From 58673a086dd56e6ac08d32c5aaaedeca47a3d02e Mon Sep 17 00:00:00 2001 From: huenrui Date: Tue, 16 Mar 2021 10:37:53 +0800 Subject: [PATCH] add reranker and reader in modelzoo/research/tprr --- model_zoo/research/nlp/tprr/README.md | 88 ++- .../nlp/tprr/reranker_and_reader_eval.py | 55 ++ .../nlp/tprr/scripts/run_eval_ascend.sh | 14 +- .../run_eval_ascend_reranker_reader.sh | 39 ++ .../nlp/tprr/src/build_reranker_data.py | 430 ++++++++++++ model_zoo/research/nlp/tprr/src/config.py | 18 +- .../nlp/tprr/src/hotpot_evaluate_v1.py | 153 ++++ model_zoo/research/nlp/tprr/src/reader.py | 73 ++ .../nlp/tprr/src/reader_albert_xxlarge.py | 263 +++++++ .../nlp/tprr/src/reader_downstream.py | 213 ++++++ .../research/nlp/tprr/src/reader_eval.py | 142 ++++ .../nlp/tprr/src/rerank_albert_xxlarge.py | 276 ++++++++ .../src/rerank_and_reader_data_generator.py | 183 +++++ .../nlp/tprr/src/rerank_and_reader_utils.py | 656 ++++++++++++++++++ .../nlp/tprr/src/rerank_downstream.py | 61 ++ model_zoo/research/nlp/tprr/src/reranker.py | 45 ++ .../research/nlp/tprr/src/reranker_eval.py | 85 +++ 17 files changed, 2765 insertions(+), 29 deletions(-) create mode 100644 model_zoo/research/nlp/tprr/reranker_and_reader_eval.py create mode 100644 model_zoo/research/nlp/tprr/scripts/run_eval_ascend_reranker_reader.sh create mode 100644 model_zoo/research/nlp/tprr/src/build_reranker_data.py create mode 100644 model_zoo/research/nlp/tprr/src/hotpot_evaluate_v1.py create mode 100644 model_zoo/research/nlp/tprr/src/reader.py create mode 100644 model_zoo/research/nlp/tprr/src/reader_albert_xxlarge.py create mode 100644 model_zoo/research/nlp/tprr/src/reader_downstream.py create mode 100644 model_zoo/research/nlp/tprr/src/reader_eval.py create mode 100644 model_zoo/research/nlp/tprr/src/rerank_albert_xxlarge.py create mode 100644 model_zoo/research/nlp/tprr/src/rerank_and_reader_data_generator.py create mode 100644 model_zoo/research/nlp/tprr/src/rerank_and_reader_utils.py create mode 100644 model_zoo/research/nlp/tprr/src/rerank_downstream.py create mode 100644 model_zoo/research/nlp/tprr/src/reranker.py create mode 100644 model_zoo/research/nlp/tprr/src/reranker_eval.py diff --git a/model_zoo/research/nlp/tprr/README.md b/model_zoo/research/nlp/tprr/README.md index 25ffac30a8..d2a02a295b 100644 --- a/model_zoo/research/nlp/tprr/README.md +++ b/model_zoo/research/nlp/tprr/README.md @@ -38,6 +38,9 @@ Wikipedia data: the 2017 English Wikipedia dump version with bidirectional hyper dev data: HotPotQA full wiki setting dev data with 7398 question-answer pairs. dev tf-idf data: the candidates for each question in dev data which is originated from top-500 retrieved from 5M paragraphs of Wikipedia through TF-IDF. +The dataset of re-ranker consists of two parts: +Wikipedia data: the 2017 English Wikipedia dump version. +dev data: HotPotQA full wiki setting dev data with 7398 question-answer pairs. # [Features](#contents) @@ -64,6 +67,7 @@ After installing MindSpore via the official website and Dataset is correctly gen ```python # run evaluation example with HotPotQA dev dataset sh run_eval_ascend.sh + sh run_eval_ascend_reranker_reader.sh ``` # [Script Description](#contents) @@ -75,25 +79,39 @@ After installing MindSpore via the official website and Dataset is correctly gen └─tprr ├─README.md ├─scripts - | ├─run_eval_ascend.sh # Launch evaluation in ascend + | ├─run_eval_ascend.sh # Launch retriever evaluation in ascend + | └─run_eval_ascend_reranker_reader # Launch re-ranker and reader evaluation in ascend | ├─src - | ├─config.py # Evaluation configurations - | ├─onehop.py # Onehop model - | ├─onehop_bert.py # Onehop bert model - | ├─process_data.py # Data preprocessing - | ├─twohop.py # Twohop model - | ├─twohop_bert.py # Twohop bert model - | └─utils.py # Utils for evaluation + | ├─build_reranker_data.py # build data for re-ranker from result of retriever + | ├─config.py # Evaluation configurations for retriever + | ├─hotpot_evaluate_v1.py # Hotpotqa evaluation script + | ├─onehop.py # Onehop model of retriever + | ├─onehop_bert.py # Onehop bert model of retriever + | ├─process_data.py # Data preprocessing for retriever + | ├─reader.py # Reader model + | ├─reader_albert_xxlarge.py # Albert-xxlarge module of reader model + | ├─reader_downstream.py # Downstream module of reader model + | ├─reader_eval.py # Reader evaluation script + | ├─rerank_albert_xxlarge.py # Albert-xxlarge module of re-ranker model + | ├─rerank_and_reader_data_generator.py # Data generator for re-ranker and reader + | ├─rerank_and_reader_utils.py # Utils for re-ranker and reader + | ├─rerank_downstream.py # Downstream module of re-ranker model + | ├─reranker.py # Re-ranker model + | ├─reranker_eval.py # Re-ranker evaluation script + | ├─twohop.py # Twohop model of retriever + | ├─twohop_bert.py # Twohop bert model of retriever + | └─utils.py # Utils for retriever | - └─retriever_eval.py # Evaluation net for retriever + ├─retriever_eval.py # Evaluation net for retriever + └─reranker_and_reader_eval.py # Evaluation net for re-ranker and reader ``` ## [Script Parameters](#contents) -Parameters for evaluation can be set in config.py. +Parameters for retriever evaluation can be set in config.py. -- config for TPRR retriever dataset +- config for TPRR retriever ```python "q_len": 64, # Max query length @@ -108,17 +126,30 @@ Parameters for evaluation can be set in config.py. config.py for more configuration. +Parameters for re-ranker and reader evaluation can be passed directly at execution time. + +- parameters for TPRR re-ranker and reader + + ```python + "seq_len": 512, # sequence length + "rerank_batch_size": 32, # batch size for re-ranker evaluation + "reader_batch_size": 448, # batch size for reader evaluation + "sp_threshold": 8 # threshold for picking supporting sentence + ``` + + config.py for more configuration. + ## [Evaluation Process](#contents) ### Evaluation -- Evaluation on Ascend +- Retriever evaluation on Ascend ```python sh run_eval_ascend.sh ``` - Evaluation result will be stored in the scripts path, whose folder name begins with "eval". You can find the result like the + Evaluation result will be stored in the scripts path, whose folder name begins with "eval_tr". You can find the result like the followings in log. ```python @@ -138,6 +169,35 @@ Parameters for evaluation can be set in config.py. evaluation time (h): 20.155506462653477 ``` +- Re-ranker and reader evaluation on Ascend + + Use the output of retriever as input of re-ranker + + ```python + sh run_eval_ascend_reranker_reader.sh + ``` + + Evaluation result will be stored in the scripts path, whose folder name begins with "eval". You can find the result like the + followings in log. + + ```python + total top1 pem: 0.8803511141120864 + + ... + em: 0.67440918298447 + f1: 0.8025625656569652 + prec: 0.8292800393689271 + recall: 0.8136908451841731 + sp_em: 0.6009453072248481 + sp_f1: 0.844555664157302 + sp_prec: 0.8640844345841021 + sp_recall: 0.8446123918845106 + joint_em: 0.4537474679270763 + joint_f1: 0.715119580346802 + joint_prec: 0.7540052057184267 + joint_recall: 0.7250240424067661 + ``` + # [Model Description](#contents) ## [Performance](#contents) @@ -154,6 +214,8 @@ Parameters for evaluation can be set in config.py. | Batch_size | 1 | | Output | inference path | | PEM | 0.9188 | +| total top1 pem | 0.88 | +| joint_f1 | 0.7151 | # [Description of random situation](#contents) diff --git a/model_zoo/research/nlp/tprr/reranker_and_reader_eval.py b/model_zoo/research/nlp/tprr/reranker_and_reader_eval.py new file mode 100644 index 0000000000..9724a52fa4 --- /dev/null +++ b/model_zoo/research/nlp/tprr/reranker_and_reader_eval.py @@ -0,0 +1,55 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""main file""" + +from mindspore import context +from src.rerank_and_reader_utils import get_parse, cal_reranker_metrics, select_reader_dev_data +from src.reranker_eval import rerank +from src.reader_eval import read +from src.hotpot_evaluate_v1 import hotpotqa_eval +from src.build_reranker_data import get_rerank_data + + +def rerank_and_retriever_eval(): + """main function""" + context.set_context(mode=context.GRAPH_MODE, device_target="Ascend") + parser = get_parse() + args = parser.parse_args() + + if args.get_reranker_data: + get_rerank_data(args) + + if args.run_reranker: + rerank(args) + + if args.cal_reranker_metrics: + total_top1_pem, _, _ = \ + cal_reranker_metrics(dev_gold_file=args.dev_gold_file, rerank_result_file=args.rerank_result_file) + print(f"total top1 pem: {total_top1_pem}") + + if args.select_reader_data: + select_reader_dev_data(args) + + if args.run_reader: + read(args) + + if args.cal_reader_metrics: + metrics = hotpotqa_eval(args.reader_result_file, args.dev_gold_file) + for k in metrics: + print(f"{k}: {metrics[k]}") + + +if __name__ == "__main__": + rerank_and_retriever_eval() diff --git a/model_zoo/research/nlp/tprr/scripts/run_eval_ascend.sh b/model_zoo/research/nlp/tprr/scripts/run_eval_ascend.sh index 4d1f2e5c1f..7d5d46b062 100644 --- a/model_zoo/research/nlp/tprr/scripts/run_eval_ascend.sh +++ b/model_zoo/research/nlp/tprr/scripts/run_eval_ascend.sh @@ -21,16 +21,16 @@ export DEVICE_NUM=1 export RANK_SIZE=$DEVICE_NUM export RANK_ID=0 -if [ -d "eval" ]; +if [ -d "eval_tr" ]; then - rm -rf ./eval + rm -rf ./eval_tr fi -mkdir ./eval +mkdir ./eval_tr -cp ../*.py ./eval -cp *.sh ./eval -cp -r ../src ./eval -cd ./eval || exit +cp ../*.py ./eval_tr +cp *.sh ./eval_tr +cp -r ../src ./eval_tr +cd ./eval_tr || exit env > env.log echo "start evaluation" diff --git a/model_zoo/research/nlp/tprr/scripts/run_eval_ascend_reranker_reader.sh b/model_zoo/research/nlp/tprr/scripts/run_eval_ascend_reranker_reader.sh new file mode 100644 index 0000000000..02435d5389 --- /dev/null +++ b/model_zoo/research/nlp/tprr/scripts/run_eval_ascend_reranker_reader.sh @@ -0,0 +1,39 @@ +#!/bin/bash +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ + +# eval script + +ulimit -u unlimited +export DEVICE_NUM=1 +export RANK_SIZE=$DEVICE_NUM +export RANK_ID=0 + +if [ -d "eval" ]; +then + rm -rf ./eval +fi +mkdir ./eval + +cp ../*.py ./eval +cp *.sh ./eval +cp -r ../src ./eval +cd ./eval || exit +env > env.log +echo "start evaluation" + +python reranker_and_reader_eval.py --get_reranker_data --run_reranker --cal_reranker_metrics --select_reader_data --run_reader --cal_reader_metrics > log_reranker_and_reader.txt 2>&1 & + +cd .. diff --git a/model_zoo/research/nlp/tprr/src/build_reranker_data.py b/model_zoo/research/nlp/tprr/src/build_reranker_data.py new file mode 100644 index 0000000000..47bc2ed6be --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/build_reranker_data.py @@ -0,0 +1,430 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""build reranker data from retriever result""" +import pickle +import gzip +from tqdm import tqdm + +from src.rerank_and_reader_utils import read_json, make_wiki_id, convert_text_to_tokens, normalize_title, \ + whitespace_tokenize, DocDB, _largest_valid_index, generate_mapping, InputFeatures, Example + +from transformers import AutoTokenizer + + +def judge_para(data): + """judge whether is valid para""" + for _, para_tokens in data["context"].items(): + if len(para_tokens) == 1: + return False + return True + + +def judge_sp(data, sent_name2id, para2id): + """judge whether is valid sp""" + for sp in data['sp']: + title = normalize_title(sp[0]) + name = normalize_title(sp[0]) + '_{}'.format(sp[1]) + if title in para2id and name not in sent_name2id: + return False + return True + + +def judge(path, path_set, reverse=False, golds=None, mode='or'): + """judge function""" + if path[0] == path[-1]: + return False + if path in path_set: + return False + if reverse and path[::-1] in path_set: + return False + if not golds: + return True + if mode == 'or': + return any(gold not in path for gold in golds) + if mode == 'and': + return all(gold not in path for gold in golds) + return False + + +def get_context_and_sents(path, doc_db): + """get context ans sentences""" + context = {} + sents = {} + for title in path: + para_info = doc_db.get_doc_info(title) + if title.endswith('_0'): + title = title[:-2] + context[title] = pickle.loads(para_info[1]) + sents[title] = pickle.loads(para_info[2]) + return context, sents + + +def gen_dev_data(dev_file, db_path, topk_file): + """generate dev data""" + # ----------------------------------------db info----------------------------------------------- + topk_data = read_json(topk_file) # path + doc_db = DocDB(db_path) # db get offset + print('load db successfully!') + + # ---------------------------------------------supervision ------------------------------------------ + dev_data = read_json(dev_file) + qid2sp = {} + qid2ans = {} + qid2type = {} + qid2path = {} + for _, data in enumerate(dev_data): + sp_facts = data['supporting_facts'] if 'supporting_facts' in data else None + qid2sp[data['_id']] = sp_facts + qid2ans[data['_id']] = data['answer'] if 'answer' in data else None + qid2type[data['_id']] = data['type'] if 'type' in data else None + qid2path[data['_id']] = list(set(list(zip(*sp_facts))[0])) if sp_facts else None + + new_dev_data = [] + + for _, data in enumerate(tqdm(topk_data)): + qid = data['q_id'] + question = data['question'] + topk_titles = data['topk_titles'] + gold_path = list(map(normalize_title, qid2path[qid])) if qid2path[qid] else None + + all_titles = [] + for titles in topk_titles: + titles = list(map(normalize_title, titles)) + if len(titles) == 1: + continue + path = titles[:2] + if judge(path, all_titles): + all_titles.append(titles[:2]) + if len(titles) == 3: + path = titles[1:] + if judge(path, all_titles): + all_titles.append(titles[1:]) + + # --------------------------------------------------process query----------------------------------- + + question = " ".join(whitespace_tokenize(question)) + question = question.strip() + q_tokens, _ = convert_text_to_tokens(question) + + gold_path = list(map(lambda x: make_wiki_id(x, 0), gold_path)) if gold_path else None + for path in all_titles: + context, sents = get_context_and_sents(path, doc_db) + ans_label = int(gold_path[0] in path and gold_path[1] in path) if gold_path else None + + new_dev_data.append({ + 'qid': qid, + 'type': qid2type[qid], + 'question': question, + 'q_tokens': q_tokens, + 'context': context, + 'sents': sents, + 'answer': qid2ans[qid], + 'sp': qid2sp[qid], + 'ans_para': None, + 'is_impossible': not ans_label == 1 + }) + + return new_dev_data + + +def read_hotpot_examples(path_data): + """reader examples""" + examples = [] + max_sent_cnt = 0 + failed = 0 + + for _, data in enumerate(path_data): + if not judge_para(data): + failed += 1 + continue + question = data['question'] + question = " ".join(whitespace_tokenize(question)) + question = question.strip() + path = list(map(normalize_title, data["context"].keys())) + qid = data['qid'] + q_tokens = data['q_tokens'] + + # -------------------------------------add para------------------------------------------------------------ + doc_tokens = [] + + para_start_end_position = [] + title_start_end_position = [] + sent_start_end_position = [] + + sent_names = [] + sent_name2id = {} + para2id = {} + + for para, para_tokens in data["context"].items(): + sents = data["sents"][para] + + para = normalize_title(para) + title_tokens = convert_text_to_tokens(para)[0] + para_node_id = len(para_start_end_position) + para2id[para] = para_node_id + + doc_offset = len(doc_tokens) + doc_tokens += title_tokens + doc_tokens += para_tokens + + title_start_end_position.append((doc_offset, doc_offset + len(title_tokens) - 1)) + + doc_offset += len(title_tokens) + para_start_end_position.append((doc_offset, doc_offset + len(para_tokens) - 1, para)) + + for idx, sent in enumerate(sents): + if sent[0] == -1 and sent[1] == -1: + continue + sent_names.append([para, idx]) # local name + sent_node_id = len(sent_start_end_position) + sent_name2id[normalize_title(para) + '_{}'.format(str(idx))] = sent_node_id + sent_start_end_position.append((doc_offset + sent[0], + doc_offset + sent[1])) + + # add sp and ans + sp_facts = [] + sup_fact_id = [] + for sp in sp_facts: + name = normalize_title(sp[0]) + '_{}'.format(sp[1]) + if name in sent_name2id: + sup_fact_id.append(sent_name2id[name]) + + sup_para_id = set() # use set + if sp_facts: + for para in list(zip(*sp_facts))[0]: + para = normalize_title(para) + if para in para2id: + sup_para_id.add(para2id[para]) + sup_para_id = list(sup_para_id) + + example = Example( + qas_id=qid, + path=path, + unique_id=qid + '_' + '_'.join(path), + question_tokens=q_tokens, + doc_tokens=doc_tokens, # multi-para tokens w/o query + sent_names=sent_names, + sup_fact_id=sup_fact_id, # global sent id + sup_para_id=sup_para_id, # global para id + para_start_end_position=para_start_end_position, + sent_start_end_position=sent_start_end_position, + title_start_end_position=title_start_end_position, + question_text=question) + + examples.append(example) + max_sent_cnt = max(max_sent_cnt, len(sent_start_end_position)) + + print(f"Maximum sentence cnt: {max_sent_cnt}") + print(f'failed examples: {failed}') + print(f'convert {len(examples)} examples successfully!') + + return examples + + +def add_sub_token(sub_tokens, idx, tok_to_orig_index, all_query_tokens): + """add sub tokens""" + for sub_token in sub_tokens: + tok_to_orig_index.append(idx) + all_query_tokens.append(sub_token) + return tok_to_orig_index, all_query_tokens + + +def get_sent_spans(example, orig_to_tok_index, orig_to_tok_back_index): + """get sentences' spans""" + sentence_spans = [] + for sent_span in example.sent_start_end_position: + sent_start_position = orig_to_tok_index[sent_span[0]] + sent_end_position = orig_to_tok_back_index[sent_span[1]] + sentence_spans.append((sent_start_position, sent_end_position + 1)) + return sentence_spans + + +def get_para_spans(example, orig_to_tok_index, orig_to_tok_back_index, all_doc_tokens, marker): + """get paragraphs' spans""" + para_spans = [] + for title_span, para_span in zip(example.title_start_end_position, example.para_start_end_position): + para_start_position = orig_to_tok_index[title_span[0]] + para_end_position = orig_to_tok_back_index[para_span[1]] + if para_end_position + 1 < len(all_doc_tokens) and all_doc_tokens[para_end_position + 1] == \ + marker['sent'][0]: + para_spans.append((para_start_position - 1, para_end_position + 1, para_span[2])) + else: + para_spans.append((para_start_position - 1, para_end_position, para_span[2])) + return para_spans + + +def build_feature(example, all_doc_tokens, doc_input_ids, doc_input_mask, doc_segment_ids, all_query_tokens, + query_input_ids, query_input_mask, query_segment_ids, para_spans, sentence_spans, tok_to_orig_index): + """build a input feature""" + feature = InputFeatures( + qas_id=example.qas_id, + path=example.path, + unique_id=example.qas_id + '_' + '_'.join(example.path), + sent_names=example.sent_names, + doc_tokens=all_doc_tokens, + doc_input_ids=doc_input_ids, + doc_input_mask=doc_input_mask, + doc_segment_ids=doc_segment_ids, + query_tokens=all_query_tokens, + query_input_ids=query_input_ids, + query_input_mask=query_input_mask, + query_segment_ids=query_segment_ids, + para_spans=para_spans, + sent_spans=sentence_spans, + token_to_orig_map=tok_to_orig_index) + return feature + + +def convert_example_to_features(tokenizer, args, examples): + """convert examples to features""" + features = [] + failed = 0 + marker = {'q': ['[q]', '[/q]'], 'para': ['', ''], 'sent': ['[s]']} + + for (_, example) in enumerate(tqdm(examples)): + + all_query_tokens = [tokenizer.cls_token, marker['q'][0]] + tok_to_orig_index = [-1, -1] # orig: query + doc tokens + ques_orig_to_tok_index = [] # start position + ques_orig_to_tok_back_index = [] # end position + q_spans = [] + + # -------------------------------------------for query--------------------------------------------- + for (idx, token) in enumerate(example.question_tokens): + sub_tokens = tokenizer.tokenize(token) + + ques_orig_to_tok_index.append(len(all_query_tokens)) + tok_to_orig_index, all_query_tokens = add_sub_token(sub_tokens, idx, tok_to_orig_index, all_query_tokens) + ques_orig_to_tok_back_index.append(len(all_query_tokens) - 1) + + all_query_tokens = all_query_tokens[:63] + tok_to_orig_index = tok_to_orig_index[:63] + all_query_tokens.append(marker['q'][-1]) + tok_to_orig_index.append(-1) + q_spans.append((1, len(all_query_tokens) - 1)) + + # ---------------------------------------add doc tokens------------------------------------------------ + all_doc_tokens = [] + orig_to_tok_index = [] # orig: token in doc + orig_to_tok_back_index = [] + title_start_mapping, title_end_mapping = generate_mapping(len(example.doc_tokens), + example.title_start_end_position) + _, sent_end_mapping = generate_mapping(len(example.doc_tokens), + example.sent_start_end_position) + all_doc_tokens += all_query_tokens + + for (idx, token) in enumerate(example.doc_tokens): + sub_tokens = tokenizer.tokenize(token) + + if title_start_mapping[idx] == 1: + all_doc_tokens.append(marker['para'][0]) + tok_to_orig_index.append(-1) + + # orig: position in doc tokens tok: global tokenized tokens (start) + orig_to_tok_index.append(len(all_doc_tokens)) + tok_to_orig_index, all_doc_tokens = add_sub_token(sub_tokens, idx + len(example.question_tokens), + tok_to_orig_index, all_doc_tokens) + orig_to_tok_back_index.append(len(all_doc_tokens) - 1) + + if title_end_mapping[idx] == 1: + all_doc_tokens.append(marker['para'][1]) + tok_to_orig_index.append(-1) + + if sent_end_mapping[idx] == 1: + all_doc_tokens.append(marker['sent'][0]) + tok_to_orig_index.append(-1) + + # -----------------------------------for sentence------------------------------------------------- + sentence_spans = get_sent_spans(example, orig_to_tok_index, orig_to_tok_back_index) + + # -----------------------------------for para------------------------------------------------------- + para_spans = get_para_spans(example, orig_to_tok_index, orig_to_tok_back_index, all_doc_tokens, marker) + + # -----------------------------------remove sent > max seq length----------------------------------------- + sent_max_index = _largest_valid_index(sentence_spans, args.seq_len) + max_sent_cnt = len(sentence_spans) + + if sent_max_index != len(sentence_spans): + if sent_max_index == 0: + failed += 0 + continue + sentence_spans = sentence_spans[:sent_max_index] + max_tok_length = sentence_spans[-1][1] # max_tok_length [s] + + # max end index: max_tok_length + para_max_index = _largest_valid_index(para_spans, max_tok_length + 1) + if para_max_index == 0: # only one para + failed += 0 + continue + if orig_to_tok_back_index[example.title_start_end_position[1][1]] + 1 >= max_tok_length: + failed += 0 + continue + max_para_span = para_spans[para_max_index] + para_spans = para_spans[:para_max_index] + para_spans.append((max_para_span[0], max_tok_length, max_para_span[2])) + + all_doc_tokens = all_doc_tokens[:max_tok_length + 1] + + sentence_spans = sentence_spans[:min(max_sent_cnt, args.max_sent_num)] + + # ----------------------------------------Padding Document----------------------------------------------------- + if len(all_doc_tokens) > args.seq_len: + st, _, title = para_spans[-1] + para_spans[-1] = (st, args.seq_len - 1, title) + all_doc_tokens = all_doc_tokens[:args.seq_len - 1] + [marker['sent'][0]] + + doc_input_ids = tokenizer.convert_tokens_to_ids(all_doc_tokens) + query_input_ids = tokenizer.convert_tokens_to_ids(all_query_tokens) + + doc_input_mask = [1] * len(doc_input_ids) + doc_segment_ids = [0] * len(query_input_ids) + [1] * (len(doc_input_ids) - len(query_input_ids)) + + doc_pad_length = args.seq_len - len(doc_input_ids) + doc_input_ids += [0] * doc_pad_length + doc_input_mask += [0] * doc_pad_length + doc_segment_ids += [0] * doc_pad_length + + # Padding Question + query_input_mask = [1] * len(query_input_ids) + query_segment_ids = [0] * len(query_input_ids) + + query_pad_length = 64 - len(query_input_ids) + query_input_ids += [0] * query_pad_length + query_input_mask += [0] * query_pad_length + query_segment_ids += [0] * query_pad_length + + feature = build_feature(example, all_doc_tokens, doc_input_ids, doc_input_mask, doc_segment_ids, + all_query_tokens, query_input_ids, query_input_mask, query_segment_ids, para_spans, + sentence_spans, tok_to_orig_index) + features.append(feature) + return features + + +def get_rerank_data(args): + """function for generating reranker's data""" + new_dev_data = gen_dev_data(dev_file=args.dev_gold_file, + db_path=args.wiki_db_file, + topk_file=args.retriever_result_file) + tokenizer = AutoTokenizer.from_pretrained(args.albert_model_path) + new_tokens = ['[q]', '[/q]', '', '', '[s]'] + tokenizer.add_tokens(new_tokens) + + examples = read_hotpot_examples(new_dev_data) + features = convert_example_to_features(tokenizer=tokenizer, args=args, examples=examples) + + with gzip.open(args.rerank_example_file, "wb") as f: + pickle.dump(examples, f) + with