def save_dev_scores(model, val, val_x):
val_probs = model.predict_proba(val_x)
score_d = {}
for pc_vector_feature, prediction in zip(val, val_probs):
cid = pc_vector_feature.claim_pers.cid
pid = pc_vector_feature.claim_pers.pid
cpid = CPID("{}_{}".format(cid, pid))
score_d[cpid] = prediction[1]
save_to_pickle(score_d, "pc_ngram_logits")
def prediction_to_dict(
prediction: List[Tuple[str, List[Dict]]]) -> Dict[CPID, float]:
output: Dict[CPID, float] = {}
for claim_id, preds in prediction:
for pred in preds:
cpid = CPID("{}_{}".format(claim_id, pred['pid']))
score = pred['score']
output[cpid] = float(score)
return output
def get_cpids_and_token_keys(
tokenizer: FullTokenizer,
claim_entry: ParagraphClaimPersFeature) -> Tuple[str, CPID]:
claim_text = claim_entry.claim_pers.claim_text
claim_tokens = tokenizer.tokenize(claim_text)
p_text = claim_entry.claim_pers.p_text
p_tokens = tokenizer.tokenize(p_text)
key = " ".join(claim_tokens) + "_" + " ".join(p_tokens)
cpid: CPID = CPID("{}_{}".format(claim_entry.claim_pers.cid,
claim_entry.claim_pers.pid))
return key, cpid
def get_cpid(data_id, info_d) -> CPID:
try:
info_1 = info_d[data_id - 1]
info_2 = info_d[data_id]
cid = info_1['cid']
pid = info_2['pid']
except KeyError:
info_1 = info_d[data_id]
info_2 = info_d[data_id + 1]
cid = info_1['cid']
pid = info_2['pid']
return CPID("{}_{}".format(cid, pid))
def collect_by_order(input_file, feature_data: List[PerspectiveCandidate]):
predictions = EstimatorPredictionViewer(input_file)
print("prediction : {}".format(predictions.data_len))
print("feature_data : {}".format(len(feature_data)))
score_d: Dict[CPID, float] = {}
for pred_entry, pc_candidate in zip(predictions, feature_data):
logits = pred_entry.get_vector("logits")
probs = softmax(logits)
score = probs[1]
cpid = CPID("{}_{}".format(pc_candidate.cid, pc_candidate.pid))
score_d[cpid] = score
return score_d
def get_predictions(
claim_and_candidate: Tuple[Dict,
List[Dict]]) -> Tuple[str, List[Dict]]:
claim_info, candidates = claim_and_candidate
nonlocal dp_not_found
for candi in candidates:
cid = candi['cid']
pid = candi['pid']
cpid = CPID("{}_{}".format(cid, pid))
if cpid in score_d:
candi['new_score'] = score_d[cpid]
else:
dp_not_found += 1
candi['new_score'] = 0.01
candi['final_score'] = candi['new_score'] + candi['score'] / 100
candi[
'rationale'] = "final_score={} cls_score={} lucene_score={}".format(
candi['final_score'], candi['new_score'], candi['score'])
candidates.sort(key=lambda c: c['final_score'], reverse=True)
return claim_info['cId'], candidates[:top_k]
def CPID_to_CPIDPair(cpid: CPID) -> CPIDPair:
cid, pid = cpid.split("_")
return CPIDPair((int(cid), int(pid)))