(ref_seq)
seq = ref_seq + seq
if ref_len < args.ref_len:
length = length + ref_len - args.ref_len
ref_len = args.ref_len
if length < 1:
ref_len = ref_len + length - 1
length = 1
mse = 0
mae = 0
acc = 0
# seq2 = []
# seen = set()
# for item in seq:
# if item.topic in seen:
# continue
# seen.add(item.topic)
# seq2.append(item)
# seq = seq2
# length = len(seq) - ref_len
pred_scores = Variable(torch.zeros(len(seq)))
s = None
h = None
for i, item in enumerate(seq):
# get last record for testing and current record for updating
if args.input_knowledge:
if item.topic in topic_dic:
knowledge = topic_dic[item.topic]
knowledge_last = topic_dic[seq[-1].topic]
else:
knowledge = zero
knowledge_last = zero
knowledge = Variable(torch.LongTensor(knowledge))
knowledge_last = Variable(torch.LongTensor(knowledge_last),
volatile=True)
if args.input_text:
text = topics.get(item.topic).content
text = Variable(torch.LongTensor(text))
text_last = topics.get(seq[-1].topic).content
text_last = Variable(torch.LongTensor(text_last),
volatile=True)
score = Variable(torch.FloatTensor([item.score]),
volatile=True)
score_last = Variable(torch.FloatTensor([round(seq[-1].score)
]),
volatile=True)
item_time = Variable(torch.FloatTensor([item.time]),
volatile=True)
time_last = Variable(torch.FloatTensor([seq[-1].time]),
volatile=True)
# test last score of each seq for seq figure
if type(model).__name__.startswith('DK'):
s, _ = model(knowledge_last, score_last, time_last, h)
elif type(model).__name__.startswith('RA'):
s, _ = model(text_last, score_last, time_last, h)
elif type(model).__name__.startswith('EK'):
s, _ = model(text_last, knowledge_last, score_last,
time_last, h)
s_last = torch.clamp(s, 0, 1)
# update student state h until the fit process reaches trainset
if ref_len > 0 and i > ref_len:
if type(model).__name__.startswith('DK'):
s, _ = model(knowledge, score, item_time, h)
elif type(model).__name__.startswith('RA'):
s, _ = model(text, score, item_time, h)
elif type(model).__name__.startswith('EK'):
s, _ = model(text, knowledge, score, item_time, h)
else:
if type(model).__name__.startswith('DK'):
s, h = model(knowledge, score, item_time, h)
elif type(model).__name__.startswith('RA'):
s, h = model(text, score, item_time, h)
elif type(model).__name__.startswith('EK'):
s, h = model(text, knowledge, score, item_time, h)
pred_scores[i] = s_last
if args.loss == 'cross_entropy':
s = F.sigmoid(s)
else:
s = torch.clamp(s, 0, 1)
if i < ref_len:
continue
mse += MSE(s, score)
m = MAE(s, score).data[0]
mae += m
acc += m < 0.5
print_seq(seq, pred_scores.data.cpu().numpy(), ref_len, f, True)
mse /= length
mae /= length
acc = float(acc) / length
total_mse += mse.data[0]
total_mae += mae
total_acc += acc
if total_seq_cnt % args.print_every != 0 and total_seq_cnt != seq_cnt:
continue
now = time.time()
duration = (now - then) / 60
logging.info(
'[%d/%d] (%.2f seqs/min) '
'rmse %.6f, mae %.6f, acc %.6f' %
(total_seq_cnt, seq_cnt,
((total_seq_cnt - 1) % args.print_every + 1) / duration,
math.sqrt(total_mse / total_seq_cnt),
total_mae / total_seq_cnt, total_acc / total_seq_cnt))
then = now
f.close()
def trainn(model, args):
logging.info('model: %s, setup: %s' %
(type(model).__name__, str(model.args)))
logging.info('loading dataset')
data = get_dataset(args.dataset)
data.random_level = args.random_level
if args.split_method == 'user':
data, _ = data.split_user(args.frac)
