def corrupt(batch, n):
"""
Corrupts the negatives of a batch of triples (in place).
Randomly corrupts either heads or tails
:param batch_size:
:param n: nr of nodes in the graph
:return:
"""
bs, ns, _ = batch.size()
# new entities to insert
corruptions = torch.randint(size=(bs * ns, ),
low=0,
high=n,
dtype=torch.long,
device=d(batch))
# boolean mask for entries to corrupt
mask = torch.bernoulli(
torch.empty(size=(bs, ns, 1), dtype=torch.float,
device=d(batch)).fill_(0.5)).to(torch.bool)
zeros = torch.zeros(size=(bs, ns, 1), dtype=torch.bool, device=d(batch))
mask = torch.cat([mask, zeros, ~mask], dim=2)
batch[mask] = corruptions
def torch_set_diag(self, t, filler=0.):
"""
Pytorch fix of fill_diagonal for batches.
We assume the input tensor has shape (bs,n,n).
"""
t_return = t.cpu().clone().detach().to(d())
ind = np.diag_indices(t.shape[-1])
t_return[:, ind[0],
ind[1]] = torch.ones(t.shape[-1], device=d()) * filler
return t_return
def reparameterize(self, mean, logvar):
"""
Reparametrization trick.
"""
self.mean = mean
self.logvar = logvar
eps = torch.normal(torch.zeros_like(mean), std=1.).to(d())
return eps * torch.exp(logvar * .5) + mean
def hungarian_batch(self, Xs):
X = Xs.clone().cpu().numpy()
# Make it a cost matrix
X = np.ones_like(X) - X
for i in range(X.shape[0]):
row_ind, col_ind = linear_sum_assignment(X[i])
M = np.zeros(X[i].shape, dtype=float)
M[row_ind, col_ind] = 1.
X[i] = M
X = torch.tensor(X, device=d())
return X
def reparameterize(self, mean_logvar):
"""
Reparametrization trick.
"""
self.mean = mean = mean_logvar[:, :, :self.z_dim]
self.logvar = logvar = mean_logvar[:, :, self.z_dim:]
if self.var:
eps = torch.normal(torch.zeros_like(mean), std=1.).to(d())
else:
eps = 1.
return eps * torch.exp(logvar * .5) + mean
def affinity(self, A, A_hat, E, E_hat, F, F_hat):
"""
Let's make some dimensionalities clear first (w/o batch dim):
A: n,n
E: n,n,d_e
F: n,d_n
A_hat: k,k
E_hat: k,k,d_e
F_hat: k,d_n
In an ideal world the target dimension n and the predictions dim k are the same.
The other two dimensions are node and edge attributes. All matrixes come in batches, which is the first dimension.
Now we are going to try to solve this with matrix multiplication, for-loops are a no-go.
My first shot would be to implement this formula without respecting the constrains:
S((i, j),(a, b)) = (E'(i,j,:)E_hat(a,b,:))A(i,j)A_hat(a,b)A_hat(a,a)A_hat(b,b) [i != j ∧ a != b] + (F'(i,:)F_hat(a,:))A_hat(a,a) [i == j ∧ a == b]
And later mask the constrained entries with zeros.
"""
n = A.shape[1]
self.n = n
k = A_hat.shape[1]
self.k = k
bs = A.shape[0] # bs stands for batch size, just to clarify.
self.bs = bs
A_hat_diag = (torch.diagonal(A_hat, dim1=-2, dim2=-1)).unsqueeze(-1)
E_norm = torch.norm(
E, p=1, dim=-1, keepdim=True
) # Division by the norm since our model can have multiple edge attributes vs. one-hot
E_norm[E_norm == 0.] = 1. # Otherwise we get nans
E_ijab = torch_batch_dot(E / E_norm, E_hat, 3,
3) # We aim for shape (batch_s,n,n,k,k).
A_ab = A_hat * self.torch_set_diag(
torch_batch_dot_v2(A_hat_diag, A_hat_diag, -1, -1, (bs, k, k)))
A_ijab = torch_batch_dot_v2((self.torch_set_diag(A)).unsqueeze(-1),
A_ab.unsqueeze(-1), -1, -1,
(bs, n, n, k, k))
A_aa = torch.bmm(torch.ones((bs, n, 1), device=d()),
torch.transpose(A_hat_diag, 1, 2))
F_ia = torch.matmul(F, torch.transpose(F_hat, 1, 2))
# S = E_ijab * A_ijab + self.set_diag_nnkk(F_ia * A_aa, bs, n, k)
# assert torch.isnan(S).any() == False
S_iaia = F_ia * A_aa
S_iajb = E_ijab * A_ijab #+ self.set_diag_nnkk(S_iaia, bs, n, k)
return (S_iajb, S_iaia)
def max_pool(self, S, n_iterations: int = 11):
"""
The famous Cho max-pooling in matrix multiplication style.
Xs: X_star meaning X in continuos space.
