src/temp_predictor.py

import spacy
import transformers
import numpy as np

class TempPredictor:
    def __init__(self, model, tokenizer, device,
                 spacy_model="en_core_web_sm"):
        self._model = model
        self._model.to(device)
        self._model.eval()
        self._tokenizer = tokenizer
        self._mtoken = self._tokenizer.mask_token
        self.unmasker = transformers.pipeline("fill-mask", model=self._model, tokenizer=self._tokenizer, device=0)
        try:
            self._spacy = spacy.load(spacy_model)
        except Exception as e:
            self._spacy = spacy.load("en_core_web_sm")
            print(f"Failed to load spacy model {spacy_model}, use default 'en_core_web_sm'\n{e}")

    def _extract_token_prob(self, arr, token, crop=1):
        for it in arr:
            if len(it["token_str"]) >= crop and (token == it["token_str"][crop:]):
                return it["score"]
        return 0.

    def _sent_lowercase(self, s):
        try:
            return s[0].lower() + s[1:]
        except:
            return s

    def _remove_punct(self, s):
        try:
            return s[:-1]
        except:
            return s

    def predict(self, e1, e2, top_k=5):
        txt = self._remove_punct(e1) + " " + self._mtoken + " " + self._sent_lowercase(e2)
        return self.unmasker(txt, top_k=top_k)

    def batch_predict(self, instances, top_k=5):
        txt = [self._remove_punct(e1) + " " + self._mtoken + " " + self._sent_lowercase(e2)
                for (e1, e2) in  instances]
        return self.unmasker(txt, top_k=top_k)


    def get_temp(self, e1, e2, top_k=5, crop=1):
        inst1 = self.predict(e1, e2, top_k)
        inst2 = self.predict(e2, e1, top_k)

        # e1 before e2
        b1 = self._extract_token_prob(inst1, "before", crop=crop)
        b2 = self._extract_token_prob(inst2, "after", crop=crop)

        # e1 after e2
        a1 = self._extract_token_prob(inst1, "after", crop=crop)
        a2 = self._extract_token_prob(inst2, "before", crop=crop)

        return (b1+b2)/2, (a1+a2)/2

    def get_temp_batch(self, instances, top_k=5, crop=1):
        reverse_instances = [(e2, e1) for (e1, e2) in instances]
        fwd_preds = self.batch_predict(instances, top_k=top_k)
        bwd_preds = self.batch_predict(reverse_instances, top_k=top_k)

        b1s = np.array([ self._extract_token_prob(pred, "before", crop=crop) for pred in fwd_preds ])
        b2s = np.array([ self._extract_token_prob(pred, "before", crop=crop) for pred in bwd_preds ])
        a1s = np.array([ self._extract_token_prob(pred, "after", crop=crop) for pred in fwd_preds ])
        a2s = np.array([ self._extract_token_prob(pred, "after", crop=crop) for pred in bwd_preds ])

        return np.array([np.array(b1s+b2s)/2, np.array(a1s+a2s)/2]).T


    def __call__(self, *args, **kwargs):
        return self.get_temp(*args, **kwargs)