def _predictDistributions(
self, in_datas: List[TacticContext]) -> torch.FloatTensor:
assert self._tokenizer
assert self._embedding
assert self.training_args
goals_batch = [
normalizeSentenceLength(self._tokenizer.toTokenList(goal),
self.training_args.max_length)
for _, _, _, goal in in_datas
]
hyps = [
get_closest_hyp(hyps, goal, self.training_args.max_length)
for _, _, hyps, goal in in_datas
]
hyp_types = [serapi_instance.get_hyp_type(hyp) for hyp in hyps]
hyps_batch = [
normalizeSentenceLength(self._tokenizer.toTokenList(hyp_type),
self.training_args.max_length)
for hyp_type in hyp_types
]
word_features_batch = [
self._get_word_features(in_data) for in_data in in_datas
]
vec_features_batch = [
self._get_vec_features(in_data) for in_data in in_datas
]
stem_distribution = self._model(LongTensor(goals_batch),
LongTensor(hyps_batch),
FloatTensor(vec_features_batch),
LongTensor(word_features_batch))
return stem_distribution
def goal_token_scores(self, stem_idxs: torch.LongTensor,
tokenized_goals: List[List[int]],
goal_masks: List[List[bool]],
) -> torch.FloatTensor:
assert self._model
assert self.training_args
batch_size = stem_idxs.size()[0]
stem_width = stem_idxs.size()[1]
goal_len = self.training_args.max_length
# The goal probabilities include the "no argument" probability
num_goal_probs = goal_len + 1
unmasked_probabilities = self._model.goal_args_model(
stem_idxs.view(batch_size * stem_width),
LongTensor(tokenized_goals).view(
batch_size, 1, goal_len)
.expand(-1, stem_width, -1).contiguous()
.view(batch_size * stem_width, goal_len))\
.view(batch_size, stem_width, num_goal_probs)
masked_probabilities = torch.where(
maybe_cuda(torch.ByteTensor(goal_masks))
.view(batch_size, 1, num_goal_probs)
.expand(-1, stem_width, -1),
unmasked_probabilities,
torch.full_like(unmasked_probabilities, -float("Inf")))
assert masked_probabilities.size() == torch.Size(
[batch_size, stem_width, num_goal_probs])
return masked_probabilities
def predictKTacticsWithLoss(self, in_data : TacticContext, k : int, correct : str) -> \
Tuple[List[Prediction], float]:
assert self.training_args
assert self._embedding
with self._lock:
prediction_distribution = self._predictDistributions([in_data])[0]
if k > self._embedding.num_tokens():
k = self._embedding.num_tokens()
correct_stem = serapi_instance.get_stem(correct)
if self._embedding.has_token(correct_stem):
output_var = maybe_cuda(
Variable(
LongTensor([self._embedding.encode_token(correct_stem)])))
loss = self._criterion(prediction_distribution.view(1, -1),
output_var).item()
else:
loss = 0
if len(in_data.hypotheses) == 0:
certainties, idxs = topk_with_filter(
prediction_distribution.view(-1), k,
lambda certainty, idx: not serapi_instance.tacticTakesHypArgs(
cast(Embedding, self._embedding).decode_token(idx)))
else:
certainties, idxs = prediction_distribution.view(-1).topk(k)
results = [
Prediction(
self.add_arg(self._embedding.decode_token(stem_idx.item()),
in_data.goal, in_data.hypotheses,
self.training_args.max_length),
math.exp(certainty.item()))
for certainty, stem_idx in zip(certainties, idxs)
]
return results, loss
def forward(self, goal_batch : torch.LongTensor, stem_batch : torch.LongTensor) \
-> torch.FloatTensor:
goal_var = maybe_cuda(Variable(goal_batch))
stem_var = maybe_cuda(Variable(stem_batch))
batch_size = goal_batch.size()[0]
initial_hidden = self._stem_embedding(stem_var)\
.view(1, batch_size, self.hidden_size)
hidden = initial_hidden
copy_likelyhoods: List[torch.FloatTensor] = []
for i in range(goal_batch.size()[1]):
token_batch = self._word_embedding(goal_var[:,i])\
.view(1, batch_size, self.hidden_size)
token_batch = F.relu(token_batch)