"""
S_iajb, S_iaia = S
Xs = torch.ones_like(S_iaia, device=d())
self.Xs = Xs
for n in range(n_iterations):
Xs = Xs * S_iaia + torch.sum(
torch.max(S_iajb * Xs.unsqueeze(1).unsqueeze(1), -1,
out=None)[0], -1)
Xs = torch.where(torch.isnan(Xs), torch.zeros_like(Xs), Xs)
Xs_norm = torch.norm(Xs, p='fro', dim=[-2, -1])
Xs = (Xs / Xs_norm.unsqueeze(-1).unsqueeze(-1))
Xs = torch.where(torch.isnan(Xs), torch.zeros_like(Xs), Xs)
return Xs
def corrupt_one(batch, candidates, target):
"""
Corrupts the negatives of a batch of triples (in place).
Corrupts either only head or only tails
:param batch_size:
:param n: nr of nodes in the graph
:param target: 0 for head, 1 for predicate, 2 for tail
:return:
"""
bs, ns, _ = batch.size()
# new entities to insert
#corruptions = torch.randint(size=(bs * ns,),low=0, high=n, dtype=torch.long, device=d(batch))
corruptions = torch.tensor(random.choices(candidates, k=bs * ns),
dtype=torch.long,
device=d(batch)).view(bs, ns)
batch[:, :, target] = corruptions
def train_lp_vembed(n_e,
n_r,
train,
test,
alltriples,
beta: int,
epochs: int,
batch_size: int,
result_dir: str,
test_batch: int = 5,
eval_int: int = 3):
"""
Source: pbloem/embed
"""
# Fix some hyperparameters
repeats = 1
sched = True
check_simple = True
negative_rate = [10, 0, 10] # No neg sampling in VAE
limit_negatives = True
loss_fn = 'bce'
reciprocal = True
reg_exp = 2
reg_eweight = None
reg_rweight = None
eval_size = None
corrupt_global = True
bias = True
patience = 1
result_file = result_dir + '/lp_log.txt'
lr = 1e-4
k = 11
model = VLinkPredictor(torch.tensor(list(alltriples)),
n_e,
n_r,
embedding=512,
decoder='distmult',
edropout=None,
rdropout=None,
init=0.85,
biases=False,
init_method='uniform',
init_parms=(-1.0, 1.0),
reciprocal=reciprocal)
wandb.watch(model)
tbw = SummaryWriter(log_dir=result_dir)
dev = 'cuda' if torch.cuda.is_available() else 'cpu'
test_mrrs = []
truedict = truedicts(alltriples)
train = torch.tensor(train).to(d())
test = torch.tensor(test).to(d())
subjects = torch.tensor(list({s
for s, _, _ in train}),
dtype=torch.long,
device=d())
predicates = torch.tensor(list({p
for _, p, _ in train}),
dtype=torch.long,
device=d())
objects = torch.tensor(list({o
for _, _, o in train}),
dtype=torch.long,
device=d())
ccandidates = (subjects, predicates, objects)
with open(result_file, 'w') as f:
print(n_e, 'nodes', file=f)
print(n_r, 'relations', file=f)
print(train.size(0), 'training triples', file=f)
print(test.size(0), 'test triples', file=f)
print(train.size(0) + test.size(0), 'total triples', file=f)
for r in tqdm.trange(repeats) if repeats > 1 else range(repeats):
if torch.cuda.is_available():
prt('Using CUDA.')
model.cuda()
# if arg.opt == 'adam':
# opt = torch.optim.Adam(model.parameters(), lr=arg.lr)
# elif arg.opt == 'adamw':
# opt = torch.optim.AdamW(model.parameters(), lr=arg.lr)
# elif arg.opt == 'adagrad':
optimizer = torch.optim.Adagrad(model.parameters(),
lr=0.15953749294870845)
# elif arg.opt == 'sgd':
# opt = torch.optim.SGD(model.parameters(), lr=arg.lr, nesterov=True, momentum=arg.momentum)
# else:
# raise Exception()
sched = torch.optim.lr_scheduler.ReduceLROnPlateau(patience=patience, optimizer=optimizer, mode='max', factor=0.95, threshold=0.0001) \
if sched else None
#-- defaults taken from libkge
# nr of negatives sampled
weight = None
seen = 0
for e in range(epochs):
seeni, sumloss = 0, 0.0
tforward = tbackward = 0
rforward = rbackward = 0
tprep = tloss = 0
tic()
for fr in trange(0, train.size(0), batch_size):
to = min(train.size(0), fr + batch_size)
model.train(True)
optimizer.zero_grad()
positives = train[fr:to].to(d())
for ctarget in [0, 1,
2]: # which part of the triple to corrupt
ng = negative_rate[ctarget]
if ng > 0:
with torch.no_grad():
bs, _ = positives.size()
tic()
if limit_negatives:
cand = ccandidates[ctarget]
mx = cand.size(0)
idx = torch.empty(bs,
ng,
dtype=torch.long,
device=d()).random_(0, mx)
corruptions = cand[idx]
else:
mx = n_r if ctarget == 1 else n_e
corruptions = torch.empty(bs,
ng,
dtype=torch.long,
device=d()).random_(
0, mx)
tprep += toc()
s, p, o = positives[:, 0:
1], positives[:, 1:
2], positives[:,
2:
3]
if ctarget == 0:
s = torch.cat([s, corruptions], dim=1)
if ctarget == 1:
p = torch.cat([p, corruptions], dim=1)
if ctarget == 2:
o = torch.cat([o, corruptions], dim=1)
