Out-of-range error while training ATEPC model - english on T-shirt dataset.

... config.model = ATEPCModelList.LCFS_ATEPC config.evaluate_begin = 5 config.num_epoch = 6 config.log_step = 100 tshirt = ABSADatasetList.TShirt

aspect_extractor = Trainer(config=config, dataset=tshirt, checkpoint_save_mode=1, auto_device=True )

Traceback - >

TShirt dataset is not found locally, search at https://github.com/yangheng95/ABSADatasets Some weights of the model checkpoint at bert-base-uncased were not used when initializing BertModel: ['cls.predictions.transform.LayerNorm.bias', 'cls.seq_relationship.weight', 'cls.predictions.transform.dense.bias', 'cls.predictions.decoder.weight', 'cls.predictions.bias', 'cls.seq_relationship.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.transform.LayerNorm.weight']

• This IS expected if you are initializing BertModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
• This IS NOT expected if you are initializing BertModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model). Using bos_token, but it is not set yet. Using eos_token, but it is not set yet. 59%|█████▊ | 1098/1870 [00:10<00:07, 100.65it/s, convert examples to features]

IndexError Traceback (most recent call last) in () 3 # from_checkpoint=checkpoint_path, 4 checkpoint_save_mode=1, ----> 5 auto_device=True 6 )

7 frames /usr/local/lib/python3.7/dist-packages/pyabsa/functional/trainer/trainer.py in init(self, config, dataset, from_checkpoint, checkpoint_save_mode, auto_device) 92 config.model_path_to_save = None 93 ---> 94 self.train() 95 96 def train(self):

/usr/local/lib/python3.7/dist-packages/pyabsa/functional/trainer/trainer.py in train(self) 103 self.config.seed = s 104 if self.checkpoint_save_mode: --> 105 model_path.append(self.train_func(self.config, self.from_checkpoint, self.logger)) 106 else: 107 # always return the last trained model if dont save trained model

/usr/local/lib/python3.7/dist-packages/pyabsa/core/atepc/training/atepc_trainer.py in train4atepc(opt, from_checkpoint_path, logger) 352 while not trainer: 353 try: --> 354 trainer = Instructor(opt, logger) 355 if from_checkpoint_path: 356 model_path = find_files(from_checkpoint_path, '.model')

/usr/local/lib/python3.7/dist-packages/pyabsa/core/atepc/training/atepc_trainer.py in init(self, opt, logger) 70 len(self.train_examples) / self.opt.batch_size / self.opt.gradient_accumulation_steps) * self.opt.num_epoch 71 train_features = convert_examples_to_features(self.train_examples, self.label_list, self.opt.max_seq_len, ---> 72 self.tokenizer, self.opt) 73 all_spc_input_ids = torch.tensor([f.input_ids_spc for f in train_features], dtype=torch.long) 74 all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)

/usr/local/lib/python3.7/dist-packages/pyabsa/core/atepc/dataset_utils/data_utils_for_training.py in convert_examples_to_features(examples, label_list, max_seq_len, tokenizer, opt) 188 text_right = '' 189 aspect = '' --> 190 prepared_inputs = prepare_input_for_atepc(opt, tokenizer, text_left, text_right, aspect) 191 lcf_cdm_vec = prepared_inputs['lcf_cdm_vec'] 192 lcf_cdw_vec = prepared_inputs['lcf_cdw_vec']

/usr/local/lib/python3.7/dist-packages/pyabsa/core/atepc/dataset_utils/atepc_utils.py in prepare_input_for_atepc(opt, tokenizer, text_left, text_right, aspect) 60 61 if 'lcfs' in opt.model_name or opt.use_syntax_based_SRD: ---> 62 syntactical_dist, _ = get_syntax_distance(text_raw, aspect, tokenizer, opt) 63 else: 64 syntactical_dist = None

/usr/local/lib/python3.7/dist-packages/pyabsa/core/apc/dataset_utils/apc_utils.py in get_syntax_distance(text_raw, aspect, tokenizer, opt) 240 # the following two functions are both designed to calculate syntax-based distances 241 if opt.srd_alignment: --> 242 syntactical_dist = syntax_distance_alignment(raw_tokens, dist, opt.max_seq_len, tokenizer) 243 else: 244 syntactical_dist = pad_syntax_based_srd(raw_tokens, dist, tokenizer, opt)[1]

/usr/local/lib/python3.7/dist-packages/pyabsa/core/apc/dataset_utils/apc_utils.py in syntax_distance_alignment(tokens, dist, max_seq_len, tokenizer) 38 if bert_tokens != text: 39 while text or bert_tokens: ---> 40 if text[0] == ' ' or text[0] == '\xa0': # bad case handle 41 text = text[1:] 42 dep_dist = dep_dist[1:]

IndexError: list index out of range

I am applying PyABASA package on amazon mobile phone reviews and its not generating attributes when the review contains numbers or emojis.

