def predict_contacts(model, x, y, use_cuda):
b = len(x)
x, order = pack_sequences(x)
x = PackedSequence(Variable(x.data), x.batch_sizes)
z = model(x) # embed the sequences
z = unpack_sequences(z, order)
logits = []
y_list = []
for i in range(b):
zi = z[i]
lp = model.predict(zi.unsqueeze(0)).view(-1)
yi = y[i].view(-1)
if use_cuda:
yi = yi.cuda()
mask = (yi < 0)
lp = lp[~mask]
yi = yi[~mask]
logits.append(lp)
y_list.append(yi)
return logits, y_list
def eval_similarity(model, test_iterator, use_cuda):
y = []
logits = []
for x0, x1, y_mb in test_iterator:
if use_cuda:
y_mb = y_mb.cuda()
y.append(y_mb.long())
b = len(x0)
x = x0 + x1
x, order = pack_sequences(x)
x = PackedSequence(Variable(x.data), x.batch_sizes)
z = model(x) # embed the sequences
z = unpack_sequences(z, order)
z0 = z[:b]
z1 = z[b:]
for i in range(b):
z_a = z0[i]
z_b = z1[i]
logits.append(model.score(z_a, z_b))
y = torch.cat(y, 0)
logits = torch.stack(logits, 0)
p = F.sigmoid(logits).data
ones = p.new(p.size(0), 1).zero_() + 1
p_ge = torch.cat([ones, p], 1)
p_lt = torch.cat([1 - p, ones], 1)
p = p_ge * p_lt
p = p / p.sum(1, keepdim=True) # make sure p is normalized
loss = F.cross_entropy(p, y).item()
_, y_hard = torch.max(p, 1)
levels = torch.arange(5).to(p.device)
y_hat = torch.sum(p * levels, 1)
accuracy = torch.mean((y == y_hard).float()).item()
mse = torch.mean((y.float() - y_hat)**2).item()
y = y.cpu().numpy()
y_hat = y_hat.cpu().numpy()
r, _ = pearsonr(y_hat, y)
rho, _ = spearmanr(y_hat, y)
return loss, accuracy, mse, r, rho
def predict_minibatch(model, x, use_cuda):
b = len(x)
x,order = pack_sequences(x)
x = PackedSequence(x.data, x.batch_sizes)
z = model(x) # embed the sequences
z = unpack_sequences(z, order)
logits = []
for i in range(b):
zi = z[i]
lp = model.predict(zi.unsqueeze(0)).view(zi.size(0), zi.size(0))
logits.append(lp)
return logits
def similarity_grad(model, x0, x1, y, use_cuda, weight=0.5):
if use_cuda:
y = y.cuda()
y = Variable(y)
b = len(x0)
x = x0 + x1
x, order = pack_sequences(x)
x = PackedSequence(Variable(x.data), x.batch_sizes)
z = model(x) # embed the sequences
z = unpack_sequences(z, order)
z0 = z[:b]
z1 = z[b:]
logits = []
for i in range(b):
z_a = z0[i]
z_b = z1[i]
logits.append(model.score(z_a, z_b))
logits = torch.stack(logits, 0)
loss = F.binary_cross_entropy_with_logits(logits, y.float())
# backprop weighted loss
w_loss = loss * weight
w_loss.backward()
# calculate minibatch performance metrics
with torch.no_grad():
p = F.sigmoid(logits)
ones = p.new(b, 1).zero_() + 1
p_ge = torch.cat([ones, p], 1)
p_lt = torch.cat([1 - p, ones], 1)
p = p_ge * p_lt
p = p / p.sum(1, keepdim=True) # make sure p is normalized
_, y_hard = torch.max(p, 1)
levels = torch.arange(5).to(p.device)
y_hat = torch.sum(p * levels, 1)
y = torch.sum(y.data, 1)
loss = F.cross_entropy(
p, y).item() # calculate cross entropy loss from p vector
correct = torch.sum((y == y_hard).float()).item()
mse = torch.mean((y.float() - y_hat)**2).item()
return loss, correct, mse, b
def __call__(self, x):
c = [torch.from_numpy(x_).long() for x_ in x]
c, order = pack_sequences(c)
if self.use_cuda:
c = c.cuda()
if self.full_features:
z = featurize(c, self.lm_embed, self.lstm_stack, self.proj)
else:
z = self.model(c) # embed the sequences
z = unpack_sequences(z, order)
return z
def __call__(self, x, y):
n = len(x)
c = [torch.from_numpy(x_).long()
for x_ in x] + [torch.from_numpy(y_).long() for y_ in y]
c, order = pack_sequences(c)
if self.use_cuda:
c = c.cuda()
with torch.no_grad():
z = self.model(c) # embed the sequences
z = unpack_sequences(z, order)
scores = np.zeros(n)
if self.mode == 'align':
for i in range(n):
z_x = z[i]
z_y = z[i + n]
logits = self.model.score(z_x, z_y)
p = F.sigmoid(logits).cpu()
p_ge = torch.ones(p.size(0) + 1)
p_ge[1:] = p
p_lt = torch.ones(p.size(0) + 1)
p_lt[:-1] = 1 - p
p = p_ge * p_lt
p = p / p.sum() # make sure p is normalized
levels = torch.arange(5).float()
scores[i] = torch.sum(p * levels).item()
elif self.mode == 'coarse':
z_x = z[:n]
z_y = z[n:]
z_x = torch.stack([z.mean(0) for z in z_x], 0)
z_y = torch.stack([z.mean(0) for z in z_y], 0)
scores[:] = -torch.sum(torch.abs(z_x - z_y), 1).cpu().numpy()
return scores
def contacts_grad(model, x, y, use_cuda, weight=0.5):
b = len(x)
x, order = pack_sequences(x)
x = PackedSequence(Variable(x.data), x.batch_sizes)
z = model(x) # embed the sequences
z = unpack_sequences(z, order)
logits = []
for i in range(b):
zi = z[i]
lp = model.predict(zi.unsqueeze(0)).view(-1)
logits.append(lp)
logits = torch.cat(logits, 0)
y = torch.cat([yi.view(-1) for yi in y])
if use_cuda:
y = y.cuda()
mask = (y < 0)
logits = logits[~mask]
y = Variable(y[~mask])
b = y.size(0)
loss = F.binary_cross_entropy_with_logits(logits, y)
# backprop weighted loss
w_loss = loss * weight
w_loss.backward()
# calculate the recall and precision
with torch.no_grad():
p_hat = F.sigmoid(logits)
tp = torch.sum(p_hat * y).item()
gp = y.sum().item()
pp = p_hat.sum().item()
return loss.item(), tp, gp, pp, b
def embed_sequences(model: nn.Module,
seqs: List[str],
batch_size: int = 32,
gpu: bool = False) -> np.ndarray:
"""embed_sequences.
Args:
model (nn.Module): model
seqs (List[str]): seqs
batch_size (int): batch_size
gpu (bool): gpu
Returns:
np.ndarray:
"""
encoded_seqs = encode_seqs(seqs)
identity_collate = lambda x: x
loader = torch.utils.data.DataLoader(encoded_seqs,
batch_size=batch_size,
collate_fn=identity_collate)
if gpu:
model = model.cuda()
return_seqs = []
with torch.no_grad():
model.eval()
#
for batch in tqdm(loader):
X, order = utils.pack_sequences(batch)
if gpu:
X = X.cuda()
out = model.forward(X)
unpacked = utils.unpack_sequences(out, order)
return_seqs.extend([i.detach().cpu().numpy() for i in unpacked])
return return_seqs