token_out, hidden = self._gru(token_batch, hidden)
copy_likelyhood = self._likelyhood_layer(F.relu(token_out))
copy_likelyhoods.append(copy_likelyhood[0])
end_token_embedded = self._word_embedding(LongTensor([EOS_token])
.expand(batch_size))\
.view(1, batch_size, self.hidden_size)
final_out, final_hidden = self._gru(F.relu(end_token_embedded), hidden)
final_likelyhood = self._likelyhood_layer(F.relu(final_out))
copy_likelyhoods.insert(0, final_likelyhood[0])
catted = torch.cat(copy_likelyhoods, dim=1)
result = self._softmax(catted)
return result
def predictKTacticsWithLoss_batch(self,
in_data : List[TacticContext],
k : int, corrects : List[str]) -> \
Tuple[List[List[Prediction]], float]:
assert self._embedding
assert self.training_args
with self._lock:
prediction_distributions = self._predictDistributions(in_data)
correct_stems = [serapi_instance.get_stem(correct) for correct in corrects]
output_var = maybe_cuda(Variable(
LongTensor([self._embedding.encode_token(correct_stem)
if self._embedding.has_token(correct_stem)
else 0
for correct_stem in correct_stems])))
loss = self._criterion(prediction_distributions, output_var).item()
if k > self._embedding.num_tokens():
k = self._embedding.num_tokens()
certainties_and_idxs_list = \
[single_distribution.view(-1).topk(k) if len(context.hypotheses) > 0 else
topk_with_filter(single_distribution.view(-1), k,
lambda certainty, idx:
not serapi_instance.tacticTakesHypArgs(
cast(Embedding, self._embedding).decode_token(idx)))
for single_distribution, context in
zip(prediction_distributions, in_data)]
results = [[Prediction(self.add_arg(self._embedding.decode_token(stem_idx.item()),
in_datum.goal, in_datum.hypotheses,
self.training_args.max_length),
math.exp(certainty.item()))
for certainty, stem_idx in zip(*certainties_and_idxs)]
for certainties_and_idxs, in_datum in
zip(certainties_and_idxs_list, in_data)]
return results, loss
def hyp_name_scores(self,
stem_idxs: torch.LongTensor,
tokenized_goal: List[int],
tokenized_premises: List[List[int]],
premise_features: List[List[float]]
) -> torch.FloatTensor:
assert self._model
assert len(stem_idxs.size()) == 1
stem_width = stem_idxs.size()[0]
num_hyps = len(tokenized_premises)
encoded_goals = self._model.goal_encoder(LongTensor([tokenized_goal]))
hyp_arg_values = self.runHypModel(stem_idxs.unsqueeze(0),
encoded_goals,
LongTensor([tokenized_premises]),
FloatTensor([premise_features]))
assert hyp_arg_values.size() == torch.Size([1, stem_width, num_hyps])
return hyp_arg_values
def train(dataset : SequenceSequenceDataset, hidden_size : int,
learning_rate : float, num_encoder_layers : int,
num_decoder_layers : int, max_length : int, num_epochs : int, batch_size : int,
print_every : int, context_vocab_size : int, tactic_vocab_size : int) -> Iterable[Checkpoint]:
print("Initializing PyTorch...")
in_stream = [inputFromSentence(datum[0], max_length) for datum in dataset]
out_stream = [inputFromSentence(datum[1], max_length) for datum in dataset]
data_loader = data.DataLoader(data.TensorDataset(torch.LongTensor(out_stream),
torch.LongTensor(in_stream)),
batch_size=batch_size, num_workers=0,
shuffle=True, pin_memory=True,
drop_last=True)
encoder = EncoderRNN(context_vocab_size, hidden_size, num_encoder_layers,
batch_size=batch_size)
decoder = DecoderRNN(hidden_size, tactic_vocab_size, num_decoder_layers,
batch_size=batch_size)
encoder_optimizer = optim.SGD(encoder.parameters(), lr=learning_rate)
decoder_optimizer = optim.SGD(decoder.parameters(), lr=learning_rate)
optimizers = [encoder_optimizer, decoder_optimizer]
criterion = maybe_cuda(nn.NLLLoss())
start = time.time()
num_items = len(dataset) * num_epochs
total_loss = 0
print("Training...")
