def predict_proba(self, inp, gpu=False, **kwargs):
"""
Compute the probability assigned by the model to an input sequence.
In the future this should use pack_padded_sequence to run prediction
on multiple input sentences at once.
Parameters:
-----------
inp: list of ints representing the input sequence
gpu: bool, whether to use the gpu
kwargs: other model parameters
Returns:
--------
float representing the log probability of the input sequence
"""
if self.training:
logging.warn("Generating in training modus!")
if isinstance(inp, list):
inp = torch.LongTensor(inp)
if gpu:
inp = inp.cuda()
outs, *_ = self(Variable(inp, volatile=True), **kwargs)
log_probs = u.select_cols(outs[:-1], inp[1:])
return log_probs.sum().data[0] / len(log_probs)
def predict_proba(self, inp, gpu=False, **kwargs):
if self.training:
logging.warn("Generating in training modus!")
inp_vec = Variable(torch.LongTensor([inp]), volatile=True)
if gpu:
inp_vec.cuda()
outs, hidden, _ = self(inp_vec, **kwargs)
outs = u.select_cols(outs, inp).sum()
return outs.data[0] / len(inp)
def predict_proba(self, inp, gpu=False, **kwargs):
if self.training:
logging.warn("Generating in training modus!")
inp_vec = Variable(torch.LongTensor([inp]), volatile=True)
if gpu:
inp_vec.cuda()
outs, _, _ = self(inp_vec, **kwargs)
log_probs = u.select_cols(outs[:-1], inp[1:])
return log_probs.sum().data[0] / len(log_probs)
def sample(self, temperature=1., seed_texts=None, max_seq_len=25,
batch_size=10, bos=False, ignore_eos=False, **kwargs):
scores, prev, hidden = self._seed(
seed_texts, batch_size, 'sample', bos, temperature=temperature)
batch = prev.size(1)
hyps = [prev.squeeze().data.tolist()]
finished = np.array([False] * batch)
for _ in range(max_seq_len):
outs, hidden, _ = self.model(prev, hidden=hidden, **kwargs)
prev = outs.div(temperature).exp_().multinomial().t()
score = u.select_cols(outs.data.cpu(), prev.squeeze().data.cpu())
hyps.append(prev.squeeze().data.tolist())
if self.eos is not None and not ignore_eos:
mask = (prev.squeeze().data == self.eos).cpu().numpy() == 1
finished[mask] = True
if all(finished == True): # nopep8
break
# 0-mask scores for finished batch examples
score[torch.ByteTensor(finished.tolist())] = 0
scores += score
return scores.tolist(), list(zip(*hyps))
def sample(self,
temperature=1.,
seed_texts=None,
max_seq_len=25,
batch_size=10,
ignore_eos=False,
bos=False,
eos=False,
**kwargs):
"""
Generate a sequence multinomially sampling from the output
distribution at each generation step. The output distribution
can be tweaked by the input parameter `temperature`.
"""
prev, hidden = self._seed(seed_texts,
batch_size,
bos,
eos,
temperature=temperature)
batch_size = prev.size(1) # not equal to input if seed_texts
hyps, scores = [], 0
mask = torch.zeros(batch_size).long() + 1
for _ in range(max_seq_len):
outs, hidden, _ = self.model(prev, hidden=hidden, **kwargs)
prev = outs.div_(temperature).exp().multinomial(1).t()
score = u.select_cols(outs.data.cpu(), prev.squeeze().data.cpu())
hyps.append(prev.squeeze().data.tolist())
if self.eos is not None and not ignore_eos:
mask = mask * (prev.squeeze().data.cpu() != self.eos).long()
if mask.sum() == 0:
break
# 0-mask scores for finished batch examples
score[mask == 0] = 0
scores += score
return scores.tolist(), list(zip(*hyps))
def sample(out):
prev = out.div(temperature).exp_().multinomial().t()
score = u.select_cols(out.data.cpu(), prev.squeeze().data.cpu())
return prev, score
def sample(out):
prev = out.div(temperature).exp_().multinomial().t()
score = u.select_cols(out.data.cpu(), prev.squeeze().data.cpu())
return prev, score