elif args.split_method == 'future':
data, _ = data.split_future(args.frac)
elif args.split_method == 'old':
data, _, _, _ = data.split()
data = data.get_seq()
if args.input_knowledge:
logging.info('loading knowledge concepts')
topic_dic = {}
kcat = Categorical(one_hot=True)
kcat.load_dict(open(model.args['knows']).read().split('\n'))
know = 'data/id_firstknow.txt' if 'first' in model.args['knows'] \
else 'data/id_know.txt'
for line in open(know):
uuid, know = line.strip().split(' ')
know = know.split(',')
topic_dic[uuid] = torch.LongTensor(kcat.apply(None,
know)).max(0)[0]
zero = [0] * len(kcat.apply(None, '<NULL>'))
if args.input_text:
logging.info('loading exercise texts')
topics = get_topics(args.dataset, model.words)
optimizer = torch.optim.Adam(model.parameters())
start_epoch = load_last_snapshot(model, args.workspace)
if use_cuda:
model.cuda()
for epoch in range(start_epoch, args.epochs):
logging.info('epoch {}:'.format(epoch))
then = time.time()
total_loss = 0
total_mae = 0
total_acc = 0
total_seq_cnt = 0
users = list(data)
random.shuffle(users)
seq_cnt = len(users)
MSE = torch.nn.MSELoss()
MAE = torch.nn.L1Loss()
for user in users:
total_seq_cnt += 1
seq = data[user]
seq_length = len(seq)
optimizer.zero_grad()
loss = 0
mae = 0
acc = 0
h = None
for i, item in enumerate(seq):
# score = round(item.score)
if args.input_knowledge:
if item.topic in topic_dic:
knowledge = topic_dic[item.topic]
else:
knowledge = zero
# knowledge = torch.LongTensor(knowledge).view(-1).type(torch.FloatTensor)
# one_index = torch.nonzero(knowledge).view(-1)
# expand_vec = torch.zeros(knowledge.size()).view(-1)
# expand_vec[one_index] = score
# cks = torch.cat([knowledge, expand_vec]).view(1, -1)
knowledge = Variable(torch.LongTensor(knowledge))
# cks = Variable(cks)
if args.input_text:
text = topics.get(item.topic).content
text = Variable(torch.LongTensor(text))
score = Variable(torch.FloatTensor([item.score]))
item_time = Variable(torch.FloatTensor([item.time]))
if type(model).__name__.startswith('DK'):
s, h = model(knowledge, score, item_time, h)
elif type(model).__name__.startswith('RA'):
s, h = model(text, score, item_time, h)
elif type(model).__name__.startswith('EK'):
s, h = model(text, knowledge, score, item_time, h)
s = s[0]
if args.loss == 'cross_entropy':
loss += F.binary_cross_entropy_with_logits(
s, score.view_as(s))
m = MAE(F.sigmoid(s), score).data[0]
else:
loss += MSE(s, score)
m = MAE(s, score).data[0]
mae += m
acc += m < 0.5
loss /= seq_length
mae /= seq_length
acc = float(acc) / seq_length
total_loss += loss.data[0]
total_mae += mae
total_acc += acc
loss.backward()
optimizer.step()
if total_seq_cnt % args.save_every == 0:
save_snapshot(model, args.workspace,
'%d.%d' % (epoch, total_seq_cnt))
if total_seq_cnt % args.print_every != 0 and total_seq_cnt != seq_cnt:
continue
now = time.time()
duration = (now - then) / 60
logging.info(
'[%d:%d/%d] (%.2f seqs/min) loss %.6f, mae %.6f, acc %.6f' %
(epoch, total_seq_cnt, seq_cnt,
((total_seq_cnt - 1) % args.print_every + 1) / duration,
total_loss / total_seq_cnt, total_mae / total_seq_cnt,
total_acc / total_seq_cnt))
then = now
save_snapshot(model, args.workspace, epoch + 1)
def default_hidden(self):
return Variable(torch.zeros(1, 1, self.topic_size))
def __init__(self):
# Variables that caputres how well you lead in context of pass/fail mission
self.leadTeam = Variable()
self.leadSucMission = Variable()
