From 951077c729b71f5d245f42dcec82a7a65a028a96 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E6=99=BA=E4=B8=9E?= Date: Thu, 16 Jun 2022 20:23:15 +0800 Subject: [PATCH] registry multi models on model and pipeline --- .../models/nlp/masked_language_model.py | 17 ++-- modelscope/pipelines/builder.py | 4 +- .../pipelines/nlp/fill_mask_pipeline.py | 29 +++--- tests/pipelines/test_fill_mask.py | 98 +++++++++++-------- 4 files changed, 87 insertions(+), 61 deletions(-) diff --git a/modelscope/models/nlp/masked_language_model.py b/modelscope/models/nlp/masked_language_model.py index 76677040..cb12a4dd 100644 --- a/modelscope/models/nlp/masked_language_model.py +++ b/modelscope/models/nlp/masked_language_model.py @@ -1,14 +1,16 @@ from typing import Any, Dict, Optional, Union import numpy as np + +from ...utils.constant import Tasks from ..base import Model, Tensor from ..builder import MODELS -from ...utils.constant import Tasks __all__ = ['MaskedLanguageModel'] @MODELS.register_module(Tasks.fill_mask, module_name=r'sbert') +@MODELS.register_module(Tasks.fill_mask, module_name=r'veco') class MaskedLanguageModel(Model): def __init__(self, model_dir: str, *args, **kwargs): @@ -17,8 +19,8 @@ class MaskedLanguageModel(Model): super().__init__(model_dir, *args, **kwargs) self.config = AutoConfig.from_pretrained(model_dir) - self.model = AutoModelForMaskedLM.from_pretrained(model_dir, config=self.config) - + self.model = AutoModelForMaskedLM.from_pretrained( + model_dir, config=self.config) def forward(self, inputs: Dict[str, Tensor]) -> Dict[str, np.ndarray]: """return the result by the model @@ -35,9 +37,8 @@ class MaskedLanguageModel(Model): 'logits': array([[-0.53860897, 1.5029076 ]], dtype=float32) # true value } """ - rst = self.model( - input_ids=inputs["input_ids"], - attention_mask=inputs['attention_mask'], - token_type_ids=inputs["token_type_ids"] - ) + rst = self.model( + input_ids=inputs['input_ids'], + attention_mask=inputs['attention_mask'], + token_type_ids=inputs['token_type_ids']) return {'logits': rst['logits'], 'input_ids': inputs['input_ids']} diff --git a/modelscope/pipelines/builder.py b/modelscope/pipelines/builder.py index 38172c99..dc7b6aa6 100644 --- a/modelscope/pipelines/builder.py +++ b/modelscope/pipelines/builder.py @@ -24,8 +24,8 @@ DEFAULT_MODEL_FOR_PIPELINE = { Tasks.image_generation: ('person-image-cartoon', 'damo/cv_unet_person-image-cartoon_compound-models'), - Tasks.fill_mask: - ('sbert') + Tasks.fill_mask: ('sbert', 'damo/nlp_structbert_fill-mask_chinese-large'), + Tasks.fill_mask: ('veco', 'damo/nlp_veco_fill-mask_large') } diff --git a/modelscope/pipelines/nlp/fill_mask_pipeline.py b/modelscope/pipelines/nlp/fill_mask_pipeline.py index ac22cf3b..aebf4e09 100644 --- a/modelscope/pipelines/nlp/fill_mask_pipeline.py +++ b/modelscope/pipelines/nlp/fill_mask_pipeline.py @@ -10,6 +10,7 @@ __all__ = ['FillMaskPipeline'] @PIPELINES.register_module(Tasks.fill_mask, module_name=r'sbert') +@PIPELINES.register_module(Tasks.fill_mask, module_name=r'veco') class FillMaskPipeline(Pipeline): def __init__(self, model: MaskedLanguageModel, @@ -36,22 +37,28 @@ class FillMaskPipeline(Pipeline): Dict[str, str]: the prediction results """ import numpy as np - logits = inputs["logits"].detach().numpy() - input_ids = inputs["input_ids"].detach().numpy() + logits = inputs['logits'].detach().numpy() + input_ids = inputs['input_ids'].detach().numpy() pred_ids = np.argmax(logits, axis=-1) - rst_ids = np.where(input_ids==self.mask_id[self.model.config.model_type], pred_ids, input_ids) + rst_ids = np.where( + input_ids == self.mask_id[self.model.config.model_type], pred_ids, + input_ids) pred_strings = [] for ids in rst_ids: if self.model.config.model_type == 'veco': - pred_string = self.tokenizer.decode(ids).split('')[0].replace("", "").replace("", "").replace("", "") - elif self.model.config.vocab_size == 21128: # zh bert + pred_string = self.tokenizer.decode(ids).split( + '')[0].replace('', + '').replace('', + '').replace('', '') + elif self.model.config.vocab_size == 21128: # zh bert pred_string = self.tokenizer.convert_ids_to_tokens(ids) - pred_string = ''.join(pred_string).replace('##','') - pred_string = pred_string.split('[SEP]')[0].replace('[CLS]', '').replace('[SEP]', '').replace('[UNK]', '') - else: ## en bert + pred_string = ''.join(pred_string).replace('##', '') + pred_string = pred_string.split('[SEP]')[0].replace( + '[CLS]', '').replace('[SEP]', '').replace('[UNK]', '') + else: ## en bert pred_string = self.tokenizer.decode(ids) - pred_string = pred_string.split('[SEP]')[0].replace('[CLS]', '').replace('[SEP]', '').replace('[UNK]', '') + pred_string = pred_string.split('[SEP]')[0].replace( + '[CLS]', '').replace('[SEP]', '').replace('[UNK]', '') pred_strings.append(pred_string) - return {'pred_string': pred_strings} - + return {'pred_string': pred_strings} diff --git a/tests/pipelines/test_fill_mask.py b/tests/pipelines/test_fill_mask.py index 12f7ff3b..8b021c03 100644 --- a/tests/pipelines/test_fill_mask.py +++ b/tests/pipelines/test_fill_mask.py @@ -5,24 +5,41 @@ import unittest from maas_hub.snapshot_download import snapshot_download +from modelscope.models import Model from modelscope.models.nlp import MaskedLanguageModel from modelscope.pipelines import FillMaskPipeline, pipeline from modelscope.preprocessors import FillMaskPreprocessor from modelscope.utils.constant import Tasks -from modelscope.models import Model from modelscope.utils.hub import get_model_cache_dir from modelscope.utils.test_utils import test_level + class FillMaskTest(unittest.TestCase): - model_id_sbert = {'zh': 'damo/nlp_structbert_fill-mask-chinese_large', - 'en': 'damo/nlp_structbert_fill-mask-english_large'} + model_id_sbert = { + 'zh': 'damo/nlp_structbert_fill-mask-chinese_large', + 'en': 'damo/nlp_structbert_fill-mask-english_large' + } model_id_veco = 'damo/nlp_veco_fill-mask_large' - ori_texts = {"zh": "段誉轻挥折扇，摇了摇头，说道：“你师父是你的师父，你师父可不是我的师父。你师父差得动你，你师父可差不动我。", - "en": "Everything in what you call reality is really just a reflection of your consciousness. Your whole universe is just a mirror reflection of your story."} + ori_texts = { + 'zh': + f'段誉轻挥折扇，摇了摇头，说道：“你师父是你的师父，你师父可不是我的师父。' + f'你师父差得动你，你师父可差不动我。', + 'en': + f'Everything in what you call reality is really just a r' + f'eflection of your consciousness. Your whole universe is' + f'just a mirror reflection of your story.' + } - test_inputs = {"zh": "段誉轻[MASK]折扇，摇了摇[MASK]，[MASK]道：“你师父是你的[MASK][MASK]，你师父可不是[MASK]的师父。你师父差得动你，你师父可[MASK]不动我。", - "en": "Everything in [MASK] you call reality is really [MASK] a reflection of your [MASK]. Your whole universe is just a mirror [MASK] of your story."} + test_inputs = { + 'zh': + f'段誉轻[MASK]折扇，摇了摇[MASK]，[MASK]道：“你师父是你的[MASK][MASK]' + f'，你师父可不是[MASK]的师父。