def test_prob_space(self):
a = torch.Tensor([[5.0, 5.0], [6.0, 2.0], [2.0, 6.0]])
b = torch.Tensor([[0.0, 0.0], [2.0, 6.0], [6.0, 2.0]])
expected = torch.Tensor([[0.5, 0.5], [0.5, 0.5], [0.5, 0.5]])
npt.assert_allclose(
pytorch_utils.average_tensors([a, b], norm_fn=F.softmax), expected
)
def test_mean(self):
a = torch.Tensor([[0.0, 2.0, 5.0], [5.0, -5.0, 6.0]])
b = torch.Tensor([[4.0, 2.0, -1.0], [5.0, 10.0, 6.0]])
c = torch.Tensor([[-1.0, 2.0, 5.0], [2.0, 10.0, 6.0]])
expected = torch.Tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
npt.assert_allclose(pytorch_utils.average_tensors([a, b, c]), expected)
third = 1.0 / 3.0
npt.assert_allclose(
pytorch_utils.average_tensors([a, b, c], weights=[third, third, third]),
expected,
)
npt.assert_allclose(
pytorch_utils.average_tensors([a, b, c], weights=[1, 1, 1]), 3 * expected
)
npt.assert_allclose(
pytorch_utils.average_tensors([a, b, c], weights=[1, 0, 0]), a
)
def forward(
self,
unprojected_outs,
src_tokens=None,
input_tokens=None,
possible_translation_tokens=None,
select_single=None,
):
weights = self.compute_weights(unprojected_outs, select_single)
weights = [weights[:, :, i:i + 1] for i in range(self.n_systems)]
averaged_unprojected = average_tensors(unprojected_outs,
weights=weights)
return self.output_projections[0](averaged_unprojected, src_tokens,
input_tokens,
possible_translation_tokens)
def forward(
self,
unprojected_outs,
src_tokens=None,
input_tokens=None,
possible_translation_tokens=None,
select_single=None,
):
return self.output_projection(
average_tensors(unprojected_outs)
if select_single is None else unprojected_outs[select_single],
src_tokens,
input_tokens,
possible_translation_tokens,
)
def forward(
self,
input_tokens,
encoder_out,
incremental_state=None,
possible_translation_tokens=None,
):
if self.unfreeze_single:
self.unfreeze_idx = (self.unfreeze_idx + 1) % self.unfreeze_mod
if self.separate_training:
unfreeze_combi_strat = len(self.decoders) == self.unfreeze_idx
for p in self.combi_strat.parameters():
p.requires_grad = unfreeze_combi_strat
self.freeze_decoders(self.unfreeze_idx)
if incremental_state is None:
incremental_state = {
decoder_id: None
for decoder_id in range(len(self.decoders))
}
decoder_outs = []
decoder_contexts = self._get_contexts(encoder_out)
for decoder_id, decoder in enumerate(self.decoders):
if decoder_id not in incremental_state:
incremental_state[decoder_id] = {}
decoder_outs.append(
decoder.forward_unprojected(
input_tokens,
decoder_contexts[decoder_id],
incremental_state=incremental_state[decoder_id],
))
mean_attn_scores = average_tensors([
decoder_outs[decoder_id][1]
for decoder_id in self.attentive_decoder_ids
])
select_single = None
if self.separate_training and not unfreeze_combi_strat:
select_single = self.unfreeze_idx
logits, possible_translation_tokens = self.combi_strat(
[x for x, _ in decoder_outs],
src_tokens=encoder_out[4],
input_tokens=input_tokens if self.training else None,
possible_translation_tokens=possible_translation_tokens,
select_single=select_single,
)
return logits, mean_attn_scores, possible_translation_tokens
def forward(
self,
unprojected_outs,
src_tokens=None,
input_tokens=None,
possible_translation_tokens=None,
select_single=None,
):
assert possible_translation_tokens is None
weights = self.compute_weights(unprojected_outs, select_single)
weights = [weights[:, :, i:i + 1] for i in range(self.n_systems)]
logits = [
p(o)[0] for p, o in zip(self.output_projections, unprojected_outs)
]
avg = average_tensors(logits, weights=weights, norm_fn=self.norm_fn)
if self.to_log:
avg.log_()
return avg, None
def forward(
self,
unprojected_outs,
src_tokens=None,
input_tokens=None,
possible_translation_tokens=None,
select_single=None,
):
assert possible_translation_tokens is None
if select_single is not None:
return self.output_projections[select_single](
unprojected_outs[select_single])
logits = [
p(o)[0] for p, o in zip(self.output_projections, unprojected_outs)
]
avg = average_tensors(logits, norm_fn=self.norm_fn)
if self.to_log:
avg.log_()
return avg, None