gzip.open(args.rerank_feature_file, "wb") as f: + pickle.dump(features, f) diff --git a/model_zoo/research/nlp/tprr/src/config.py b/model_zoo/research/nlp/tprr/src/config.py index 9af4e5d46f..edbe493968 100644 --- a/model_zoo/research/nlp/tprr/src/config.py +++ b/model_zoo/research/nlp/tprr/src/config.py @@ -33,14 +33,14 @@ def ThinkRetrieverConfig(): parser.add_argument("--batch_size", type=int, default=1, help="batch size") parser.add_argument("--device_id", type=int, default=0, help="device id") parser.add_argument("--save_name", type=str, default='doc_path', help='name of output') - parser.add_argument("--save_path", type=str, default='./', help='path of output') - parser.add_argument("--vocab_path", type=str, default='./scripts/vocab.txt', help="vocab path") - parser.add_argument("--wiki_path", type=str, default='./scripts/db_docs_bidirection_new.pkl', help="wiki path") - parser.add_argument("--dev_path", type=str, default='./scripts/hotpot_dev_fullwiki_v1_for_retriever.json', + parser.add_argument("--save_path", type=str, default='../', help='path of output') + parser.add_argument("--vocab_path", type=str, default='../vocab.txt', help="vocab path") + parser.add_argument("--wiki_path", type=str, default='../db_docs_bidirection_new.pkl', help="wiki path") + parser.add_argument("--dev_path", type=str, default='../hotpot_dev_fullwiki_v1_for_retriever.json', help="dev path") - parser.add_argument("--dev_data_path", type=str, default='./scripts/dev_tf_idf_data_raw.pkl', help="dev data path") - parser.add_argument("--onehop_bert_path", type=str, default='./scripts/onehop.ckpt', help="onehop bert ckpt path") - parser.add_argument("--onehop_mlp_path", type=str, default='./scripts/onehop_mlp.ckpt', help="onehop mlp ckpt path") - parser.add_argument("--twohop_bert_path", type=str, default='./scripts/twohop.ckpt', help="twohop bert ckpt path") - parser.add_argument("--twohop_mlp_path", type=str, default='./scripts/twohop_mlp.ckpt', help="twohop mlp ckpt path") + parser.add_argument("--dev_data_path", type=str, default='../dev_tf_idf_data_raw.pkl', help="dev data path") + parser.add_argument("--onehop_bert_path", type=str, default='../onehop.ckpt', help="onehop bert ckpt path") + parser.add_argument("--onehop_mlp_path", type=str, default='../onehop_mlp.ckpt', help="onehop mlp ckpt path") + parser.add_argument("--twohop_bert_path", type=str, default='../twohop.ckpt', help="twohop bert ckpt path") + parser.add_argument("--twohop_mlp_path", type=str, default='../twohop_mlp.ckpt', help="twohop mlp ckpt path") return parser.parse_args() diff --git a/model_zoo/research/nlp/tprr/src/hotpot_evaluate_v1.py b/model_zoo/research/nlp/tprr/src/hotpot_evaluate_v1.py new file mode 100644 index 0000000000..ecb25a0250 --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/hotpot_evaluate_v1.py @@ -0,0 +1,153 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""hotpotqa evaluate script""" + +import re +import string +from collections import Counter +import ujson as json + + +def normalize_answer(s): + """normalize answer""" + def remove_articles(text): + """remove articles""" + return re.sub(r'\b(a|an|the)\b', ' ', text) + + def white_space_fix(text): + """fix whitespace""" + return ' '.join(text.split()) + + def remove_punc(text): + """remove punctuation from text""" + exclude = set(string.punctuation) + return ''.join(ch for ch in text if ch not in exclude) + + def lower(text): + """lower text""" + return text.lower() + + return white_space_fix(remove_articles(remove_punc(lower(s)))) + + +def f1_score(prediction, ground_truth): + """calculate f1 score""" + normalized_prediction = normalize_answer(prediction) + normalized_ground_truth = normalize_answer(ground_truth) + + ZERO_METRIC = (0, 0, 0) + + if normalized_prediction in ['yes', 'no', 'noanswer'] and normalized_prediction != normalized_ground_truth: + return ZERO_METRIC + if normalized_ground_truth in ['yes', 'no', 'noanswer'] and normalized_prediction != normalized_ground_truth: + return ZERO_METRIC + + prediction_tokens = normalized_prediction.split() + ground_truth_tokens = normalized_ground_truth.split() + common = Counter(prediction_tokens) & Counter(ground_truth_tokens) + num_same = sum(common.values()) + if num_same == 0: + return ZERO_METRIC + precision = 1.0 * num_same / len(prediction_tokens) + recall = 1.0 * num_same / len(ground_truth_tokens) + f1 = (2 * precision * recall) / (precision + recall) + return f1, precision, recall + + +def exact_match_score(prediction, ground_truth): + """calculate exact match score""" + return normalize_answer(prediction) == normalize_answer(ground_truth) + + +def update_answer(metrics, prediction, gold): + """update answer""" + em = exact_match_score(prediction, gold) + f1, prec, recall = f1_score(prediction, gold) + metrics['em'] += float(em) + metrics['f1'] += f1 + metrics['prec'] += prec + metrics['recall'] += recall + return em, prec, recall + + +def update_sp(metrics, prediction, gold): + """update supporting sentences""" + cur_sp_pred = set(map(tuple, prediction)) + gold_sp_pred = set(map(tuple, gold)) + tp, fp, fn = 0, 0, 0 + for e in cur_sp_pred: + if e in gold_sp_pred: + tp += 1 + else: + fp += 1 + for e in gold_sp_pred: + if e not in cur_sp_pred: + fn += 1 + prec = 1.0 * tp / (tp + fp) if tp + fp > 0 else 0.0 + recall = 1.0 * tp / (tp + fn) if tp + fn > 0 else 0.0 + f1 = 2 * prec * recall / (prec + recall) if prec + recall > 0 else 0.0 + em = 1.0 if fp + fn == 0 else 0.0 + metrics['sp_em'] += em + metrics['sp_f1'] += f1 + metrics['sp_prec'] += prec + metrics['sp_recall'] += recall + return em, prec, recall + + +def hotpotqa_eval(prediction_file, gold_file): + """hotpotqa evaluate function""" + with open(prediction_file) as f: + prediction = json.load(f) + with open(gold_file) as f: + gold = json.load(f) + + metrics = {'em': 0, 'f1': 0, 'prec': 0, 'recall': 0, + 'sp_em': 0, 'sp_f1': 0, 'sp_prec': 0, 'sp_recall': 0, + 'joint_em': 0, 'joint_f1': 0, 'joint_prec': 0, 'joint_recall': 0} + for dp in gold: + cur_id = dp['_id'] + can_eval_joint = True + if cur_id not in prediction['answer']: + print('missing answer {}'.format(cur_id)) + can_eval_joint = False + else: + em, prec, recall = update_answer( + metrics, prediction['answer'][cur_id], dp['answer']) + if cur_id not in prediction['sp']: + print('missing sp fact {}'.format(cur_id)) + can_eval_joint = False + else: + sp_em, sp_prec, sp_recall = update_sp( + metrics, prediction['sp'][cur_id], dp['supporting_facts']) + + if can_eval_joint: + joint_prec = prec * sp_prec + joint_recall = recall * sp_recall + if joint_prec + joint_recall > 0: + joint_f1 = 2 * joint_prec * joint_recall / (joint_prec + joint_recall) + else: + joint_f1 = 0. + joint_em = em * sp_em + + metrics['joint_em'] += joint_em + metrics['joint_f1'] += joint_f1 + metrics['joint_prec'] += joint_prec + metrics['joint_recall'] += joint_recall + + num = len(gold) + for k in metrics: + metrics[k] /= num + + return metrics diff --git a/model_zoo/research/nlp/tprr/src/reader.py b/model_zoo/research/nlp/tprr/src/reader.py new file mode 100644 index 0000000000..73d6fe8ac7 --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/reader.py @@ -0,0 +1,73 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""Reader model""" + +import mindspore.nn as nn +from mindspore import load_checkpoint, load_param_into_net +from mindspore.ops import BatchMatMul +from mindspore import ops +from mindspore import dtype as mstype +from src.reader_albert_xxlarge import Reader_Albert +from src.reader_downstream import Reader_Downstream + + +dst_type = mstype.float16 +dst_type2 = mstype.float32 + + +class Reader(nn.Cell): + """Reader model""" + def __init__(self, batch_size, encoder_ck_file, downstream_ck_file): + """init function""" + super(Reader, self).__init__(auto_prefix=False) + + self.encoder = Reader_Albert(batch_size) + param_dict = load_checkpoint(encoder_ck_file) + not_load_params = load_param_into_net(self.encoder, param_dict) + print(f"not loaded: {not_load_params}") + + self.downstream = Reader_Downstream() + param_dict = load_checkpoint(downstream_ck_file) + not_load_params = load_param_into_net(self.downstream, param_dict) + print(f"not loaded: {not_load_params}") + + self.bmm = BatchMatMul() + + def construct(self, input_ids, attn_mask, token_type_ids, + context_mask, square_mask, packing_mask, cache_mask, + para_start_mapping, sent_end_mapping): + """construct function""" + state = self.encoder(attn_mask, input_ids, token_type_ids) + + para_state = self.bmm(ops.Cast()(para_start_mapping, dst_type), ops.Cast()(state, dst_type)) # [B, 2, D] + sent_state = self.bmm(ops.Cast()(sent_end_mapping, dst_type), ops.Cast()(state, dst_type)) # [B, max_sent, D] + + q_type, start, end, para_logit, sent_logit = self.downstream(ops.Cast()(para_state, dst_type2), + ops.Cast()(sent_state, dst_type2), + state, + context_mask) + + outer = start[:, :, None] + end[:, None] + + outer_mask = cache_mask + outer_mask = square_mask * outer_mask[None] + outer = outer - 1e30 * (1 - outer_mask) + outer = outer - 1e30 * packing_mask[:, :, None] + max_row = ops.ReduceMax()(outer, 2) + y1 = ops.Argmax()(max_row) + max_col = ops.ReduceMax()(outer, 1) + y2 = ops.Argmax()(max_col) + + return start, end, q_type, para_logit, sent_logit, y1, y2 diff --git a/model_zoo/research/nlp/tprr/src/reader_albert_xxlarge.py b/model_zoo/research/nlp/tprr/src/reader_albert_xxlarge.py new file mode 100644 index 0000000000..8eb7f75850 --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/reader_albert_xxlarge.py @@ -0,0 +1,263 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""albert-xxlarge Model for reader""" + +import numpy as np +from mindspore import nn, ops +from mindspore import Tensor, Parameter +from mindspore.ops import operations as P +from mindspore import dtype as mstype + +dst_type = mstype.float16 +dst_type2 = mstype.float32 + + +class LayerNorm(nn.Cell): + """LayerNorm layer""" + def __init__(self, mul_7_w_shape, add_8_bias_shape): + """init function""" + super(LayerNorm, self).__init__() + self.reducemean_0 = P.ReduceMean(keep_dims=True) + self.sub_1 = P.Sub() + self.pow_2 = P.Pow() + self.pow_2_input_weight = 2.0 + self.reducemean_3 = P.ReduceMean(keep_dims=True) + self.add_4 = P.Add() + self.add_4_bias = 9.999999960041972e-13 + self.sqrt_5 = P.Sqrt() + self.div_6 = P.Div() + self.mul_7 = P.Mul() + self.mul_7_w = Parameter(Tensor(np.random.uniform(0, 1, mul_7_w_shape).astype(np.float32)), name=None) + self.add_8 = P.Add() + self.add_8_bias = Parameter(Tensor(np.random.uniform(0, 1, add_8_bias_shape).astype(np.float32)), name=None) + + def construct(self, x): + """construct function""" + opt_reducemean_0 = self.reducemean_0(x, -1) + opt_sub_1 = self.sub_1(x, opt_reducemean_0) + opt_pow_2 = self.pow_2(opt_sub_1, self.pow_2_input_weight) + opt_reducemean_3 = self.reducemean_3(opt_pow_2, -1) + opt_add_4 = self.add_4(opt_reducemean_3, self.add_4_bias) + opt_sqrt_5 = self.sqrt_5(opt_add_4) + opt_div_6 = self.div_6(opt_sub_1, opt_sqrt_5) + opt_mul_7 = self.mul_7(opt_div_6, self.mul_7_w) + opt_add_8 = self.add_8(opt_mul_7, self.add_8_bias) + return opt_add_8 + + +class Linear(nn.Cell): + """Linear layer""" + def __init__(self, matmul_0_weight_shape, add_1_bias_shape): + """init function""" + super(Linear, self).__init__() + self.matmul_0 = nn.MatMul() + self.matmul_0_w = Parameter(Tensor(np.random.uniform(0, 1, matmul_0_weight_shape).astype(np.float32)), + name=None) + self.add_1 = P.Add() + self.add_1_bias = Parameter(Tensor(np.random.uniform(0, 1, add_1_bias_shape).astype(np.float32)), name=None) + + def construct(self, x): + """construct function""" + opt_matmul_0 = self.matmul_0(ops.Cast()(x, dst_type), ops.Cast()(self.matmul_0_w, dst_type)) + opt_add_1 = self.add_1(ops.Cast()(opt_matmul_0, dst_type2), self.add_1_bias) + return opt_add_1 + + +class MultiHeadAttn(nn.Cell): + """Multi-head attention layer""" + def __init__(self, batch_size): + """init function""" + super(MultiHeadAttn, self).