2017 年 8 月 6 日發布的 pytorch 0.2.0 哪個特性最吸引你？

def pad_sentences(doc):
"""Converts a list of indices to a padded torch Variable"""
n_sents = len(doc)
len_ix = np.argsort([-len(sent) for sent in doc], kind="mergesort")
doc = [doc[ix] for ix in len_ix]
lens = [len(sent) for sent in doc]
out = doc[0].new(n_sents, lens[0]).fill_(Constants.PAD)
for i in range(n_sents):
sent_length = doc[i].size(0)
out[i].narrow(0, 0, sent_length).copy_(doc[i])

inv_ix = np.zeros_like(len_ix)
inv_ix[len_ix] = np.arange(len(len_ix))
inv_ix = Variable(torch.LongTensor(inv_ix),
requires_grad=False).detach()
out = Variable(out, requires_grad=False)

if out.is_cuda:
inv_ix = inv_ix.cuda()

return out, lens, inv_ix

def _make_detached_long(x, cuda=False):
y = Variable(torch.LongTensor(x), requires_grad=False).detach()
if cuda:
y = y.cuda()
return y

def forward(self, X_train):
# first, collect all sentences in the batch
all_sents = [sent for doc in X_train for sent in doc]
lengths = [len(sent) for sent in all_sents]
offsets = np.cumsum([len(doc) for doc in X_train])
all_sents_padded, sorted_lens, inv_ix = pad_sentences(all_sents)
all_sents_embed = self.embed(all_sents_padded)

all_sents_pack = pack_padded_sequence(all_sents_embed, sorted_lens,
batch_first=True)
all_sents_enc_pack, _ = self.word_encoder(all_sents_pack)
all_sents_enc, _ = pad_packed_sequence(all_sents_enc_pack,
batch_first=True)

# undo sorting
all_sents_enc = all_sents_enc.index_select(0, inv_ix)

# test unsorting
# for sent, sent_enc in zip(all_sents, all_sents_enc):
# assert len(sent) == torch.prod(sent_enc != 0, 1).sum()

# word-level attention
_, sents_repr = self.word_attn(all_sents_enc, lengths)

# group sentences by doc
doc_ixs = np.split(np.arange(len(all_sents)), offsets[:-1])
doc_ixs = [_make_detached_long(ix, cuda=sents_repr.is_cuda)
for ix in doc_ixs]
sents_by_doc = [sents_repr.index_select(0, ix) for ix in doc_ixs]
# check correctness
# assert len(sents_by_doc) == len(X_train)
# for doc, sents_ in zip(X_train, sents_by_doc):
# assert(len(doc) == len(sents_))

# pad, pack and apply sentence-level RNN
doc_lens = [len(doc) for doc in X_train]
docs_padded, doc_lens_sorted, doc_inv_ix = pad_docs(sents_by_doc)
docs_pack = pack_padded_sequence(docs_padded, doc_lens_sorted,
batch_first=True)
docs_enc_pack, _ = self.sent_encoder(docs_pack)
docs_enc, _ = pad_packed_sequence(docs_enc_pack, batch_first=True)

# undo sorting
docs_enc = docs_enc.index_select(0, doc_inv_ix)

# sentence-level attention
_, docs_repr = self.sent_attn(docs_enc, doc_lens)