for epoch in range(num_epochs):
print("Epoch {}".format(epoch))
adjustLearningRates(learning_rate, optimizers, epoch)
for batch_num, (output_batch, input_batch) in enumerate(data_loader):
target_length = output_batch.size()[1]
encoder_optimizer.zero_grad()
decoder_optimizer.zero_grad()
predictor_output = decoder.run_teach(encoder
.run(cast(SomeLongTensor, input_batch)),
cast(SomeLongTensor, output_batch))
loss = maybe_cuda(Variable(LongTensor(0)))
output_var = maybe_cuda(Variable(output_batch))
for i in range(target_length):
loss += criterion(predictor_output[i], output_var[:,i])
loss.backward()
encoder_optimizer.step()
decoder_optimizer.step()
total_loss += (loss.data[0] / target_length) * batch_size
if (batch_num + 1) % print_every == 0:
items_processed = (batch_num + 1) * batch_size + epoch * len(dataset)
progress = items_processed / num_items
print("{} ({} {:.2f}%) {:.4f}".
format(timeSince(start, progress),
items_processed, progress * 100,
total_loss / items_processed))
yield encoder.state_dict(), decoder.state_dict()
def _predictStemDistributions(self, in_datas : List[TacticContext]) \
-> torch.FloatTensor:
word_features_batch = LongTensor(
[self._get_word_features(in_data) for in_data in in_datas])
vec_features_batch = FloatTensor(
[self._get_vec_features(in_data) for in_data in in_datas])
encoded_word_features = self._model.word_features_encoder(
word_features_batch)
stem_distribution = \
self._softmax(self._model.features_classifier(torch.cat((
encoded_word_features, vec_features_batch), dim=1)))
return stem_distribution
def predictKTactics(self, in_data : TacticContext, k : int) -> \
List[Prediction]:
in_sentence = LongTensor(inputFromSentence(
self.context_tokenizer.toTokenList(in_data.goal),
self.max_length)).view(1, -1)
feature_vector = self.encoder.run(in_sentence)
prediction_sentences = decodeKTactics(self.decoder,
feature_vector,
self.beam_width,
self.max_length)[:k]
return [Prediction(self.tactic_tokenizer.toString(sentence), .5 **i)
for sentence, i in zip(prediction_sentences, itertools.count())]
def predict_stems(self, k: int,
word_features: List[List[int]],
vec_features: List[List[float]]
) -> Tuple[torch.FloatTensor, torch.LongTensor]:
assert self._model
assert len(word_features) == len(vec_features)
batch_size = len(word_features)
stem_distribution = self._model.stem_classifier(
LongTensor(word_features), FloatTensor(vec_features))
stem_probs, stem_idxs = stem_distribution.topk(k)
assert stem_probs.size() == torch.Size([batch_size, k])
assert stem_idxs.size() == torch.Size([batch_size, k])
return stem_probs, stem_idxs
def _predictFromStemDistributionWithLoss(
self, beam_width: int, stem_distribution: torch.FloatTensor,
in_datas: List[TacticContext], corrects: List[str],
k: int) -> Tuple[List[List[Prediction]], float]:
assert self.training_args
assert self._embedding
stem_probs, likely_correct_stems = stem_distribution.topk(
min(beam_width,
stem_distribution.size()[1]))
all_prob_batches, index_mapping = \
self._predictCompositeDistributionFromStemDistribution(
beam_width, stem_distribution, in_datas)
correct_idxs = LongTensor([[
max(0, arg_idx - 1) + (stem_idx * self.training_args.max_length)
for stem_idx, arg_idx in [
get_stem_and_arg_idx(
self.training_args.max_length, self._embedding,
serapi_instance.normalizeNumericArgs(
ScrapedTactic(
[], in_data.prev_tactics,
ProofContext(
[Obligation(in_data.hypotheses, in_data.goal)],
[], [], []), correct)))
]
][0] for in_data, correct in zip(in_datas, corrects)])
loss = self._criterion(all_prob_batches, correct_idxs)
final_probs, final_idxs = all_prob_batches.topk(beam_width)
row_length = self.training_args.max_length
indices = final_idxs / row_length
stem_idxs = [
