import tensorflow as tf
from transformers import TFAutoModelForSequenceClassification
from capreolus import ConfigOption, Dependency
from capreolus.reranker import Reranker
[docs]class TFElectraRelevanceHead(tf.keras.layers.Layer):
"""BERT-style ClassificationHead (i.e., out_proj only -- no dense). See transformers.TFElectraClassificationHead"""
def __init__(self, dropout, out_proj, *args, **kwargs):
super().__init__(*args, **kwargs)
self.dropout = dropout
self.out_proj = out_proj
[docs] def call(self, inputs, **kwargs):
x = inputs[:, 0, :] # take <s> token (equiv. to [CLS])
x = self.dropout(x)
x = self.out_proj(x)
return x
[docs]class TFBERTMaxP_Class(tf.keras.layers.Layer):
def __init__(self, extractor, config, *args, **kwargs):
super(TFBERTMaxP_Class, self).__init__(*args, **kwargs)
self.extractor = extractor
# TODO hidden prob missing below?
if config["pretrained"] == "electra-base-msmarco":
self.bert = TFAutoModelForSequenceClassification.from_pretrained("Capreolus/electra-base-msmarco")
dropout, fc = self.bert.classifier.dropout, self.bert.classifier.out_proj
self.bert.classifier = TFElectraRelevanceHead(dropout, fc)
elif config["pretrained"] == "electra-base":
self.bert = TFAutoModelForSequenceClassification.from_pretrained("google/electra-base-discriminator")
dropout, fc = self.bert.classifier.dropout, self.bert.classifier.out_proj
self.bert.classifier = TFElectraRelevanceHead(dropout, fc)
elif config["pretrained"] == "bert-base-msmarco":
self.bert = TFAutoModelForSequenceClassification.from_pretrained("Capreolus/bert-base-msmarco")
else:
self.bert = TFAutoModelForSequenceClassification.from_pretrained(
config["pretrained"], hidden_dropout_prob=config["hidden_dropout_prob"], from_pt=True
)
self.config = config
[docs] def call(self, x, **kwargs):
"""
Returns logits of shape [2]
"""
doc_bert_input, doc_mask, doc_seg = x[0], x[1], x[2]
if "roberta" in self.config["pretrained"]:
doc_seg = tf.zeros_like(doc_mask) # since roberta does not have segment input
passage_scores = self.bert(doc_bert_input, attention_mask=doc_mask, token_type_ids=doc_seg)[0]
return passage_scores
[docs] def predict_step(self, data):
"""
Scores each passage and applies max pooling over it.
"""
posdoc_bert_input, posdoc_mask, posdoc_seg, negdoc_bert_input, negdoc_mask, negdoc_seg = data
batch_size = tf.shape(posdoc_bert_input)[0]
num_passages = self.extractor.config["numpassages"]
maxseqlen = self.extractor.config["maxseqlen"]
passage_position = tf.reduce_sum(posdoc_mask * posdoc_seg, axis=-1) # (B, P)
passage_mask = tf.cast(tf.greater(passage_position, 5), tf.float32) # (B, P)
posdoc_bert_input = tf.reshape(posdoc_bert_input, [batch_size * num_passages, maxseqlen])
posdoc_mask = tf.reshape(posdoc_mask, [batch_size * num_passages, maxseqlen])
posdoc_seg = tf.reshape(posdoc_seg, [batch_size * num_passages, maxseqlen])
passage_scores = self.call((posdoc_bert_input, posdoc_mask, posdoc_seg), training=False)[:, 1]
passage_scores = tf.reshape(passage_scores, [batch_size, num_passages])
if self.config["aggregation"] == "max":
passage_scores = tf.math.reduce_max(passage_scores, axis=1)
elif self.config["aggregation"] == "first":
passage_scores = passage_scores[:, 0]
elif self.config["aggregation"] == "sum":
passage_scores = tf.math.reduce_sum(passage_mask * passage_scores, axis=1)
elif self.config["aggregation"] == "avg":
passage_scores = tf.math.reduce_sum(passage_mask * passage_scores, axis=1) / tf.reduce_sum(passage_mask)
else:
raise ValueError("Unknown aggregation method: {}".format(self.config["aggregation"]))
return passage_scores
[docs] def score(self, x, **kwargs):
posdoc_bert_input, posdoc_mask, posdoc_seg, negdoc_bert_input, negdoc_mask, negdoc_seg = x
return self.call((posdoc_bert_input, posdoc_mask, posdoc_seg), **kwargs)
[docs] def score_pair(self, x, **kwargs):
posdoc_bert_input, posdoc_mask, posdoc_seg, negdoc_bert_input, negdoc_mask, negdoc_seg = x
pos_score = self.call((posdoc_bert_input, posdoc_mask, posdoc_seg), **kwargs)[:, 1]
neg_score = self.call((negdoc_bert_input, negdoc_mask, negdoc_seg), **kwargs)[:, 1]
return pos_score, neg_score
@Reranker.register
[docs]class TFBERTMaxP(Reranker):
"""
TensorFlow implementation of BERT-MaxP.
Deeper Text Understanding for IR with Contextual Neural Language Modeling. Zhuyun Dai and Jamie Callan. SIGIR 2019.
https://arxiv.org/pdf/1905.09217.pdf
"""
[docs] module_name = "TFBERTMaxP"
[docs] dependencies = [
Dependency(key="extractor", module="extractor", name="bertpassage"),
Dependency(key="trainer", module="trainer", name="tensorflow"),
]
[docs] config_spec = [
ConfigOption(
"pretrained",
"bert-base-uncased",
"Pretrained model: bert-base-uncased, bert-base-msmarco, electra-base-msmarco, or HuggingFace supported models",
),
ConfigOption("aggregation", "max"),
ConfigOption("hidden_dropout_prob", 0.1, "The dropout probability of BERT-like model's hidden layers."),
]
[docs] def build_model(self):
self.model = TFBERTMaxP_Class(self.extractor, self.config)
return self.model