def from_pytt(cls, fields, *, is_grad=False) -> "Activations":
"""Create Activations from the output tuples produced by PyTorch Transformers.
Includes converting torch tensors to xp, and handling missing values.
"""
# lh: last hidden
# po: pooler_output
# ah: all_hidden
# aa: all_attention
if len(fields) != 4:
lh = fields[0]
po = tuple()
ah = []
aa = []
else:
lh, po, ah, aa = fields
# Convert last_hidden_state to xp
lh = torch2xp(lh)
xp = get_array_module(lh)
# Normalize "None" value for pooler output
if isinstance(po, tuple):
po = xp.zeros((0, ), dtype=lh.dtype)
else:
po = torch2xp(po)
ah = list(map(torch2xp, ah))
aa = list(map(torch2xp, aa))
return cls(lh, po, ah, aa, is_grad=is_grad)
def join(cls, sub_acts: List["Activations"]) -> "Activations":
"""Concatenate activations from subsequences."""
xp = get_array_module(sub_acts[0].lh)
lh: Array = xp.vstack([x.lh for x in sub_acts])
po: Array = xp.vstack([x.po for x in sub_acts])
# Transpose the lists, so that the inner list items refer
# to the subsequences. Then we can vstack those.
ah = list(map(xp.vstack, zip(*[x.ah for x in sub_acts])))
#aa = list(map(xp.vstack, zip(*[x.aa for x in sub_acts])))
aa = []
return cls(lh, po, ah, aa, is_grad=sub_acts[0].is_grad)
def pad_batch_activations(batch: List[Activations],
*,
to: int = 0) -> Activations:
if not batch:
return Activations.blank()
xp = get_array_module(batch[0])
lh = pad_batch([x.lh for x in batch], xp=xp, to=to)
if lh.size:
lh = lh.reshape((len(batch), -1, lh.shape[-1]))
po = pad_batch([x.po for x in batch], xp=xp, to=to)
if po.size:
po = po.reshape((len(batch), -1, po.shape[-1]))
# Transpose the lists, and then pad_batch the items
ah = [
pad_batch(list(seq), xp=xp, to=to)
for seq in zip(*[x.ah for x in batch])
]
aa = [
pad_batch(list(seq), xp=xp, to=to)
for seq in zip(*[x.aa for x in batch])
]
return Activations(lh, po, ah, aa, is_grad=batch[0].is_grad)
def get_similarity_via_tensor(doc1, doc2):
v1 = doc1.vector
v2 = doc2.vector
xp = get_array_module(v1)
return xp.dot(v1, v2) / (doc1.vector_norm * doc2.vector_norm)
def set_annotations(self, docs, activations):
"""Assign the extracted features to the Doc objects and overwrite the
vector and similarity hooks.
docs (iterable): A batch of `Doc` objects.
activations (iterable): A batch of activations.
"""
for doc, doc_acts in zip(docs, activations):
xp = get_array_module(doc_acts.lh)
wp_tensor = doc_acts.lh
doc.tensor = self.model.ops.allocate((len(doc), self.model.nO))
doc._.pytt_last_hidden_state = wp_tensor
doc._.pytt_pooler_output = doc_acts.po
doc._.pytt_all_hidden_states = doc_acts.ah
doc._.pytt_all_attentions = doc_acts.aa
doc._.pytt_d_last_hidden_state = xp.zeros((0, ),
dtype=wp_tensor.dtype)
doc._.pytt_d_pooler_output = xp.zeros((0, ), dtype=wp_tensor.dtype)
doc._.pytt_d_all_hidden_states = []
doc._.pytt_d_all_attentions = []
if wp_tensor.shape != (len(doc._.pytt_word_pieces), self.model.nO):
print("# word pieces: ", len(doc._.pytt_word_pieces))
print("# tensor rows: ", wp_tensor.shape[0])
for sent in doc.sents:
if sent._.pytt_start is None or sent._.pytt_end is None:
print("Text: ", sent.text)
print("WPs: ", sent._.pytt_word_pieces_)
print(sent._.pytt_start, sent._.pytt_end)
raise ValueError(
"Mismatch between tensor shape and word pieces. This usually "
"means we did something wrong in the sentence reshaping, "
"or possibly finding the separator tokens.")
# Count how often each word-piece token is represented. This allows
# a weighted sum, so that we can make sure doc.tensor.sum()
# equals wp_tensor.sum().
# TODO: Obviously incrementing the rows individually is bad. Need
# to make this more efficient. Maybe just copy to CPU, do our stuff,
# copy back to GPU?
align_sizes = [0 for _ in range(len(doc._.pytt_word_pieces))]
for word_piece_slice in doc._.pytt_alignment:
for i in word_piece_slice:
align_sizes[i] += 1
for i, word_piece_slice in enumerate(doc._.pytt_alignment):
for j in word_piece_slice:
doc.tensor[i] += wp_tensor[j] / align_sizes[j]
# To make this weighting work, we "align" the boundary tokens against
# every token in their sentence.
if doc.tensor.sum() != wp_tensor.sum():
for sent in doc.sents:
if sent._.pytt_start is not None and sent._.pytt_end is not None:
cls_vector = wp_tensor[sent._.pytt_start]
sep_vector = wp_tensor[sent._.pytt_end]
doc.tensor[sent.start:sent.end +
1] += cls_vector / len(sent)
doc.tensor[sent.start:sent.end +
1] += sep_vector / len(sent)
doc.user_hooks["vector"] = get_doc_vector_via_tensor
doc.user_span_hooks["vector"] = get_span_vector_via_tensor
doc.user_token_hooks["vector"] = get_token_vector_via_tensor
doc.user_hooks["similarity"] = get_similarity_via_tensor
doc.user_span_hooks["similarity"] = get_similarity_via_tensor
doc.user_token_hooks["similarity"] = get_similarity_via_tensor
def join(cls, sub_acts: List["Activations"]) -> "Activations":
"""Concatenate activations from subsequences."""
xp = get_array_module(sub_acts[0].lh)
lh: Array = xp.vstack([x.lh for x in sub_acts])
return cls(lh, [], [], [], is_grad=sub_acts[0].is_grad)
def pad_batch_activations(batch: List[Activations]) -> Activations:
xp = get_array_module(batch[0])
lh = pad_batch([x.lh for x in batch], xp=xp)
lh = lh.reshape((len(batch), -1, lh.shape[-1]))
return Activations(lh, [], [], [], is_grad=batch[0].is_grad)