index_map.index_select(0, indices1)
for index_map, indices1 in zip(index_mapping, indices)
]
arg_idxs = final_idxs % row_length
return [[
Prediction(
self._embedding.decode_token(stem_idx.item()) + " " +
get_arg_from_token_idx(in_data.goal, arg_idx.item()) + ".",
math.exp(final_prob))
for stem_idx, arg_idx, final_prob in islice(
zip(stem_list, arg_list, final_list), k)
] for in_data, stem_list, arg_list, final_list in zip(
in_datas, stem_idxs, arg_idxs, final_probs)], loss
def predictKTacticsWithLoss(self, in_data : TacticContext, k : int,
correct : str) -> Tuple[List[Prediction], float]:
with self._lock:
distribution, hyp_var = self._predictDistribution(in_data)
correct_stem = serapi_instance.get_stem(correct)
if self._embedding.has_token(correct_stem):
loss = self._criterion(distribution.view(1, -1),
Variable(LongTensor([self._embedding.encode_token(correct_stem)]))).item()
else:
loss = float("+inf")
indices, probabilities = list_topk(list(distribution), k)
predictions : List[Prediction] = []
for certainty, idx in zip(probabilities, indices):
stem = self._embedding.decode_token(idx)
if serapi_instance.tacticTakesHypArgs(stem):
predictions.append(Prediction(stem + " " + hyp_var + ".",
math.exp(certainty)))
else:
predictions.append(Prediction(stem + ".", math.exp(certainty)))
return predictions, loss
def forward(self, stem_batch: torch.LongTensor, goal_batch: torch.LongTensor) \
-> torch.FloatTensor:
goal_var = maybe_cuda(Variable(goal_batch))
stem_var = maybe_cuda(Variable(stem_batch))
batch_size = goal_batch.size()[0]
assert stem_batch.size()[0] == batch_size
initial_hidden = self._stem_embedding(stem_var)\
.view(1, batch_size, self.hidden_size)
hidden = initial_hidden
encoded_tokens: List[torch.FloatTensor] = []
for i in range(goal_batch.size()[1]):
token_batch = self._token_embedding(goal_var[:, i])\
.view(1, batch_size, self.hidden_size)
token_batch2 = F.relu(token_batch)
token_out, hidden = self._gru(token_batch2, hidden)
encoded_tokens.append(token_out.squeeze(dim=0).unsqueeze(1))
end_token_embedded = self._token_embedding(LongTensor([EOS_token])
.expand(batch_size))\
.view(1, batch_size, self.hidden_size)
final_out, _final_hidden = self._gru(F.relu(end_token_embedded), hidden)
encoded_tokens.insert(0, final_out.squeeze(dim=0).unsqueeze(1))
catted = torch.cat(encoded_tokens, dim=1)
return catted
def forward(self, stem_batch : torch.LongTensor, goal_batch : torch.LongTensor) \
-> torch.FloatTensor:
goal_var = maybe_cuda(Variable(goal_batch))
stem_var = maybe_cuda(Variable(stem_batch))
batch_size = goal_batch.size()[0]
assert stem_batch.size()[0] == batch_size
initial_hidden = self._stem_embedding(stem_var)\
.view(1, batch_size, self.hidden_size)
hidden = initial_hidden
copy_likelyhoods: List[torch.FloatTensor] = []
for i in range(goal_batch.size()[1]):
try:
token_batch = self._token_embedding(goal_var[:,i])\
.view(1, batch_size, self.hidden_size)
token_batch2 = F.relu(token_batch)
token_out, hidden = self._gru(token_batch2, hidden)
copy_likelyhood = self._likelyhood_layer(F.relu(token_out))
copy_likelyhoods.append(copy_likelyhood[0])
except RuntimeError:
eprint("Tokenized goal:")
for j in range(goal_batch.size()[0]):
eprint(goal_batch[j, i].item(), end=" ")
assert goal_batch[j, i] < 123
eprint()
eprint(f"goal_var: {goal_var}")
eprint("Token batch")
eprint(token_batch)
raise
end_token_embedded = self._token_embedding(LongTensor([EOS_token])
.expand(batch_size))\
.view(1, batch_size, self.hidden_size)
final_out, final_hidden = self._gru(F.relu(end_token_embedded), hidden)
final_likelyhood = self._likelyhood_layer(F.relu(final_out))
copy_likelyhoods.insert(0, final_likelyhood[0])
catted = torch.cat(copy_likelyhoods, dim=1)