__init__() + self.batch_size = batch_size + self.matmul_0 = nn.MatMul() + self.matmul_0_w = Parameter(Tensor(np.random.uniform(0, 1, (4096, 4096)).astype(np.float32)), name=None) + self.matmul_1 = nn.MatMul() + self.matmul_1_w = Parameter(Tensor(np.random.uniform(0, 1, (4096, 4096)).astype(np.float32)), name=None) + self.matmul_2 = nn.MatMul() + self.matmul_2_w = Parameter(Tensor(np.random.uniform(0, 1, (4096, 4096)).astype(np.float32)), name=None) + self.add_3 = P.Add() + self.add_3_bias = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.add_4 = P.Add() + self.add_4_bias = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.add_5 = P.Add() + self.add_5_bias = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.reshape_6 = P.Reshape() + self.reshape_6_shape = tuple([batch_size, 512, 64, 64]) + self.reshape_7 = P.Reshape() + self.reshape_7_shape = tuple([batch_size, 512, 64, 64]) + self.reshape_8 = P.Reshape() + self.reshape_8_shape = tuple([batch_size, 512, 64, 64]) + self.transpose_9 = P.Transpose() + self.transpose_10 = P.Transpose() + self.transpose_11 = P.Transpose() + self.matmul_12 = nn.MatMul() + self.div_13 = P.Div() + self.div_13_w = 8.0 + self.add_14 = P.Add() + self.softmax_15 = nn.Softmax(axis=3) + self.matmul_16 = nn.MatMul() + self.transpose_17 = P.Transpose() + self.matmul_18 = P.MatMul() + self.matmul_18_weight = Parameter(Tensor(np.random.uniform(0, 1, (64, 64, 4096)).astype(np.float32)), name=None) + self.add_19 = P.Add() + self.add_19_bias = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + + def construct(self, x, x0): + """construct function""" + opt_matmul_0 = self.matmul_0(ops.Cast()(x, dst_type), ops.Cast()(self.matmul_0_w, dst_type)) + opt_matmul_1 = self.matmul_1(ops.Cast()(x, dst_type), ops.Cast()(self.matmul_1_w, dst_type)) + opt_matmul_2 = self.matmul_2(ops.Cast()(x, dst_type), ops.Cast()(self.matmul_2_w, dst_type)) + opt_add_3 = self.add_3(ops.Cast()(opt_matmul_0, dst_type2), self.add_3_bias) + opt_add_4 = self.add_4(ops.Cast()(opt_matmul_1, dst_type2), self.add_4_bias) + opt_add_5 = self.add_5(ops.Cast()(opt_matmul_2, dst_type2), self.add_5_bias) + opt_reshape_6 = self.reshape_6(opt_add_3, self.reshape_6_shape) + opt_reshape_7 = self.reshape_7(opt_add_4, self.reshape_7_shape) + opt_reshape_8 = self.reshape_8(opt_add_5, self.reshape_8_shape) + opt_transpose_9 = self.transpose_9(opt_reshape_6, (0, 2, 1, 3)) + opt_transpose_10 = self.transpose_10(opt_reshape_7, (0, 2, 3, 1)) + opt_transpose_11 = self.transpose_11(opt_reshape_8, (0, 2, 1, 3)) + opt_matmul_12 = self.matmul_12(ops.Cast()(opt_transpose_9, dst_type), ops.Cast()(opt_transpose_10, dst_type)) + opt_div_13 = self.div_13(ops.Cast()(opt_matmul_12, dst_type2), ops.Cast()(self.div_13_w, dst_type2)) + opt_add_14 = self.add_14(opt_div_13, x0) + opt_softmax_15 = self.softmax_15(opt_add_14) + opt_matmul_16 = self.matmul_16(ops.Cast()(opt_softmax_15, dst_type), ops.Cast()(opt_transpose_11, dst_type)) + opt_transpose_17 = self.transpose_17(ops.Cast()(opt_matmul_16, dst_type2), (0, 2, 1, 3)) + opt_matmul_18 = self.matmul_18(ops.Cast()(opt_transpose_17, dst_type).view(self.batch_size * 512, -1), + ops.Cast()(self.matmul_18_weight, dst_type).view(-1, 4096))\ + .view(self.batch_size, 512, 4096) + + opt_add_19 = self.add_19(ops.Cast()(opt_matmul_18, dst_type2), self.add_19_bias) + return opt_add_19 + + +class NewGeLU(nn.Cell): + """new gelu layer""" + def __init__(self): + """init function""" + super(NewGeLU, self).__init__() + self.mul_0 = P.Mul() + self.mul_0_w = 0.5 + self.pow_1 = P.Pow() + self.pow_1_input_weight = 3.0 + self.mul_2 = P.Mul() + self.mul_2_w = 0.044714998453855515 + self.add_3 = P.Add() + self.mul_4 = P.Mul() + self.mul_4_w = 0.7978845834732056 + self.tanh_5 = nn.Tanh() + self.add_6 = P.Add() + self.add_6_bias = 1.0 + self.mul_7 = P.Mul() + + def construct(self, x): + """construct function""" + opt_mul_0 = self.mul_0(x, self.mul_0_w) + opt_pow_1 = self.pow_1(x, self.pow_1_input_weight) + opt_mul_2 = self.mul_2(opt_pow_1, self.mul_2_w) + opt_add_3 = self.add_3(x, opt_mul_2) + opt_mul_4 = self.mul_4(opt_add_3, self.mul_4_w) + opt_tanh_5 = self.tanh_5(opt_mul_4) + opt_add_6 = self.add_6(opt_tanh_5, self.add_6_bias) + opt_mul_7 = self.mul_7(opt_mul_0, opt_add_6) + return opt_mul_7 + + +class TransformerLayer(nn.Cell): + """Transformer layer""" + def __init__(self, batch_size, layernorm1_0_mul_7_w_shape, layernorm1_0_add_8_bias_shape, + linear3_0_matmul_0_weight_shape, linear3_0_add_1_bias_shape, linear3_1_matmul_0_weight_shape, + linear3_1_add_1_bias_shape): + """init function""" + super(TransformerLayer, self).__init__() + self.multiheadattn_0 = MultiHeadAttn(batch_size) + self.add_0 = P.Add() + self.layernorm1_0 = LayerNorm(mul_7_w_shape=layernorm1_0_mul_7_w_shape, + add_8_bias_shape=layernorm1_0_add_8_bias_shape) + self.linear3_0 = Linear(matmul_0_weight_shape=linear3_0_matmul_0_weight_shape, + add_1_bias_shape=linear3_0_add_1_bias_shape) + self.newgelu2_0 = NewGeLU() + self.linear3_1 = Linear(matmul_0_weight_shape=linear3_1_matmul_0_weight_shape, + add_1_bias_shape=linear3_1_add_1_bias_shape) + self.add_1 = P.Add() + + def construct(self, x, x0): + """construct function""" + multiheadattn_0_opt = self.multiheadattn_0(x, x0) + opt_add_0 = self.add_0(x, multiheadattn_0_opt) + layernorm1_0_opt = self.layernorm1_0(opt_add_0) + linear3_0_opt = self.linear3_0(layernorm1_0_opt) + newgelu2_0_opt = self.newgelu2_0(linear3_0_opt) + linear3_1_opt = self.linear3_1(newgelu2_0_opt) + opt_add_1 = self.add_1(linear3_1_opt, layernorm1_0_opt) + return opt_add_1 + + +class Reader_Albert(nn.Cell): + """Albert model for reader""" + def __init__(self, batch_size): + """init function""" + super(Reader_Albert, self).__init__() + self.expanddims_0 = P.ExpandDims() + self.expanddims_0_axis = 1 + self.expanddims_3 = P.ExpandDims() + self.expanddims_3_axis = 2 + self.cast_5 = P.Cast() + self.cast_5_to = mstype.float32 + self.sub_7 = P.Sub() + self.sub_7_bias = 1.0 + self.mul_9 = P.Mul() + self.mul_9_w = -10000.0 + self.gather_1_input_weight = Parameter(Tensor(np.random.uniform(0, 1, (30005, 128)).astype(np.float32)), + name=None) + self.gather_1_axis = 0 + self.gather_1 = P.Gather() + self.gather_2_input_weight = Parameter(Tensor(np.random.uniform(0, 1, (2, 128)).astype(np.float32)), name=None) + self.gather_2_axis = 0 + self.gather_2 = P.Gather() + self.add_4 = P.Add() + self.add_6 = P.Add() + self.add_6_bias = Parameter(Tensor(np.random.uniform(0, 1, (1, 512, 128)).astype(np.float32)), name=None) + self.layernorm1_0 = LayerNorm(mul_7_w_shape=(128,), add_8_bias_shape=(128,)) + self.linear3_0 = Linear(matmul_0_weight_shape=(128, 4096), add_1_bias_shape=(4096,)) + self.module34_0 = TransformerLayer(batch_size, + layernorm1_0_mul_7_w_shape=(4096,), + layernorm1_0_add_8_bias_shape=(4096,), + linear3_0_matmul_0_weight_shape=(4096, 16384), + linear3_0_add_1_bias_shape=(16384,), + linear3_1_matmul_0_weight_shape=(16384, 4096), + linear3_1_add_1_bias_shape=(4096,)) + self.layernorm1_1 = LayerNorm(mul_7_w_shape=(4096,), add_8_bias_shape=(4096,)) + + def construct(self, x, x0, x1): + """construct function""" + opt_expanddims_0 = self.expanddims_0(x, self.expanddims_0_axis) + opt_expanddims_3 = self.expanddims_3(opt_expanddims_0, self.expanddims_3_axis) + opt_cast_5 = self.cast_5(opt_expanddims_3, self.cast_5_to) + opt_sub_7 = self.sub_7(self.sub_7_bias, opt_cast_5) + opt_mul_9 = self.mul_9(opt_sub_7, self.mul_9_w) + opt_gather_1_axis = self.gather_1_axis + opt_gather_1 = self.gather_1(self.gather_1_input_weight, x0, opt_gather_1_axis) + opt_gather_2_axis = self.gather_2_axis + opt_gather_2 = self.gather_2(self.gather_2_input_weight, x1, opt_gather_2_axis) + opt_add_4 = self.add_4(opt_gather_1, opt_gather_2) + opt_add_6 = self.add_6(opt_add_4, self.add_6_bias) + layernorm1_0_opt = self.layernorm1_0(opt_add_6) + linear3_0_opt = self.linear3_0(layernorm1_0_opt) + module34_0_opt = self.module34_0(linear3_0_opt, opt_mul_9) + out = self.layernorm1_1(module34_0_opt) + for _ in range(11): + out = self.module34_0(out, opt_mul_9) + out = self.layernorm1_1(out) + return out diff --git a/model_zoo/research/nlp/tprr/src/reader_downstream.py b/model_zoo/research/nlp/tprr/src/reader_downstream.py new file mode 100644 index 0000000000..f971b0f522 --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/reader_downstream.py @@ -0,0 +1,213 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""downstream Model for reader""" + +import numpy as np +from mindspore import nn, ops +from mindspore import Tensor, Parameter +from mindspore.ops import operations as P +from mindspore import dtype as mstype + + +dst_type = mstype.float16 +dst_type2 = mstype.float32 + + +class Module15(nn.Cell): + """module of reader downstream""" + def __init__(self, matmul_0_weight_shape, add_1_bias_shape): + """init function""" + super(Module15, self).__init__() + self.matmul_0 = nn.MatMul() + self.matmul_0_w = Parameter(Tensor(np.random.uniform(0, 1, matmul_0_weight_shape).astype(np.float32)), + name=None) + self.add_1 = P.Add() + self.add_1_bias = Parameter(Tensor(np.random.uniform(0, 1, add_1_bias_shape).astype(np.float32)), name=None) + self.relu_2 = nn.ReLU() + + def construct(self, x): + """construct function""" + opt_matmul_0 = self.matmul_0(ops.Cast()(x, dst_type), ops.Cast()(self.matmul_0_w, dst_type)) + opt_add_1 = self.add_1(ops.Cast()(opt_matmul_0, dst_type2), self.add_1_bias) + opt_relu_2 = self.relu_2(opt_add_1) + return opt_relu_2 + + +class NormModule(nn.Cell): + """Normalization module of reader downstream""" + def __init__(self, mul_8_w_shape, add_9_bias_shape): + """init function""" + super(NormModule, self).__init__() + self.reducemean_0 = P.ReduceMean(keep_dims=True) + self.sub_1 = P.Sub() + self.sub_2 = P.Sub() + self.pow_3 = P.Pow() + self.pow_3_input_weight = 2.0 + self.reducemean_4 = P.ReduceMean(keep_dims=True) + self.add_5 = P.Add() + self.add_5_bias = 9.999999960041972e-13 + self.sqrt_6 = P.Sqrt() + self.div_7 = P.Div() + self.mul_8 = P.Mul() + self.mul_8_w = Parameter(Tensor(np.random.uniform(0, 1, mul_8_w_shape).astype(np.float32)), name=None) + self.add_9 = P.Add() + self.add_9_bias = Parameter(Tensor(np.random.uniform(0, 1, add_9_bias_shape).astype(np.float32)), name=None) + + def construct(self, x): + """construct function""" + opt_reducemean_0 = self.reducemean_0(x, -1) + opt_sub_1 = self.sub_1(x, opt_reducemean_0) + opt_sub_2 = self.sub_2(x, opt_reducemean_0) + opt_pow_3 = self.pow_3(opt_sub_1, self.pow_3_input_weight) + opt_reducemean_4 = self.reducemean_4(opt_pow_3, -1) + opt_add_5 = self.add_5(opt_reducemean_4, self.add_5_bias) + opt_sqrt_6 = self.sqrt_6(opt_add_5) + opt_div_7 = self.div_7(opt_sub_2, opt_sqrt_6) + opt_mul_8 = self.mul_8(self.mul_8_w, opt_div_7) + opt_add_9 = self.add_9(opt_mul_8, self.add_9_bias) + return opt_add_9 + + +class Module16(nn.Cell): + """module of reader downstream""" + def __init__(self, module15_0_matmul_0_weight_shape, module15_0_add_1_bias_shape, normmodule_0_mul_8_w_shape, + normmodule_0_add_9_bias_shape): + """init function""" + super(Module16, self).__init__() + self.module15_0 = Module15(matmul_0_weight_shape=module15_0_matmul_0_weight_shape, + add_1_bias_shape=module15_0_add_1_bias_shape) + self.normmodule_0 = NormModule(mul_8_w_shape=normmodule_0_mul_8_w_shape, + add_9_bias_shape=normmodule_0_add_9_bias_shape) + self.matmul_0 = nn.MatMul() + self.matmul_0_w = Parameter(Tensor(np.random.uniform(0, 1, (8192, 1)).astype(np.float32)), name=None) + + def construct(self, x): + """construct function""" + module15_0_opt = self.module15_0(x) + normmodule_0_opt = self.normmodule_0(module15_0_opt) + opt_matmul_0 = self.matmul_0(ops.Cast()(normmodule_0_opt, dst_type), ops.Cast()(self.matmul_0_w, dst_type)) + return ops.Cast()(opt_matmul_0, dst_type2) + + +class Module17(nn.Cell): + """module of reader downstream""" + def __init__(self, module15_0_matmul_0_weight_shape, module15_0_add_1_bias_shape, normmodule_0_mul_8_w_shape, + normmodule_0_add_9_bias_shape): + """init function""" + super(Module17, self).__init__() + self.module15_0 = Module15(matmul_0_weight_shape=module15_0_matmul_0_weight_shape, + add_1_bias_shape=module15_0_add_1_bias_shape) + self.normmodule_0 = NormModule(mul_8_w_shape=normmodule_0_mul_8_w_shape, + add_9_bias_shape=normmodule_0_add_9_bias_shape) + self.matmul_0 = nn.MatMul() + self.matmul_0_w = Parameter(Tensor(np.random.uniform(0, 1, (4096, 1)).astype(np.float32)), name=None) + self.add_1 = P.Add() + self.add_1_bias = Parameter(Tensor(np.random.uniform(0, 1, (1,)).astype(np.float32)), name=None) + + def construct(self, x): + """construct function""" + module15_0_opt = self.module15_0(x) + normmodule_0_opt = self.normmodule_0(module15_0_opt) + opt_matmul_0 = self.matmul_0(ops.Cast()(normmodule_0_opt, dst_type), ops.Cast()(self.matmul_0_w, dst_type)) + opt_add_1 = self.add_1(ops.Cast()(opt_matmul_0, dst_type2), self.add_1_bias) + return opt_add_1 + + +class Module5(nn.Cell): + """module of reader downstream""" + def __init__(self): + """init function""" + super(Module5, self).