你师父差得动你，你师父可[MASK]不动我。', + 'en': + f'Everything in [MASK] you call reality is really [MASK] a ' + f'reflection of your [MASK]. Your whole universe is just a ' + f'mirror [MASK] of your story.' + } #def test_run(self): # # sbert @@ -37,51 +54,52 @@ class FillMaskTest(unittest.TestCase): # ori_text = self.ori_texts[language] # test_input = self.test_inputs[language] # print( - # f'ori_text: {ori_text}\ninput: {test_input}\npipeline1: {pipeline1(test_input)}\npipeline2: {pipeline2(test_input)}' + # f'ori_text: {ori_text}\ninput: {test_input}\npipeline1: ' + # f'{pipeline1(test_input)}\npipeline2: {pipeline2(test_input)}' # ) - ## veco - #model_dir = snapshot_download(self.model_id_veco) - #preprocessor = FillMaskPreprocessor( - # model_dir, first_sequence='sentence', second_sequence=None) - #model = MaskedLanguageModel(model_dir) - #pipeline1 = FillMaskPipeline(model, preprocessor) - #pipeline2 = pipeline( - # Tasks.fill_mask, model=model, preprocessor=preprocessor) - #for language in ["zh", "en"]: - # ori_text = self.ori_texts[language] - # test_input = self.test_inputs["zh"].replace("[MASK]", "") - # print( - # f'ori_text: {ori_text}\ninput: {test_input}\npipeline1: {pipeline1(test_input)}\npipeline2: {pipeline2(test_input)}' - + ## veco + #model_dir = snapshot_download(self.model_id_veco) + #preprocessor = FillMaskPreprocessor( + # model_dir, first_sequence='sentence', second_sequence=None) + #model = MaskedLanguageModel(model_dir) + #pipeline1 = FillMaskPipeline(model, preprocessor) + #pipeline2 = pipeline( + # Tasks.fill_mask, model=model, preprocessor=preprocessor) + #for language in ["zh", "en"]: + # ori_text = self.ori_texts[language] + # test_input = self.test_inputs["zh"].replace("[MASK]", "") + # print( + # f'ori_text: {ori_text}\ninput: {test_input}\npipeline1: ' + # f'{pipeline1(test_input)}\npipeline2: {pipeline2(test_input)}' def test_run_with_model_from_modelhub(self): - for language in ["zh"]: + for language in ['zh']: print(self.model_id_sbert[language]) model = Model.from_pretrained(self.model_id_sbert[language]) - print("model", model.model_dir) + print('model', model.model_dir) preprocessor = FillMaskPreprocessor( - model.model_dir, first_sequence='sentence', second_sequence=None) + model.model_dir, + first_sequence='sentence', + second_sequence=None) pipeline_ins = pipeline( - task=Tasks.fill_mask, model=model, preprocessor=preprocessor) - print(pipeline_ins(self_test_inputs[language])) - + task=Tasks.fill_mask, model=model, preprocessor=preprocessor) + print(pipeline_ins(self.test_inputs[language])) #def test_run_with_model_name(self): - ## veco - #pipeline_ins = pipeline( - # task=Tasks.fill_mask, model=self.model_id_veco) - #for language in ["zh", "en"]: - # input_ = self.test_inputs[language].replace("[MASK]", "") - # print(pipeline_ins(input_)) + ## veco + #pipeline_ins = pipeline( + # task=Tasks.fill_mask, model=self.model_id_veco) + #for language in ["zh", "en"]: + # input_ = self.test_inputs[language].replace("[MASK]", "") + # print(pipeline_ins(input_)) - ## structBert - #for language in ["zh"]: - # pipeline_ins = pipeline( - # task=Tasks.fill_mask, model=self.model_id_sbert[language]) - # print(pipeline_ins(self_test_inputs[language])) + ## structBert + #for language in ["zh"]: + # pipeline_ins = pipeline( + # task=Tasks.fill_mask, model=self.model_id_sbert[language]) + # print(pipeline_ins(self_test_inputs[language])) if __name__ == '__main__': unittest.main() -