return catted
def predictKTactics(self, context: TacticContext,
k: int) -> List[Prediction]:
assert self.training_args
assert self._model
num_stem_poss = get_num_tokens(self._metadata)
stem_width = min(self.training_args.max_beam_width, num_stem_poss,
k**2)
tokenized_premises, hyp_features, \
nhyps_batch, tokenized_goal, \
goal_mask, \
word_features, vec_features = \
sample_fpa(extract_dataloader_args(self.training_args),
self._metadata,
context.relevant_lemmas,
context.prev_tactics,
context.hypotheses,
context.goal)
num_hyps = nhyps_batch[0]
stem_distribution = self._model.stem_classifier(
LongTensor(word_features), FloatTensor(vec_features))
stem_certainties, stem_idxs = stem_distribution.topk(stem_width)
goals_batch = LongTensor(tokenized_goal)
goal_arg_values = self._model.goal_args_model(
stem_idxs.view(1 * stem_width),
goals_batch.view(1, 1, self.training_args.max_length)
.expand(-1, stem_width, -1).contiguous()
.view(1 * stem_width,
self.training_args.max_length))\
.view(1, stem_width,
self.training_args.max_length + 1)
goal_arg_values = torch.where(
torch.ByteTensor(goal_mask).view(1, 1,
self.training_args.max_length +
1).expand(-1, stem_width, -1),
goal_arg_values, torch.full_like(goal_arg_values, -float("Inf")))
assert goal_arg_values.size() == torch.Size([1, stem_width,
self.training_args.max_length + 1]),\
"goal_arg_values.size(): {}; stem_width: {}".format(goal_arg_values.size(),
stem_width)
num_probs = 1 + num_hyps + self.training_args.max_length
goal_symbols = get_fpa_words(context.goal)
num_valid_probs = (1 + num_hyps + len(goal_symbols)) * stem_width
if num_hyps > 0:
encoded_goals = self._model.goal_encoder(goals_batch)\
.view(1, 1, self.training_args.hidden_size)
hyps_batch = LongTensor(tokenized_premises)
assert hyps_batch.size() == torch.Size([1, num_hyps,
self.training_args.max_length]), \
(hyps_batch.size(),
num_hyps,
self.training_args.max_length)
hypfeatures_batch = FloatTensor(hyp_features)
assert hypfeatures_batch.size() == \
torch.Size([1, num_hyps, hypFeaturesSize()]), \
(hypfeatures_batch.size(), num_hyps, hypFeaturesSize())
hyp_arg_values = self.runHypModel(stem_idxs, encoded_goals,
hyps_batch, hypfeatures_batch)
assert hyp_arg_values.size() == \
torch.Size([1, stem_width, num_hyps])
total_values = torch.cat((goal_arg_values, hyp_arg_values), dim=2)
else:
total_values = goal_arg_values
all_prob_batches = self._softmax((total_values +
stem_certainties.view(1, stem_width, 1)
.expand(-1, -1, num_probs))
.contiguous()
.view(1, stem_width * num_probs))\
.view(stem_width * num_probs)
final_probs, final_idxs = all_prob_batches.topk(k)
assert not torch.isnan(final_probs).any()
assert final_probs.size() == torch.Size([k])
row_length = self.training_args.max_length + num_hyps + 1
stem_keys = final_idxs // row_length
assert stem_keys.size() == torch.Size([k])
assert stem_idxs.size() == torch.Size([1,
stem_width]), stem_idxs.size()
prediction_stem_idxs = stem_idxs.view(stem_width).index_select(
0, stem_keys)
assert prediction_stem_idxs.size() == torch.Size([k]), \
prediction_stem_idxs.size()
arg_idxs = final_idxs % row_length
assert arg_idxs.size() == torch.Size([k])
if self.training_args.lemma_args:
all_hyps = context.hypotheses + context.relevant_lemmas
else:
all_hyps = context.hypotheses
return [
Prediction(
decode_fpa_result(extract_dataloader_args(self.training_args),
self._metadata, all_hyps, context.goal,
stem_idx.item(), arg_idx.item()),
math.exp(prob)) for stem_idx, arg_idx, prob in islice(
zip(prediction_stem_idxs, arg_idxs, final_probs),
min(k, num_valid_probs))
]
def predictKTactics_batch(self, context_batch: List[TacticContext],
k: int) \
-> List[List[Prediction]]:
assert self.training_args