__init__() + self.sub_0 = P.Sub() + self.sub_0_bias = 1.0 + self.mul_1 = P.Mul() + self.mul_1_w = 1.0000000150474662e+30 + + def construct(self, x): + """construct function""" + opt_sub_0 = self.sub_0(self.sub_0_bias, x) + opt_mul_1 = self.mul_1(opt_sub_0, self.mul_1_w) + return opt_mul_1 + + +class Module10(nn.Cell): + """module of reader downstream""" + def __init__(self): + """init function""" + super(Module10, self).__init__() + self.squeeze_0 = P.Squeeze(2) + self.module5_0 = Module5() + self.sub_1 = P.Sub() + + def construct(self, x, x0): + """construct function""" + opt_squeeze_0 = self.squeeze_0(x) + module5_0_opt = self.module5_0(x0) + opt_sub_1 = self.sub_1(opt_squeeze_0, module5_0_opt) + return opt_sub_1 + + +class Reader_Downstream(nn.Cell): + """Downstream model for reader""" + def __init__(self): + """init function""" + super(Reader_Downstream, self).__init__() + self.module16_0 = Module16(module15_0_matmul_0_weight_shape=(4096, 8192), + module15_0_add_1_bias_shape=(8192,), + normmodule_0_mul_8_w_shape=(8192,), + normmodule_0_add_9_bias_shape=(8192,)) + self.add_74 = P.Add() + self.add_74_bias = Parameter(Tensor(np.random.uniform(0, 1, (1,)).astype(np.float32)), name=None) + self.module16_1 = Module16(module15_0_matmul_0_weight_shape=(4096, 8192), + module15_0_add_1_bias_shape=(8192,), + normmodule_0_mul_8_w_shape=(8192,), + normmodule_0_add_9_bias_shape=(8192,)) + self.add_75 = P.Add() + self.add_75_bias = Parameter(Tensor(np.random.uniform(0, 1, (1,)).astype(np.float32)), name=None) + self.module17_0 = Module17(module15_0_matmul_0_weight_shape=(4096, 4096), + module15_0_add_1_bias_shape=(4096,), + normmodule_0_mul_8_w_shape=(4096,), + normmodule_0_add_9_bias_shape=(4096,)) + self.module10_0 = Module10() + self.module17_1 = Module17(module15_0_matmul_0_weight_shape=(4096, 4096), + module15_0_add_1_bias_shape=(4096,), + normmodule_0_mul_8_w_shape=(4096,), + normmodule_0_add_9_bias_shape=(4096,)) + self.module10_1 = Module10() + self.gather_6_input_weight = Tensor(np.array(0)) + self.gather_6_axis = 1 + self.gather_6 = P.Gather() + self.dense_13 = nn.Dense(in_channels=4096, out_channels=4096, has_bias=True) + self.relu_18 = nn.ReLU() + self.normmodule_0 = NormModule(mul_8_w_shape=(4096,), add_9_bias_shape=(4096,)) + self.dense_73 = nn.Dense(in_channels=4096, out_channels=3, has_bias=True) + + def construct(self, x, x0, x1, x2): + """construct function""" + module16_0_opt = self.module16_0(x) + opt_add_74 = self.add_74(module16_0_opt, self.add_74_bias) + module16_1_opt = self.module16_1(x0) + opt_add_75 = self.add_75(module16_1_opt, self.add_75_bias) + module17_0_opt = self.module17_0(x1) + opt_module10_0 = self.module10_0(module17_0_opt, x2) + module17_1_opt = self.module17_1(x1) + opt_module10_1 = self.module10_1(module17_1_opt, x2) + opt_gather_6_axis = self.gather_6_axis + opt_gather_6 = self.gather_6(x1, self.gather_6_input_weight, opt_gather_6_axis) + opt_dense_13 = self.dense_13(opt_gather_6) + opt_relu_18 = self.relu_18(opt_dense_13) + normmodule_0_opt = self.normmodule_0(opt_relu_18) + opt_dense_73 = self.dense_73(normmodule_0_opt) + return opt_dense_73, opt_module10_0, opt_module10_1, opt_add_74, opt_add_75 diff --git a/model_zoo/research/nlp/tprr/src/reader_eval.py b/model_zoo/research/nlp/tprr/src/reader_eval.py new file mode 100644 index 0000000000..e39a9055fb --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/reader_eval.py @@ -0,0 +1,142 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""execute reader""" + +from collections import defaultdict +import random +from time import time +import json +from tqdm import tqdm +import numpy as np + +from transformers import AlbertTokenizer + +from mindspore import Tensor, ops +from mindspore import dtype as mstype + +from src.rerank_and_reader_data_generator import DataGenerator +from src.rerank_and_reader_utils import convert_to_tokens, make_wiki_id, DocDB +from src.reader import Reader + + +def read(args): + """reader function""" + db_file = args.wiki_db_file + reader_feature_file = args.reader_feature_file + reader_example_file = args.reader_example_file + encoder_ck_file = args.reader_encoder_ck_file + downstream_ck_file = args.reader_downstream_ck_file + albert_model_path = args.albert_model_path + reader_result_file = args.reader_result_file + seed = args.seed + sp_threshold = args.sp_threshold + seq_len = args.seq_len + batch_size = args.reader_batch_size + para_limit = args.max_para_num + sent_limit = args.max_sent_num + + random.seed(seed) + np.random.seed(seed) + + t1 = time() + + doc_db = DocDB(db_file) + + generator = DataGenerator(feature_file_path=reader_feature_file, + example_file_path=reader_example_file, + batch_size=batch_size, seq_len=seq_len, + para_limit=para_limit, sent_limit=sent_limit, + task_type="reader") + example_dict = generator.example_dict + feature_dict = generator.feature_dict + answer_dict = defaultdict(lambda: defaultdict(list)) + new_answer_dict = {} + total_sp_dict = defaultdict(list) + new_total_sp_dict = defaultdict(list) + + tokenizer = AlbertTokenizer.from_pretrained(albert_model_path) + new_tokens = ['[q]', '[/q]', '', '', '[s]'] + tokenizer.add_tokens(new_tokens) + + reader = Reader(batch_size=batch_size, + encoder_ck_file=encoder_ck_file, + downstream_ck_file=downstream_ck_file) + + print("start reading ...") + + for _, batch in tqdm(enumerate(generator)): + input_ids = Tensor(batch["context_idxs"], mstype.int32) + attn_mask = Tensor(batch["context_mask"], mstype.int32) + token_type_ids = Tensor(batch["segment_idxs"], mstype.int32) + context_mask = Tensor(batch["context_mask"], mstype.float32) + square_mask = Tensor(batch["square_mask"], mstype.float32) + packing_mask = Tensor(batch["query_mapping"], mstype.float32) + para_start_mapping = Tensor(batch["para_start_mapping"], mstype.float32) + sent_end_mapping = Tensor(batch["sent_end_mapping"], mstype.float32) + unique_ids = batch["unique_ids"] + sent_names = batch["sent_names"] + cache_mask = Tensor(np.tril(np.triu(np.ones((seq_len, seq_len)), 0), 30), mstype.float32) + + _, _, q_type, _, sent_logit, y1, y2 = reader(input_ids, attn_mask, token_type_ids, + context_mask, square_mask, packing_mask, cache_mask, + para_start_mapping, sent_end_mapping) + + type_prob = ops.Softmax()(q_type).asnumpy() + + answer_dict_ = convert_to_tokens(example_dict, + feature_dict, + batch['ids'], + y1.asnumpy().tolist(), + y2.asnumpy().tolist(), + type_prob, + tokenizer, + sent_logit.asnumpy(), + sent_names, + unique_ids) + for q_id in answer_dict_: + answer_dict[q_id] = answer_dict_[q_id] + + for q_id in answer_dict: + res = answer_dict[q_id] + answer_text_ = res[0] + sent_ = res[1] + sent_names_ = res[2] + new_answer_dict[q_id] = answer_text_ + + predict_support_np = ops.Sigmoid()(Tensor(sent_, mstype.float32)).asnumpy() + + for j in range(predict_support_np.shape[0]): + if j >= len(sent_names_): + break + if predict_support_np[j] > sp_threshold: + total_sp_dict[q_id].append(sent_names_[j]) + + for _id in total_sp_dict: + _sent_names = total_sp_dict[_id] + for para in _sent_names: + title = make_wiki_id(para[0], 0) + para_original_title = doc_db.get_doc_info(title)[-1] + para[0] = para_original_title + new_total_sp_dict[_id].append(para) + + prediction = {'answer': new_answer_dict, + 'sp': new_total_sp_dict} + + with open(reader_result_file, 'w') as f: + json.dump(prediction, f, indent=4) + + t2 = time() + + print(f"reader cost time: {t2-t1} s") diff --git a/model_zoo/research/nlp/tprr/src/rerank_albert_xxlarge.py b/model_zoo/research/nlp/tprr/src/rerank_albert_xxlarge.py new file mode 100644 index 0000000000..965b9b352a --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/rerank_albert_xxlarge.py @@ -0,0 +1,276 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""albert-xxlarge Model for reranker""" + +import numpy as np +from mindspore import nn, ops +from mindspore import Tensor, Parameter +from mindspore.ops import operations as P +from mindspore import dtype as mstype + + +dst_type = mstype.float16 +dst_type2 = mstype.float32 + + +class LayerNorm(nn.Cell): + """LayerNorm layer""" + def __init__(self, passthrough_w_0, passthrough_w_1): + """init function""" + super(LayerNorm, self).__init__() + self.reducemean_0 = P.ReduceMean(keep_dims=True) + self.sub_1 = P.Sub() + self.pow_2 = P.Pow() + self.pow_2_input_weight = 2.0 + self.reducemean_3 = P.ReduceMean(keep_dims=True) + self.add_4 = P.Add() + self.add_4_bias = 9.999999960041972e-13 + self.sqrt_5 = P.Sqrt() + self.div_6 = P.Div() + self.mul_7 = P.Mul() + self.mul_7_w = passthrough_w_0 + self.add_8 = P.Add() + self.add_8_bias = passthrough_w_1 + + def construct(self, x): + """construct function""" + opt_reducemean_0 = self.reducemean_0(x, -1) + opt_sub_1 = self.sub_1(x, opt_reducemean_0) + opt_pow_2 = self.pow_2(opt_sub_1, self.pow_2_input_weight) + opt_reducemean_3 = self.reducemean_3(opt_pow_2, -1) + opt_add_4 = self.add_4(opt_reducemean_3, self.add_4_bias) + opt_sqrt_5 = self.sqrt_5(opt_add_4) + opt_div_6 = self.div_6(opt_sub_1, opt_sqrt_5) + opt_mul_7 = self.mul_7(opt_div_6, self.mul_7_w) + opt_add_8 = self.add_8(opt_mul_7, self.add_8_bias) + return opt_add_8 + + +class Linear(nn.Cell): + """Linear layer""" + def __init__(self, matmul_0_w_shape, passthrough_w_0): + """init function""" + super(Linear, self).__init__() + self.matmul_0 = nn.MatMul() + self.matmul_0_w = Parameter(Tensor(np.random.uniform(0, 1, matmul_0_w_shape).astype(np.float32)), name=None) + self.add_1 = P.Add() + self.add_1_bias = passthrough_w_0 + + def construct(self, x): + """construct function""" + opt_matmul_0 = self.matmul_0(ops.Cast()(x, dst_type), ops.Cast()(self.matmul_0_w, dst_type)) + opt_add_1 = self.add_1(ops.Cast()(opt_matmul_0, dst_type2), self.add_1_bias) + return opt_add_1 + + +class MultiHeadAttn(nn.Cell): + """Multi-head attention layer""" + def __init__(self, batch_size, passthrough_w_0, passthrough_w_1, passthrough_w_2): + """init function""" + super(MultiHeadAttn, self).__init__() + self.batch_size = batch_size + self.matmul_0 = nn.MatMul() + self.matmul_0_w = Parameter(Tensor(np.random.uniform(0, 1, (4096, 4096)).astype(np.float32)), name=None) + self.matmul_1 = nn.MatMul() + self.matmul_1_w = Parameter(Tensor(np.random.uniform(0, 1, (4096, 4096)).astype(np.float32)), name=None) + self.matmul_2 = nn.MatMul() + self.matmul_2_w = Parameter(Tensor(np.random.uniform(0, 1, (4096, 4096)).astype(np.float32)), name=None) + self.add_3 = P.Add() + self.add_3_bias = passthrough_w_0 + self.add_4 = P.Add() + self.add_4_bias = passthrough_w_1 + self.add_5 = P.Add() + self.add_5_bias = passthrough_w_2 + self.reshape_6 = P.Reshape() + self.reshape_6_shape = tuple([batch_size, 512, 64, 64]) + self.reshape_7 = P.Reshape() + self.reshape_7_shape = tuple([batch_size, 512, 64, 64]) + self.reshape_8 = P.Reshape() + self.reshape_8_shape = tuple([batch_size, 512, 64, 64]) + self.transpose_9 = P.Transpose() + self.transpose_10 = P.Transpose() + self.transpose_11 = P.Transpose() + self.matmul_12 = nn.MatMul() + self.div_13 = P.Div() + self.div_13_w = 8.0 + self.add_14 = P.Add() + self.softmax_15 = nn.Softmax(axis=3) + self.matmul_16 = nn.MatMul() + self.transpose_17 = P.Transpose() + self.matmul_18 = P.MatMul() + self.matmul_18_weight = Parameter(Tensor(np.random.uniform(0, 1, (64, 64, 4096)).astype(np.float32)), name=None) + self.add_19 = P.Add() + self.add_19_bias = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + + def construct(self, x, x0): + """construct function""" + opt_matmul_0 = self.matmul_0(ops.Cast()(x, dst_type), ops.Cast()(self.matmul_0_w, dst_type)) + opt_matmul_1 = self.matmul_1(ops.Cast()(x, dst_type), ops.Cast()(self.matmul_1_w, dst_type)) + opt_matmul_2 = self.matmul_2(ops.Cast()(x, dst_type), ops.Cast()(self.matmul_2_w, dst_type)) + opt_add_3 = self.add_3(ops.Cast()(opt_matmul_0, dst_type2), self.add_3_bias) + opt_add_4 = self.add_4(ops.Cast()(opt_matmul_1, dst_type2), self.add_4_bias) + opt_add_5 = self.add_5(ops.Cast()(opt_matmul_2, dst_type2), self.add_5_bias) + opt_reshape_6 = self.reshape_6(opt_add_3, self.reshape_6_shape) + opt_reshape_7 = self.reshape_7(opt_add_4, self.reshape_7_shape) + opt_reshape_8 = self.reshape_8(opt_add_5, self.reshape_8_shape) + opt_transpose_9 = self.transpose_9(opt_reshape_6, (0, 2, 1, 3)) + opt_transpose_10 = self.transpose_10(opt_reshape_7, (0, 2, 3, 1)) + opt_transpose_11 = self.transpose_11(opt_reshape_8, (0, 2, 1, 3)) + opt_matmul_12 = self.matmul_12(ops.Cast()(opt_transpose_9, dst_type), ops.Cast()(opt_transpose_10, dst_type)) + opt_div_13 = self.div_13(ops.Cast()(opt_matmul_12, dst_type2), ops.Cast()(self.div_13_w, dst_type2)) + opt_add_14 = self.add_14(opt_div_13, x0) + opt_softmax_15 = self.softmax_15(opt_add_14) + opt_matmul_16 = self.matmul_16(ops.Cast()(opt_softmax_15, dst_type), ops.Cast()(opt_transpose_11, dst_type)) + opt_transpose_17 = self.transpose_17(ops.Cast()(opt_matmul_16, dst_type2), (0, 2, 1, 3)) + opt_matmul_18 = self.matmul_18(ops.Cast()(opt_transpose_17, dst_type).view(self.batch_size * 512, -1), + ops.Cast()(self.matmul_18_weight, dst_type).view(-1, 4096))\ + .view(self.batch_size, 512, 4096) + opt_add_19 = self.add_19(ops.Cast()(opt_matmul_18, dst_type2), self.add_19_bias) + return opt_add_19 + + +class NewGeLU(nn.Cell): + """Gelu layer""" + def __init__(self): + """init function""" + super(NewGeLU, self).