self.leadFailMission = Variable()
# Variables that capture how well you voted in context of pass/fail mission
self.voted = Variable()
self.votedYes = Variable()
self.votedNo = Variable()
self.votedYesPass = Variable()
self.votedYesFail = Variable()
self.votedNoPass = Variable()
self.votedNoFail = Variable()
# Variables that capture how well you perform when selected in context of pass/ fail mission
self.selected = Variable()
self.selectedPass = Variable()
self.selectedFail = Variable()
# Variables that capture how much sabotaging/passing you were involved in
self.sabotages = Variable()
self.passeds = Variable()
# Label True/False either spy or not
self.Label = None
def train(model, args):
logging.info('args: %s' % str(args))
logging.info('model: %s, setup: %s' %
(type(model).__name__, str(model.args)))
logging.info('loading dataset')
data = get_dataset(args.dataset)
data.random_level = args.random_level
if args.split_method == 'user':
data, _ = data.split_user(args.frac)
elif args.split_method == 'future':
data, _ = data.split_future(args.frac)
elif args.split_method == 'old':
data, _, _, _ = data.split()
data = data.get_seq()
if type(model).__name__.startswith('DK'):
topic_dic = {}
kcat = Categorical(one_hot=True)
kcat.load_dict(open('data/know_list.txt').read().split('\n'))
for line in open('data/id_know.txt'):
uuid, know = line.strip().split(' ')
know = know.split(',')
topic_dic[uuid] = \
torch.LongTensor(kcat.apply(None, know)) \
.max(0)[0] \
.type(torch.LongTensor)
zero = [0] * len(kcat.apply(None, '<NULL>'))
else:
topics = get_topics(args.dataset, model.words)
optimizer = torch.optim.Adam(model.parameters())
start_epoch = load_last_snapshot(model, args.workspace)
if use_cuda:
model.cuda()
for epoch in range(start_epoch, args.epochs):
logging.info(('epoch {}:'.format(epoch)))
then = time.time()
total_loss = 0
total_mae = 0
total_acc = 0
total_seq_cnt = 0
users = list(data)
random.shuffle(users)
seq_cnt = len(users)
MSE = torch.nn.MSELoss()
MAE = torch.nn.L1Loss()
for user in users:
total_seq_cnt += 1
seq = data[user]
length = len(seq)
optimizer.zero_grad()
loss = 0
mae = 0
acc = 0
h = None
for i, item in enumerate(seq):
if type(model).__name__.startswith('DK'):
if item.topic in topic_dic:
x = topic_dic[item.topic]
else:
x = zero
else:
x = topics.get(item.topic).content
x = Variable(torch.LongTensor(x))
# print(x.size())
score = Variable(torch.FloatTensor([round(item.score)]))
t = Variable(torch.FloatTensor([item.time]))
s, h = model(x, score, t, h)
if args.loss == 'cross_entropy':
loss += F.binary_cross_entropy_with_logits(
s, score.view_as(s))
m = MAE(F.sigmoid(s), score).data[0]
else:
loss += MSE(s, score)
m = MAE(s, score).data[0]
mae += m
acc += m < 0.5
loss /= length
mae /= length
acc /= length
total_loss += loss.data[0]
total_mae += mae
total_acc += acc
loss.backward()
optimizer.step()
if total_seq_cnt % args.save_every == 0:
save_snapshot(model, args.workspace,
'%d.%d' % (epoch, total_seq_cnt))
if total_seq_cnt % args.print_every != 0 and \
total_seq_cnt != seq_cnt:
continue
now = time.time()
duration = (now - then) / 60
logging.info(
'[%d:%d/%d] (%.2f seqs/min) '
'loss %.6f, mae %.6f, acc %.6f' %
(epoch, total_seq_cnt, seq_cnt,
((total_seq_cnt - 1) % args.print_every + 1) / duration,
total_loss / total_seq_cnt, total_mae / total_seq_cnt,
total_acc / total_seq_cnt))
then = now
save_snapshot(model, args.workspace, epoch + 1)