# -- NB: two of the index vectors s, p o are now size (bs, 1) and the other is (bs, ng+1)
# We will let the model broadcast these to give us a score tensor of (bs, ng+1)
# In most cases we can optimize the decoder to broadcast late for better speed.
if loss_fn == 'bce':
labels = torch.cat([
torch.ones(bs, 1, device=d()),
torch.zeros(bs, ng, device=d())
],
dim=1)
elif loss_fn == 'ce':
labels = torch.zeros(bs,
dtype=torch.long,
device=d())
# -- CE loss treats the problem as a multiclass classification problem: for a positive triple,
# together with its k corruptions, identify which is the true triple. This is always triple 0.
# (It may seem like the model could easily cheat by always choosing triple 0, but the score
# function is order equivariant, so it can't choose by ordering.)
recip = None if not reciprocal else (
'head' if ctarget == 0 else 'tail')
# -- We use the tail relations if the target is the relation (usually p-corruption is not used)
tic()
out = model.forward(s, p, o)
tforward += toc()
assert out.size() == (
bs, ng + 1), f'{out.size()=} {(bs, ng + 1)=}'
tic()
if loss_fn == 'bce':
out = F.sigmoid(out)
recon_loss = F.binary_cross_entropy_with_logits(
out, labels, weight=weight)
wandb.log({"recon_loss": recon_loss})
reg_loss = kl_divergence(model.encoder.mean,
model.encoder.logvar)
loss = torch.mean(beta * reg_loss - recon_loss)
# elif loss_fn == 'ce':
# loss = F.cross_entropy(out, labels)
wandb.log({"loss": loss.item()})
assert not torch.isnan(loss), 'Loss has become NaN'
sumloss += float(loss.item())
seen += bs
seeni += bs
tloss += toc()
tic()
loss.backward()
tbackward += toc()
# No step yet, we accumulate the gradients over all corruptions.
# -- this causes problems with modules like batchnorm, so be careful when porting.
tic()
regloss = None
# if reg_eweight is not None:
# regloss = model.penalty(which='entities', p=reg_exp, rweight=reg_eweight)
# if reg_rweight is not None:
# regloss = model.penalty(which='relations', p=reg_exp, rweight=reg_rweight)
# rforward += toc()
# tic()
# if regloss is not None:
# sumloss += float(regloss.item())
# regloss.backward()
# wandb.log({"regloss": regloss})
# rbackward += toc()
optimizer.step()
if e == 0:
print(
'\n pred: forward {tforward:.4}, backward {tbackward:.4}')
print(f' prep {tprep:.4}, loss {tloss:.4}')
print(f' total: {toc():.4}')
# -- NB: these numbers will not be accurate for GPU runs unless CUDA_LAUNCH_BLOCKING is set to 1
# Evaluate
if ((e + 1) % eval_int == 0) or e == epochs - 1:
with torch.no_grad():
model.train(False)
if eval_size is None:
testsub = test
else:
testsub = test[random.sample(range(test.size(0)),
k=eval_size)]
mrr, hits, ranks = util.eval(model=model,
valset=testsub,
truedicts=truedict,
n=n_e,
batch_size=test_batch,
verbose=True)
# if check_simple: # double-check using a separate, slower implementation
# mrrs, hitss, rankss = util.eval_simple(
# model=model, valset=testsub, alltriples=alltriples, n=n_e, verbose=True)
# assert ranks == rankss
# assert mrr == mrrs
with open(result_file, 'a+') as f:
print(
f'epoch {e}: MRR {mrr:.4}\t hits@1 {hits[0]:.4}\t hits@3 {hits[1]:.4}\t hits@10 {hits[2]:.4}',
file=f)
print(
f'epoch {e}: MRR {mrr:.4}\t hits@1 {hits[0]:.4}\t hits@3 {hits[1]:.4}\t hits@10 {hits[2]:.4}'
)
tbw.add_scalar('biases/mrr', mrr, e)
tbw.add_scalar('biases/h@1', hits[0], e)
tbw.add_scalar('biases/h@3', hits[1], e)
tbw.add_scalar('biases/h@10', hits[2], e)
wandb.log({
"mrr": mrr,
"h@1": hits[0],
"h@3": hits[1],
"h@10": hits[2]
})
if sched is not None:
sched.step(mrr) # reduce lr if mrr stalls
test_mrrs.append(mrr)
with open(result_file, 'a+') as f:
print('training finished.', file=f)
print('training finished.')
temrrs = torch.tensor(test_mrrs)
with open(result_file, 'a+') as f:
print(
f'mean test MRR {temrrs.mean():.3} ({temrrs.std():.3}) \t{test_mrrs}',
file=f)
print(
f'mean test MRR {temrrs.mean():.3} ({temrrs.std():.3}) \t{test_mrrs}'
)