__init__() + self.mul_0 = P.Mul() + self.mul_0_w = 0.5 + self.pow_1 = P.Pow() + self.pow_1_input_weight = 3.0 + self.mul_2 = P.Mul() + self.mul_2_w = 0.044714998453855515 + self.add_3 = P.Add() + self.mul_4 = P.Mul() + self.mul_4_w = 0.7978845834732056 + self.tanh_5 = nn.Tanh() + self.add_6 = P.Add() + self.add_6_bias = 1.0 + self.mul_7 = P.Mul() + + def construct(self, x): + """construct function""" + opt_mul_0 = self.mul_0(x, self.mul_0_w) + opt_pow_1 = self.pow_1(x, self.pow_1_input_weight) + opt_mul_2 = self.mul_2(opt_pow_1, self.mul_2_w) + opt_add_3 = self.add_3(x, opt_mul_2) + opt_mul_4 = self.mul_4(opt_add_3, self.mul_4_w) + opt_tanh_5 = self.tanh_5(opt_mul_4) + opt_add_6 = self.add_6(opt_tanh_5, self.add_6_bias) + opt_mul_7 = self.mul_7(opt_mul_0, opt_add_6) + return opt_mul_7 + + +class TransformerLayerWithLayerNorm(nn.Cell): + """Transformer layer with LayerNOrm""" + def __init__(self, batch_size, linear3_0_matmul_0_w_shape, linear3_1_matmul_0_w_shape, passthrough_w_0, + passthrough_w_1, passthrough_w_2, passthrough_w_3, passthrough_w_4, passthrough_w_5, passthrough_w_6): + """init function""" + super(TransformerLayerWithLayerNorm, self).__init__() + self.multiheadattn_0 = MultiHeadAttn(batch_size=batch_size, + passthrough_w_0=passthrough_w_0, + passthrough_w_1=passthrough_w_1, + passthrough_w_2=passthrough_w_2) + self.add_0 = P.Add() + self.layernorm1_0 = LayerNorm(passthrough_w_0=passthrough_w_3, passthrough_w_1=passthrough_w_4) + self.linear3_0 = Linear(matmul_0_w_shape=linear3_0_matmul_0_w_shape, passthrough_w_0=passthrough_w_5) + self.newgelu2_0 = NewGeLU() + self.linear3_1 = Linear(matmul_0_w_shape=linear3_1_matmul_0_w_shape, passthrough_w_0=passthrough_w_6) + self.add_1 = P.Add() + + def construct(self, x, x0): + """construct function""" + multiheadattn_0_opt = self.multiheadattn_0(x, x0) + opt_add_0 = self.add_0(x, multiheadattn_0_opt) + layernorm1_0_opt = self.layernorm1_0(opt_add_0) + linear3_0_opt = self.linear3_0(layernorm1_0_opt) + newgelu2_0_opt = self.newgelu2_0(linear3_0_opt) + linear3_1_opt = self.linear3_1(newgelu2_0_opt) + opt_add_1 = self.add_1(linear3_1_opt, layernorm1_0_opt) + return opt_add_1 + + +class Rerank_Albert(nn.Cell): + """Albert model for rerank""" + def __init__(self, batch_size): + """init function""" + super(Rerank_Albert, self).__init__() + self.passthrough_w_0 = Parameter(Tensor(np.random.uniform(0, 1, (128,)).astype(np.float32)), name=None) + self.passthrough_w_1 = Parameter(Tensor(np.random.uniform(0, 1, (128,)).astype(np.float32)), name=None) + self.passthrough_w_2 = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.passthrough_w_3 = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.passthrough_w_4 = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.passthrough_w_5 = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.passthrough_w_6 = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.passthrough_w_7 = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.passthrough_w_8 = Parameter(Tensor(np.random.uniform(0, 1, (16384,)).astype(np.float32)), name=None) + self.passthrough_w_9 = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.passthrough_w_10 = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.passthrough_w_11 = Parameter(Tensor(np.random.uniform(0, 1, (4096,)).astype(np.float32)), name=None) + self.expanddims_0 = P.ExpandDims() + self.expanddims_0_axis = 1 + self.expanddims_3 = P.ExpandDims() + self.expanddims_3_axis = 2 + self.cast_5 = P.Cast() + self.cast_5_to = mstype.float32 + self.sub_7 = P.Sub() + self.sub_7_bias = 1.0 + self.mul_9 = P.Mul() + self.mul_9_w = -10000.0 + self.gather_1_input_weight = Parameter(Tensor(np.random.uniform(0, 1, (30005, 128)).astype(np.float32)), + name=None) + self.gather_1_axis = 0 + self.gather_1 = P.Gather() + self.gather_2_input_weight = Parameter(Tensor(np.random.uniform(0, 1, (2, 128)).astype(np.float32)), name=None) + self.gather_2_axis = 0 + self.gather_2 = P.Gather() + self.add_4 = P.Add() + self.add_4_bias = Parameter(Tensor(np.random.uniform(0, 1, (1, 512, 128)).astype(np.float32)), name=None) + self.add_6 = P.Add() + self.layernorm1_0 = LayerNorm(passthrough_w_0=self.passthrough_w_0, passthrough_w_1=self.passthrough_w_1) + self.linear3_0 = Linear(matmul_0_w_shape=(128, 4096), passthrough_w_0=self.passthrough_w_2) + self.module34_0 = TransformerLayerWithLayerNorm(batch_size=batch_size, + linear3_0_matmul_0_w_shape=(4096, 16384), + linear3_1_matmul_0_w_shape=(16384, 4096), + passthrough_w_0=self.passthrough_w_3, + passthrough_w_1=self.passthrough_w_4, + passthrough_w_2=self.passthrough_w_5, + passthrough_w_3=self.passthrough_w_6, + passthrough_w_4=self.passthrough_w_7, + passthrough_w_5=self.passthrough_w_8, + passthrough_w_6=self.passthrough_w_9) + self.layernorm1_1 = LayerNorm(passthrough_w_0=self.passthrough_w_10, passthrough_w_1=self.passthrough_w_11) + + def construct(self, input_ids, attention_mask, token_type_ids): + """construct function""" + opt_expanddims_0 = self.expanddims_0(attention_mask, self.expanddims_0_axis) + opt_expanddims_3 = self.expanddims_3(opt_expanddims_0, self.expanddims_3_axis) + opt_cast_5 = self.cast_5(opt_expanddims_3, self.cast_5_to) + opt_sub_7 = self.sub_7(self.sub_7_bias, opt_cast_5) + opt_mul_9 = self.mul_9(opt_sub_7, self.mul_9_w) + opt_gather_1_axis = self.gather_1_axis + opt_gather_1 = self.gather_1(self.gather_1_input_weight, input_ids, opt_gather_1_axis) + opt_gather_2_axis = self.gather_2_axis + opt_gather_2 = self.gather_2(self.gather_2_input_weight, token_type_ids, opt_gather_2_axis) + opt_add_4 = self.add_4(opt_gather_1, self.add_4_bias) + opt_add_6 = self.add_6(opt_add_4, opt_gather_2) + layernorm1_0_opt = self.layernorm1_0(opt_add_6) + linear3_0_opt = self.linear3_0(layernorm1_0_opt) + opt = self.module34_0(linear3_0_opt, opt_mul_9) + opt = self.layernorm1_1(opt) + for _ in range(11): + opt = self.module34_0(opt, opt_mul_9) + opt = self.layernorm1_1(opt) + return opt diff --git a/model_zoo/research/nlp/tprr/src/rerank_and_reader_data_generator.py b/model_zoo/research/nlp/tprr/src/rerank_and_reader_data_generator.py new file mode 100644 index 0000000000..fd9d83921e --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/rerank_and_reader_data_generator.py @@ -0,0 +1,183 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""define a data generator""" + +import gzip +import pickle +import random +import numpy as np + + +random.seed(42) +np.random.seed(42) + + +class DataGenerator: + """data generator for reranker and reader""" + def __init__(self, feature_file_path, example_file_path, batch_size, seq_len, + para_limit=None, sent_limit=None, task_type=None): + """init function""" + self.example_ptr = 0 + self.bsz = batch_size + self.seq_length = seq_len + self.para_limit = para_limit + self.sent_limit = sent_limit + self.task_type = task_type + + self.feature_file_path = feature_file_path + self.example_file_path = example_file_path + self.features = self.load_features() + self.examples = self.load_examples() + self.feature_dict = self.get_feature_dict() + self.example_dict = self.get_example_dict() + + self.features = self.padding_feature(self.features, self.bsz) + + def load_features(self): + """load features from feature file""" + with gzip.open(self.feature_file_path, 'rb') as fin: + features = pickle.load(fin) + print("load features successful !!!") + return features + + def padding_feature(self, features, bsz): + """padding features as multiples of batch size""" + padding_num = ((len(features) // bsz + 1) * bsz - len(features)) + print(f"features padding num is {padding_num}") + new_features = features + features[:padding_num] + return new_features + + def load_examples(self): + """laod examples from file""" + if self.example_file_path: + with gzip.open(self.example_file_path, 'rb') as fin: + examples = pickle.load(fin) + print("load examples successful !!!") + return examples + return {} + + def get_feature_dict(self): + """build a feature dict""" + return {f.unique_id: f for f in self.features} + + def get_example_dict(self): + """build a example dict""" + if self.example_file_path: + return {e.unique_id: e for e in self.examples} + return {} + + def common_process_single_case(self, i, case, context_idxs, context_mask, segment_idxs, ids, path, unique_ids): + """common process for a single case""" + context_idxs[i] = np.array(case.doc_input_ids) + context_mask[i] = np.array(case.doc_input_mask) + segment_idxs[i] = np.array(case.doc_segment_ids) + + ids.append(case.qas_id) + path.append(case.path) + unique_ids.append(case.unique_id) + + return context_idxs, context_mask, segment_idxs, ids, path, unique_ids + + def reader_process_single_case(self, i, case, sent_names, square_mask, query_mapping, ques_start_mapping, + para_start_mapping, sent_end_mapping): + """process for a single case about reader""" + sent_names.append(case.sent_names) + prev_position = None + for cur_position, token_id in enumerate(case.doc_input_ids): + if token_id >= 30000: + if prev_position: + square_mask[i, prev_position + 1: cur_position, prev_position + 1: cur_position] = 1.0 + prev_position = cur_position + if case.sent_spans: + for j in range(case.sent_spans[0][0] - 1): + query_mapping[i, j] = 1 + ques_start_mapping[i, 0, 1] = 1 + for j, para_span in enumerate(case.para_spans[:self.para_limit]): + start, end, _ = para_span + if start <= end: + para_start_mapping[i, j, start] = 1 + for j, sent_span in enumerate(case.sent_spans[:self.sent_limit]): + start, end = sent_span + if start <= end: + end = min(end, self.seq_length - 1) + sent_end_mapping[i, j, end] = 1 + return sent_names, square_mask, query_mapping, ques_start_mapping, para_start_mapping, sent_end_mapping + + def __iter__(self): + """iteration function""" + while True: + if self.example_ptr >= len(self.features): + break + start_id = self.example_ptr + cur_bsz = min(self.bsz, len(self.features) - start_id) + cur_batch = self.features[start_id: start_id + cur_bsz] + # BERT input + context_idxs = np.zeros((cur_bsz, self.seq_length)) + context_mask = np.zeros((cur_bsz, self.seq_length)) + segment_idxs = np.zeros((cur_bsz, self.seq_length)) + + # others + ids = [] + path = [] + unique_ids = [] + + if self.task_type == "reader": + # Mappings + ques_start_mapping = np.zeros((cur_bsz, 1, self.seq_length)) + query_mapping = np.zeros((cur_bsz, self.seq_length)) + para_start_mapping = np.zeros((cur_bsz, self.para_limit, self.seq_length)) + sent_end_mapping = np.zeros((cur_bsz, self.sent_limit, self.seq_length)) + square_mask = np.zeros((cur_bsz, self.seq_length, self.seq_length)) + sent_names = [] + + for i, case in enumerate(cur_batch): + context_idxs, context_mask, segment_idxs, ids, path, unique_ids = \ + self.common_process_single_case(i, case, context_idxs, context_mask, segment_idxs, ids, path, + unique_ids) + if self.task_type == "reader": + sent_names, square_mask, query_mapping, ques_start_mapping, para_start_mapping, sent_end_mapping = \ + self.reader_process_single_case(i, case, sent_names, square_mask, query_mapping, + ques_start_mapping, para_start_mapping, sent_end_mapping) + + self.example_ptr += cur_bsz + + if self.task_type == "reranker": + yield { + "context_idxs": context_idxs, + "context_mask": context_mask, + "segment_idxs": segment_idxs, + + "ids": ids, + "unique_ids": unique_ids, + "path": path + } + elif self.task_type == "reader": + yield { + "context_idxs": context_idxs, + "context_mask": context_mask, + "segment_idxs": segment_idxs, + "query_mapping": query_mapping, + "para_start_mapping": para_start_mapping, + "sent_end_mapping": sent_end_mapping, + "square_mask": square_mask, + "ques_start_mapping": ques_start_mapping, + + "ids": ids, + "unique_ids": unique_ids, + "sent_names": sent_names, + "path": path + } + else: + print(f"data generator received a error type: {self.task_type} !!!") diff --git a/model_zoo/research/nlp/tprr/src/rerank_and_reader_utils.py b/model_zoo/research/nlp/tprr/src/rerank_and_reader_utils.py new file mode 100644 index 0000000000..bd4dba3762 --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/rerank_and_reader_utils.py @@ -0,0 +1,656 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""utils""" + +import re +import argparse +from urllib.parse import unquote +from collections import defaultdict +import collections +import logging +import unicodedata +import json +import gzip +import string +import pickle +import sqlite3 +from tqdm import tqdm + +import numpy as np +from transformers import BasicTokenizer + + +logger = logging.getLogger(__name__) + + +class Example: + """A single example of data""" + def __init__(self, + qas_id, + path, + unique_id, + question_tokens, + doc_tokens, + sent_names, + sup_fact_id, + sup_para_id, + para_start_end_position, + sent_start_end_position, + question_text, + title_start_end_position=None): + """init function""" + self.qas_id = qas_id + self.path = path + self.unique_id = unique_id + self.question_tokens = question_tokens + self.doc_tokens = doc_tokens + self.question_text = question_text + self.sent_names = sent_names + self.sup_fact_id = sup_fact_id + self.sup_para_id = sup_para_id + self.para_start_end_position = para_start_end_position + self.sent_start_end_position = sent_start_end_position + self.title_start_end_position = title_start_end_position + + +class InputFeatures: + """A single set of features of data.""" + + def __init__(self, + unique_id, + qas_id, + path, + sent_names, + doc_tokens, + doc_input_ids, + doc_input_mask, + doc_segment_ids, + query_tokens, + query_input_ids, + query_input_mask, + query_segment_ids, + para_spans, + sent_spans, + token_to_orig_map): + """init function""" + self.qas_id = qas_id + self.doc_tokens = doc_tokens + self.doc_input_ids = doc_input_ids + self.doc_input_mask = doc_input_mask + self.doc_segment_ids = doc_segment_ids + self.path = path + self.unique_id = unique_id + self.sent_names = sent_names + + self.query_tokens = query_tokens + self.query_input_ids = query_input_ids + self.query_input_mask = query_input_mask + self.query_segment_ids = query_segment_ids + + self.para_spans = para_spans + self.sent_spans = sent_spans + + self.token_to_orig_map = token_to_orig_map + + +class DocDB: + """ + Sqlite backed document storage. + Implements get_doc_text(doc_id). + """ + + def __init__(self, db_path): + """init function""" + self.path = db_path + self.connection = sqlite3.connect(self.path, check_same_thread=False) + + def __enter__(self): + """enter function""" + return self + + def __exit__(self, *args): + """exit function""" + self.close() + + def close(self): + """Close the connection to the database.""" + self.connection.close() + + def get_doc_ids(self): + """Fetch all ids of docs stored in the db.""" + cursor = self.connection.cursor() + cursor.execute("SELECT id FROM documents") + results = [r[0] for r in cursor.fetchall()] + cursor.close() + return results + + def get_doc_info(self, doc_id): + """get docment information""" + if not doc_id.endswith('_0'): + doc_id += '_0' + cursor = self.connection.cursor() + cursor.execute( + "SELECT * FROM documents WHERE id = ?", + (normalize_title(doc_id),) + ) + result = cursor.fetchall() + cursor.close() + return result if result is None else result[0] + + +def get_parse(): + """get parse function""" + parser = argparse.ArgumentParser() + + # Environment + parser.add_argument('--seed', type=int, default=42, + help="random seed for initialization") + parser.add_argument('--seq_len', type=int, default=512, + help="max sentence length") + parser.add_argument("--get_reranker_data", + action='store_true', + help="Set this flag if you want to get reranker data from retrieved result") + parser.add_argument("--run_reranker", + action='store_true', + help="Set this flag if you want to run reranker") + parser.add_argument("--cal_reranker_metrics", + action='store_true', + help="Set this flag if you want to calculate rerank metrics") + parser.add_argument("--select_reader_data", + action='store_true', + help="Set this flag if you want to select reader data") + parser.add_argument("--run_reader", + action='store_true', + help="Set this flag if you want to run reader") + parser.add_argument("--cal_reader_metrics", + action='store_true', + help="Set this flag if you want to calculate reader metrics") + parser.add_argument('--dev_gold_file', + type=str, + default="../hotpot_dev_fullwiki_v1.json", + help='file of dev ground truth') + parser.add_argument('--wiki_db_file', + type=str, + default="../enwiki_offset.db", + help='wiki_database_file') + parser.add_argument('--albert_model_path', + type=str, + default="../albert-xxlarge/", + help='model path of huggingface albert-xxlarge') + + # Retriever + parser.add_argument('--retriever_result_file', + type=str, + default="../doc_path", + help='file of retriever result') + + # Rerank + parser.add_argument('--rerank_batch_size', type=int, default=32, + help="rerank batchsize for evaluating") + parser.add_argument('--rerank_feature_file', + type=str, + default="../reranker_feature_file.pkl.gz", + help='file of rerank feature') + parser.add_argument('--rerank_example_file', + type=str, + default="../reranker_example_file.pkl.gz", + help='file of rerank example') + parser.add_argument('--rerank_result_file', + type=str, + default="../rerank_result.json", + help='file of rerank result') + parser.add_argument('--rerank_encoder_ck_file', + type=str, + default="../rerank_albert_12.ckpt", + help='checkpoint of rerank albert-xxlarge') + parser.add_argument('--rerank_downstream_ck_file', + type=str, + default="../rerank_downstream.ckpt", + help='checkpoint of rerank downstream') + + # Reader + parser.add_argument('--reader_batch_size', type=int, default=32, + help="reader batchsize for evaluating") + parser.add_argument('--reader_feature_file', + type=str, + default="../reader_feature_file.pkl.gz", + help='file of reader feature') + parser.add_argument('--reader_example_file', + type=str, + default="../reader_example_file.pkl.gz", + help='file of reader example') + parser.add_argument('--reader_encoder_ck_file', + type=str, + default="../albert_12_layer.ckpt", + help='checkpoint of reader albert-xxlarge') + parser.add_argument('--reader_downstream_ck_file', + type=str, + default="../reader_downstream.ckpt", + help='checkpoint of reader downstream') + parser.add_argument('--reader_result_file', + type=str, + default="../reader_result_file.json", + help='file of reader result') + parser.add_argument('--sp_threshold', type=float, default=0.65, + help="threshold for selecting supporting sentences") + parser.add_argument("--max_para_num", default=2, type=int) + parser.add_argument("--max_sent_num", default=40, type=int) + + return parser + + +def select_reader_dev_data(args): + """select reader dev data from result of retriever based on result of reranker""" + rerank_result_file = args.rerank_result_file + rerank_feature_file = args.rerank_feature_file + rerank_example_file = args.rerank_example_file + reader_feature_file = args.reader_feature_file + reader_example_file = args.reader_example_file + + with gzip.open(rerank_example_file, "rb") as f: + dev_examples = pickle.load(f) + with gzip.open(rerank_feature_file, "rb") as f: + dev_features = pickle.load(f) + with open(rerank_result_file, "r") as f: + rerank_result = json.load(f) + + new_dev_examples = [] + new_dev_features = [] + + rerank_unique_ids = defaultdict(int) + feature_unique_ids = defaultdict(int) + + for _, res in tqdm(rerank_result.items(), desc="get rerank unique ids"): + rerank_unique_ids[res[0]] = True + print(f"rerank result num is {len(rerank_unique_ids)}") + + for feature in tqdm(dev_features, desc="select rerank top1 feature"): + if feature.unique_id in rerank_unique_ids: + feature_unique_ids[feature.unique_id] = True + new_dev_features.append(feature) + print(f"new feature num is {len(new_dev_features)}") + + for example in tqdm(dev_examples, desc="select rerank top1 example"): + if example.unique_id in rerank_unique_ids and example.unique_id in feature_unique_ids: + new_dev_examples.append(example) + print(f"new examples num is {len(new_dev_examples)}") + + print("start save new examples ......") + with gzip.open(reader_example_file, "wb") as f: + pickle.dump(new_dev_examples, f) + + print("start save new features ......") + with gzip.open(reader_feature_file, "wb") as f: + pickle.dump(new_dev_features, f) + print("finish selecting reader data !!!") + + +def get_final_text(pred_text, orig_text, do_lower_case, verbose_logging=False): + """Project the tokenized prediction back to the original text.""" + def _strip_spaces(text): + ns_chars = [] + ns_to_s_map = collections.OrderedDict() + for (i, c) in enumerate(text): + if c == " ": + continue + ns_to_s_map[len(ns_chars)] = i + ns_chars.append(c) + ns_text = "".join(ns_chars) + return (ns_text, ns_to_s_map) + + tokenizer = BasicTokenizer(do_lower_case=do_lower_case) + + tok_text = " ".join(tokenizer.tokenize(orig_text)) + + start_position = tok_text.find(pred_text) + if start_position == -1: + if verbose_logging: + print("Unable to find text: '%s' in '%s'" % (pred_text, orig_text)) + return orig_text + end_position = start_position + len(pred_text) - 1 + + (orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text) + (tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text) + + if len(orig_ns_text) != len(tok_ns_text): + if verbose_logging: + logger.info("Length not equal after stripping spaces: '%s' vs '%s'", + orig_ns_text, tok_ns_text) + return orig_text + + # We then project the characters in `pred_text` back to `orig_text` using + # the character-to-character alignment. + tok_s_to_ns_map = {} + for (i, tok_index) in tok_ns_to_s_map.items(): + tok_s_to_ns_map[tok_index] = i + + orig_start_position = None + if start_position in tok_s_to_ns_map: + ns_start_position = tok_s_to_ns_map[start_position] + if ns_start_position in orig_ns_to_s_map: + orig_start_position = orig_ns_to_s_map[ns_start_position] + + if orig_start_position is None: + if verbose_logging: + print("Couldn't map start position") + return orig_text + + orig_end_position = None + if end_position in tok_s_to_ns_map: + ns_end_position = tok_s_to_ns_map[end_position] + if ns_end_position in orig_ns_to_s_map: + orig_end_position = orig_ns_to_s_map[ns_end_position] + + if orig_end_position is None: + if verbose_logging: + print("Couldn't map end position") + return orig_text + + output_text = orig_text[orig_start_position:(orig_end_position + 1)] + return output_text + + +def get_ans_from_pos(tokenizer, examples, features, y1, y2, unique_id): + """get answer text from predicted position""" + feature = features[unique_id] + example = examples[unique_id] + tok_to_orig_map = feature.token_to_orig_map + orig_all_tokens = example.question_tokens + example.doc_tokens + + final_text = " " + if y1 < len(tok_to_orig_map) and y2 < len(tok_to_orig_map): + orig_tok_start = tok_to_orig_map[y1] + orig_tok_end = tok_to_orig_map[y2] + # -----------------orig all tokens----------------------------------- + orig_tokens = orig_all_tokens[orig_tok_start: (orig_tok_end + 1)] + tok_tokens = feature.doc_tokens[y1: y2 + 1] + tok_text = tokenizer.convert_tokens_to_string(tok_tokens) + # Clean whitespace + tok_text = tok_text.strip() + tok_text = " ".join(tok_text.split()) + orig_text = " ".join(orig_tokens) + final_text = get_final_text(tok_text, orig_text, True, False) + # print("final_text: " + final_text) + return final_text + + +def convert_to_tokens(examples, features, ids, y1, y2, q_type_prob, tokenizer, sent, sent_names, + unique_ids): + """get raw answer text and supporting sentences""" + answer_dict = defaultdict(list) + + q_type = np.argmax(q_type_prob, 1) + + for i, qid in enumerate(ids): + unique_id = unique_ids[i] + + if q_type[i] == 0: + answer_text = 'yes' + elif q_type[i] == 1: + answer_text = 'no' + elif q_type[i] == 2: + answer_text = get_ans_from_pos(tokenizer, examples, features, y1[i], y2[i], unique_id) + else: + raise ValueError("question type error") + + answer_dict[qid].append(answer_text) + answer_dict[qid].append(sent[i]) + answer_dict[qid].append(sent_names[i]) + + return answer_dict + + +def normalize_title(text): + """Resolve different type of unicode encodings / capitarization in HotpotQA data.""" + text = unicodedata.normalize('NFD', text) + return text[0].capitalize() + text[1:] + + +def make_wiki_id(title, para_index): + """make wiki id""" + title_id = "{0}_{1}".format(normalize_title(title), para_index) + return title_id + + +def cal_reranker_metrics(dev_gold_file, rerank_result_file): + """function for calculating reranker's metrics""" + with open(dev_gold_file, 'rb') as f: + gt = json.load(f) + with open(rerank_result_file, 'rb') as f: + rerank_result = json.load(f) + + cnt = 0 + all_ = len(gt) + + cnt_c = 0 + cnt_b = 0 + all_c = 0 + all_b = 0 + + for item in tqdm(gt, desc="get com and bridge "): + q_type = item["type"] + if q_type == "comparison": + all_c += 1 + elif q_type == "bridge": + all_b += 1 + else: + print(f"{q_type} is a error question type!!!") + + for item in tqdm(gt, desc="cal pem"): + _id = item["_id"] + + if _id in rerank_result: + pred = rerank_result[_id][1] + sps = item["supporting_facts"] + q_type = item["type"] + gold = [] + for t in sps: + gold.append(normalize_title(t[0])) + gold = set(gold) + flag = True + for t in gold: + if t not in pred: + flag = False + break + if flag: + cnt += 1 + if q_type == "comparison": + cnt_c += 1 + elif q_type == "bridge": + cnt_b += 1 + else: + print(f"{q_type} is a error question type!!!") + + return cnt/all_, cnt_c/all_c, cnt_b/all_b + + +def whitespace_tokenize(text): + """Runs basic whitespace cleaning and splitting on a piece of text.""" + text = text.strip() + if not text: + return [] + tokens = text.split() + return tokens + + +def find_hyper_linked_titles(text_w_links): + """find hyperlinked titles""" + titles = re.findall(r'href=[\'"]?