assert self._model
num_stem_poss = get_num_tokens(self._metadata)
stem_width = min(self.training_args.max_beam_width, num_stem_poss,
k**2)
batch_size = len(context_batch)
tprems_batch, pfeat_batch, \
nhyps_batch, tgoals_batch, \
goal_masks_batch, \
wfeats_batch, vfeats_batch = \
sample_fpa_batch(extract_dataloader_args(self.training_args),
self._metadata,
[context_py2r(context)
for context in context_batch])
for tprem, pfeat, nhyp, tgoal, masks, wfeat, vfeat, context \
in zip(tprems_batch, pfeat_batch,
nhyps_batch, tgoals_batch,
goal_masks_batch, wfeats_batch,
vfeats_batch, context_batch):
s_tprem, s_pfeat, s_nhyp, s_tgoal, s_masks, s_wfeat, s_vfeat = \
sample_fpa(extract_dataloader_args(self.training_args),
self._metadata,
context.relevant_lemmas,
context.prev_tactics,
context.hypotheses,
context.goal)
assert len(s_tprem) == 1
assert len(tprem) == len(s_tprem[0])
for p1, p2 in zip(tprem, s_tprem[0]):
assert p1 == p2, (p1, p2)
assert len(s_pfeat) == 1
assert len(pfeat) == len(s_pfeat[0])
for f1, f2 in zip(pfeat, s_pfeat[0]):
assert f1 == f2, (f1, f2)
assert s_nhyp[0] == nhyp, (s_nhyp[0], nhyp)
assert s_tgoal[0] == tgoal, (s_tgoal[0], tgoal)
assert s_masks[0] == masks, (s_masks[0], masks)
assert s_wfeat[0] == wfeat, (s_wfeat[0], wfeat)
assert s_vfeat[0] == vfeat, (s_vfeat[0], vfeat)
stem_distribution = self._model.stem_classifier(
LongTensor(wfeats_batch), FloatTensor(vfeats_batch))
stem_certainties_batch, stem_idxs_batch = stem_distribution.topk(
stem_width)
goals_batch = LongTensor(tgoals_batch)
goal_arg_values = self._model.goal_args_model(
stem_idxs_batch.view(batch_size * stem_width),
goals_batch.view(batch_size, 1, self.training_args.max_length)
.expand(-1, stem_width, -1).contiguous()
.view(batch_size * stem_width,
self.training_args.max_length))\
.view(batch_size, stem_width,
self.training_args.max_length + 1)
goal_arg_values = torch.where(
torch.ByteTensor(goal_masks_batch).view(
batch_size, 1,
self.training_args.max_length + 1).expand(-1, stem_width, -1),
goal_arg_values, torch.full_like(goal_arg_values, -float("Inf")))
encoded_goals_batch = self._model.goal_encoder(goals_batch)
stems_expanded_batch = torch.cat([
stem_idxs.view(1, stem_width).expand(
num_hyps, stem_width).contiguous().view(num_hyps * stem_width)
for stem_idxs, num_hyps, in zip(stem_idxs_batch, nhyps_batch)
])
egoals_expanded_batch = torch.cat([
encoded_goal.view(1, self.training_args.hidden_size).expand(
num_hyps * stem_width, -1).contiguous()
for encoded_goal, num_hyps in zip(encoded_goals_batch, nhyps_batch)
])
tprems_expanded_batch = torch.cat([
LongTensor(tpremises).view(
1, -1, self.training_args.max_length).expand(
stem_width, -1,
-1).contiguous().view(-1, self.training_args.max_length)
for tpremises in tprems_batch
])
pfeat_expanded_batch = torch.cat([
FloatTensor(premise_features).view(1, num_hyps, 2).expand(
stem_width, -1, -1).contiguous().view(num_hyps * stem_width, 2)
for premise_features, num_hyps in zip(pfeat_batch, nhyps_batch)
])
prem_arg_values = self._model.hyp_model(stems_expanded_batch,
egoals_expanded_batch,
tprems_expanded_batch,
pfeat_expanded_batch)
prem_arg_values_split = prem_arg_values.split(
[num_hyps * stem_width for num_hyps in nhyps_batch])
total_values_list = [
torch.cat((goal_values, prem_values.view(stem_width, -1)),
dim=1) for goal_values, prem_values in zip(
goal_arg_values, prem_arg_values_split)
]
all_probs_list = [
self._softmax((total_values + stem_certainties.view(
stem_width, 1).expand_as(total_values)).contiguous().view(
1, -1)).view(-1) for total_values, stem_certainties in zip(
total_values_list, stem_certainties_batch)
]
final_probs_list, final_idxs_list = zip(
*[probs.topk(k) for probs in all_probs_list])