([^\'" >]+)', text_w_links) + titles = [unquote(title) for title in titles] + titles = [title[0].capitalize() + title[1:] for title in titles] + return titles + + +def normalize_text(text): + """Resolve different type of unicode encodings / capitarization in HotpotQA data.""" + text = unicodedata.normalize('NFD', text) + return text + + +def convert_char_to_token_offset(orig_text, start_offset, end_offset, char_to_word_offset, doc_tokens): + """build characters' offset""" + length = len(orig_text) + assert start_offset + length == end_offset + assert end_offset <= len(char_to_word_offset) + + start_position = char_to_word_offset[start_offset] + end_position = char_to_word_offset[start_offset + length - 1] + + actual_text = " ".join( + doc_tokens[start_position:(end_position + 1)]) + + assert actual_text.lower().find(orig_text.lower()) != -1 + return start_position, end_position + + +def _is_whitespace(c): + """check whitespace""" + if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F: + return True + return False + + +def convert_text_to_tokens(context_text, return_word_start=False): + """convert text to tokens""" + doc_tokens = [] + char_to_word_offset = [] + words_start_idx = [] + prev_is_whitespace = True + + for idx, c in enumerate(context_text): + if _is_whitespace(c): + prev_is_whitespace = True + else: + if prev_is_whitespace: + doc_tokens.append(c) + words_start_idx.append(idx) + else: + doc_tokens[-1] += c + prev_is_whitespace = False + char_to_word_offset.append(len(doc_tokens) - 1) + + if not return_word_start: + return doc_tokens, char_to_word_offset + return doc_tokens, char_to_word_offset, words_start_idx + + +def read_json(eval_file_name): + """reader json files""" + print("loading examples from {0}".format(eval_file_name)) + with open(eval_file_name) as reader: + lines = json.load(reader) + return lines + + +def write_json(data, out_file_name): + """write json files""" + print("writing {0} examples to {1}".format(len(data), out_file_name)) + with open(out_file_name, 'w') as writer: + json.dump(data, writer, indent=4) + + +def get_edges(sentence): + """get edges""" + EDGE_XY = re.compile(r'(.*?)<\/a>') + ret = EDGE_XY.findall(sentence) + return [(unquote(x), y) for x, y in ret] + + +def relocate_tok_span(orig_to_tok_index, orig_to_tok_back_index, word_tokens, subword_tokens, + orig_start_position, orig_end_position, orig_text, tokenizer, tok_to_orig_index=None): + """relocate tokens' span""" + if orig_start_position is None: + return 0, 0 + + tok_start_position = orig_to_tok_index[orig_start_position] + if tok_start_position >= len(subword_tokens): + return 0, 0 + + if orig_end_position < len(word_tokens) - 1: + tok_end_position = orig_to_tok_back_index[orig_end_position] + if tok_to_orig_index and tok_to_orig_index[tok_end_position + 1] == -1: + assert tok_end_position <= orig_to_tok_index[orig_end_position + 1] - 2 + else: + assert tok_end_position == orig_to_tok_index[orig_end_position + 1] - 1 + else: + tok_end_position = orig_to_tok_back_index[orig_end_position] + return _improve_answer_span( + subword_tokens, tok_start_position, tok_end_position, tokenizer, orig_text) + + +def generate_mapping(length, positions): + """generate mapping""" + start_mapping = [0] * length + end_mapping = [0] * length + for _, (start, end) in enumerate(positions): + start_mapping[start] = 1 + end_mapping[end] = 1 + return start_mapping, end_mapping + + +def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer, orig_answer_text): + """Returns tokenized answer spans that better match the annotated answer.""" + tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text, add_prefix_space=True)) + + for new_start in range(input_start, input_end + 1): + for new_end in range(input_end, new_start - 1, -1): + text_span = " ".join(doc_tokens[new_start: (new_end + 1)]) + if text_span == tok_answer_text: + return new_start, new_end + + return input_start, input_end + + +def _largest_valid_index(spans, limit): + """return largest valid index""" + for idx, _ in enumerate(spans): + if spans[idx][1] >= limit: + return idx + return len(spans) + + +def remove_punc(text): + """remove punctuation""" + if text == " ": + return '' + exclude = set(string.punctuation) + return ''.join(ch for ch in text if ch not in exclude) + + +def check_text_include_ans(ans, text): + """check whether text include answer""" + if normalize_answer(ans) in normalize_answer(text): + return True + return False + + +def remove_articles(text): + """remove articles""" + return re.sub(r'\b(a|an|the)\b', ' ', text) + + +def white_space_fix(text): + """fix whitespace""" + return ' '.join(text.split()) + + +def lower(text): + """lower text""" + return text.lower() + + +def normalize_answer(s): + """Lower text and remove punctuation, articles and extra whitespace.""" + return white_space_fix(remove_articles(remove_punc(lower(s)))) diff --git a/model_zoo/research/nlp/tprr/src/rerank_downstream.py b/model_zoo/research/nlp/tprr/src/rerank_downstream.py new file mode 100644 index 0000000000..29bc41a7a5 --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/rerank_downstream.py @@ -0,0 +1,61 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""downstream Model for reranker""" + +import numpy as np +from mindspore import nn +from mindspore import Tensor, Parameter +from mindspore.ops import operations as P + + +class Rerank_Downstream(nn.Cell): + """Downstream model for rerank""" + def __init__(self): + """init function""" + super(Rerank_Downstream, self).__init__() + self.dense_0 = nn.Dense(in_channels=4096, out_channels=8192, has_bias=True) + self.relu_1 = nn.ReLU() + self.reducemean_2 = P.ReduceMean(keep_dims=True) + self.sub_3 = P.Sub() + self.sub_4 = P.Sub() + self.pow_5 = P.Pow() + self.pow_5_input_weight = 2.0 + self.reducemean_6 = P.ReduceMean(keep_dims=True) + self.add_7 = P.Add() + self.add_7_bias = 9.999999960041972e-13 + self.sqrt_8 = P.Sqrt() + self.div_9 = P.Div() + self.mul_10 = P.Mul() + self.mul_10_w = Parameter(Tensor(np.random.uniform(0, 1, (8192,)).astype(np.float32)), name=None) + self.add_11 = P.Add() + self.add_11_bias = Parameter(Tensor(np.random.uniform(0, 1, (8192,)).astype(np.float32)), name=None) + self.dense_12 = nn.Dense(in_channels=8192, out_channels=2, has_bias=True) + + def construct(self, x): + """construct function""" + opt_dense_0 = self.dense_0(x) + opt_relu_1 = self.relu_1(opt_dense_0) + opt_reducemean_2 = self.reducemean_2(opt_relu_1, -1) + opt_sub_3 = self.sub_3(opt_relu_1, opt_reducemean_2) + opt_sub_4 = self.sub_4(opt_relu_1, opt_reducemean_2) + opt_pow_5 = self.pow_5(opt_sub_3, self.pow_5_input_weight) + opt_reducemean_6 = self.reducemean_6(opt_pow_5, -1) + opt_add_7 = self.add_7(opt_reducemean_6, self.add_7_bias) + opt_sqrt_8 = self.sqrt_8(opt_add_7) + opt_div_9 = self.div_9(opt_sub_4, opt_sqrt_8) + opt_mul_10 = self.mul_10(self.mul_10_w, opt_div_9) + opt_add_11 = self.add_11(opt_mul_10, self.add_11_bias) + opt_dense_12 = self.dense_12(opt_add_11) + return opt_dense_12 diff --git a/model_zoo/research/nlp/tprr/src/reranker.py b/model_zoo/research/nlp/tprr/src/reranker.py new file mode 100644 index 0000000000..ef732b696d --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/reranker.py @@ -0,0 +1,45 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""Reranker Model""" + +import mindspore.nn as nn +from mindspore import load_checkpoint, load_param_into_net +from src.rerank_albert_xxlarge import Rerank_Albert +from src.rerank_downstream import Rerank_Downstream + + +class Reranker(nn.Cell): + """Reranker model""" + def __init__(self, batch_size, encoder_ck_file, downstream_ck_file): + """init function""" + super(Reranker, self).__init__(auto_prefix=False) + + self.encoder = Rerank_Albert(batch_size) + param_dict = load_checkpoint(encoder_ck_file) + not_load_params_1 = load_param_into_net(self.encoder, param_dict) + print(f"not loaded albert: {not_load_params_1}") + + self.no_answer_mlp = Rerank_Downstream() + param_dict = load_checkpoint(downstream_ck_file) + not_load_params_2 = load_param_into_net(self.no_answer_mlp, param_dict) + print(f"not loaded downstream: {not_load_params_2}") + + def construct(self, input_ids, attn_mask, token_type_ids): + """construct function""" + state = self.encoder(input_ids, attn_mask, token_type_ids) + state = state[:, 0, :] + + no_answer = self.no_answer_mlp(state) + return no_answer diff --git a/model_zoo/research/nlp/tprr/src/reranker_eval.py b/model_zoo/research/nlp/tprr/src/reranker_eval.py new file mode 100644 index 0000000000..4c2d89451d --- /dev/null +++ b/model_zoo/research/nlp/tprr/src/reranker_eval.py @@ -0,0 +1,85 @@ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""execute reranker""" + +import json +import random +from collections import defaultdict +from time import time +from tqdm import tqdm +import numpy as np + +from mindspore import Tensor, ops +from mindspore import dtype as mstype + +from src.rerank_and_reader_data_generator import DataGenerator +from src.reranker import Reranker + + +def rerank(args): + """rerank function""" + rerank_feature_file = args.rerank_feature_file + rerank_result_file = args.rerank_result_file + encoder_ck_file = args.rerank_encoder_ck_file + downstream_ck_file = args.rerank_downstream_ck_file + seed = args.seed + seq_len = args.seq_len + batch_size = args.rerank_batch_size + + random.seed(seed) + np.random.seed(seed) + + t1 = time() + + generator = DataGenerator(feature_file_path=rerank_feature_file, + example_file_path=None, + batch_size=batch_size, seq_len=seq_len, + task_type="reranker") + gather_dict = defaultdict(lambda: defaultdict(list)) + + reranker = Reranker(batch_size=batch_size, + encoder_ck_file=encoder_ck_file, + downstream_ck_file=downstream_ck_file) + + print("start re-ranking ...") + + for _, batch in tqdm(enumerate(generator)): + input_ids = Tensor(batch["context_idxs"], mstype.int32) + attn_mask = Tensor(batch["context_mask"], mstype.int32) + token_type_ids = Tensor(batch["segment_idxs"], mstype.int32) + + no_answer = reranker(input_ids, attn_mask, token_type_ids) + + no_answer_prob = ops.Softmax()(no_answer).asnumpy() + no_answer_prob = no_answer_prob[:, 0] + + for i in range(len(batch['ids'])): + qas_id = batch['ids'][i] + gather_dict[qas_id][no_answer_prob[i]].append(batch['unique_ids'][i]) + gather_dict[qas_id][no_answer_prob[i]].append(batch['path'][i]) + + rerank_result = {} + for qas_id in tqdm(gather_dict, desc="get top1 path from re-rank result"): + all_paths = gather_dict[qas_id] + all_paths = sorted(all_paths.items(), key=lambda item: item[0]) + assert qas_id not in rerank_result + rerank_result[qas_id] = all_paths[0][1] + + with open(rerank_result_file, 'w') as f: + json.dump(rerank_result, f) + + t2 = time() + + print(f"re-rank cost time: {t2-t1} s")