stem_keys_list = [
final_idxs // (self.training_args.max_length + num_hyps + 1)
for final_idxs, num_hyps in zip(final_idxs_list, nhyps_batch)
]
stem_idxs_list = [
stem_idxs.view(stem_width).index_select(0, stem_keys)
for stem_idxs, stem_keys in zip(stem_idxs_batch, stem_keys_list)
]
arg_idxs_list = [
final_idxs % (self.training_args.max_length + num_hyps + 1)
for final_idxs, num_hyps in zip(final_idxs_list, nhyps_batch)
]
predictions = [[
Prediction(
decode_fpa_result(extract_dataloader_args(self.training_args),
self._metadata,
context.hypotheses + context.relevant_lemmas,
context.goal,
stem_idx.item(), arg_idx.item()),
math.exp(prob)) for stem_idx, arg_idx, prob in islice(
zip(stem_idxs, arg_idxs, final_probs),
min(k, 1 + num_hyps + len(get_fpa_words(context.goal))))
] for stem_idxs, arg_idxs, final_probs, context, num_hyps in zip(
stem_idxs_list, arg_idxs_list, final_probs_list, context_batch,
nhyps_batch)]
for context, pred_list in zip(context_batch, predictions):
for batch_pred, single_pred in zip(
pred_list, self.predictKTactics(context, k)):
assert batch_pred.prediction == single_pred.prediction, \
(batch_pred, single_pred)
return predictions
def initInput(self) -> SomeLongTensor:
return Variable(LongTensor([[SOS_token] * self.batch_size]))
def decodeKTactics(decoder : DecoderRNN, encoder_hidden : torch.FloatTensor,
beam_width : int, max_length : int):
v = decoder.output_size
pos_index = Variable(LongTensor([0]) * beam_width).view(-1, 1)
hidden = _inflate(encoder_hidden, beam_width)
sequence_scores = FloatTensor(beam_width, 1)
sequence_scores.fill_(-float('Inf'))
sequence_scores.index_fill_(0, LongTensor([0]), 0.0)
sequence_scores = Variable(sequence_scores)
input_var = Variable(LongTensor([[SOS_token] * beam_width]))
stored_predecessors = list()
stored_emitted_symbols = list()
for j in range(max_length):
decoder_output, hidden = decoder(input_var, hidden)
sequence_scores = _inflate(sequence_scores, v)
sequence_scores += decoder_output
scores, candidates = sequence_scores.view(1, -1).topk(beam_width)
input_var = (candidates % v).view(1, beam_width)
sequence_scores = scores.view(beam_width, 1)
predecessors = (candidates / v +
pos_index.expand_as(candidates)).view(beam_width, 1)
hidden = hidden.index_select(1, cast(torch.LongTensor, predecessors.squeeze()))
eos_indices = input_var.data.eq(EOS_token)
if eos_indices.nonzero().dim() > 0:
sequence_scores.data.masked_fill_(torch.transpose(eos_indices, 0, 1),
-float('inf'))
stored_predecessors.append(predecessors)
stored_emitted_symbols.append(torch.transpose(input_var, 0, 1))
# Trace back from the final three highest scores
_, next_idxs = sequence_scores.view(beam_width).sort(descending=True)
seqs = [] # type: List[List[SomeLongTensor]]
eos_found = 0
for i in range(max_length - 1, -1, -1):
# The next column of symbols from the end
next_symbols = stored_emitted_symbols[i].view(beam_width) \
.index_select(0, next_idxs).data
# The predecessors of that column
next_idxs = stored_predecessors[i].view(beam_width).index_select(0, next_idxs)
# Handle sequences that ended early
eos_indices = stored_emitted_symbols[i].data.squeeze(1).eq(EOS_token).nonzero()
if eos_indices.dim() > 0:
for j in range(eos_indices.size(0)-1, -1, -1):
idx = eos_indices[j]
res_k_idx = beam_width - (eos_found % beam_width) - 1
eos_found += 1
res_idx = res_k_idx
next_idxs[res_idx] = stored_predecessors[i][idx[0]]
next_symbols[res_idx] = stored_emitted_symbols[i][idx[0]].data[0]
# Commit the result
seqs.insert(0, next_symbols)
# Transpose
int_seqs = [[data[i][0] for data in seqs] for i in range(beam_width)]
# Cut off EOS tokens
int_seqs = [list(takewhile(lambda x: x != EOS_token, seq)) for seq in int_seqs]
return int_seqs