For example : iPhone 12. Best phone 😍 Genuine product thanks a lot amazon I purchase this divice 20 jan 2022 almost work fine. Best one

For above reviews and similar ones its not generating entities with sentiment. I really appreciate if this issue can be resolved.

Question Hi, it's great works you'd been made on this project.

I used this project for training on custom dataset, it has around 2000 examples. Label count is a little imbalance.Finally, I trained a model with 100 apoach and achieved apc_acc around 90 score. But the predict resullt is always Positive on all the aspect.

thanks very much you any advice?

Hi, thanks for the nice work. Recently I try to use the multilingual pretrained model for inference. I found that if the model predicts both of 2 consecutive words as (B-ASP). There will be a 'empty separator' error while inferencing. Is there any advice for avoiding this situation? Thanks again !

Hello, excuse me

1. It is not convenient for the party to write a document listing the versions of each package used by the framework.
2. One more question, will the packages used by the framework be updated in a timely manner? For example, if the torch is upgraded to 1.11.0, will the framework be updated in a timely manner?

你好，冒昧打扰，作者辛苦了，谢谢维护这个项目，在使用过程中遇到如下问题： 版本：1.16.5 文本如下： Let me begin by saying that there are two kinds of people, those who will give the Tokyo Hotel 5 stars and rave about it to everyone they know, or... people who can't get past the broken phone, blood stains, beeping fire alarms, peg-legged receptionist, lack of water pressure, cracked walls, strange smells, questionable elevator, televisions left to die after the digital conversion, and the possibility that the air conditioner may fall out the window at any moment. That being said, I whole-heartedly give the Tokyo Hotel 5 stars. This is not a place to quietly slip in and out of with nothing to show but a faint memory of the imitation Thomas Kinkade painting bolted to the wall above your bed. And, there is no continental breakfast or coffee in the lobby. There are a few vending machines, but I'm pretty sure they wont take change minted after 1970. Here your senses will be assaulted, and after you leave you will have enough memories to compete with a 1,000 mile road-trip. I beg anyone who is even mildly considering staying here to give it a chance. The location is prime. We were able to walk down Michigan Ave and the river-walk in the middle of the night, all without straying too far from the hotel. There is a grocery store a block away and parking (which may cost more that your hotel room) across the street. Besides, this place is cheap. Super-cheap for downtown Chicago. The closest price we found in the area was four times as expensive. But, be sure to grab some cash. They don't accept credit cards. Some rules though: - Say hello to Clifton Jackson, the homeless guy by Jewel-Osco. - Buy him a drink, some chicken and look him up on Facebook. - Stay on the 17 floor. All the way at the top. - Go out the fire escape (be sure to prop the door open or you'll have a looong walk down) - Be very very careful. - Explore. (Yes, that ladder will hold your weight) - Be very very careful. - Don't be alarmed by any weird noises you hear. - Spend the night on the roof. 17 stories up, in the heart of Chicago. - Write your own Yelp review. I want to see that others are getting the Tokyo Hotel Experience. - Check out is at noon. Be sure to drink lots of water. - Spend the next day hung over. And... Please be careful on the roof. 使用的预训练好的模型：fast_lcf_atepc_Multilingual_cdw_apcacc_88.96_apcf1_81.58_atef1_81.92 得到的结果：'aspect': [], 'position': [], 'sentiment': [], 'probs': [], 'confidence': [] 从结果看出，无法分析文本的细粒度情感，请问这种情况出现的原因是文本造成的还是模型的原因 关于预训练好的模型，我在hugging face上看到你有更新一些checkpoint，请问那些模型是不是可以直接用来加载使用？

Environment pyabsa: v1.1.22

Question atepc prediction result is array, but its length is not equal with inputs. For example: inputs examples = ['我就想问，这个真的用清水可以清洗的干净的吗？洗完之后油的吹不太干……难不成我昨晚发膜还要拿洗发水再洗一遍？那请问意义何在了……实在是很尴尬']*20

outputs [{'sentence': '我 就 想 问 ， 这 个 真 的 用 清 水 可 以 清 洗 的 干 净 的 吗 ？ 洗 完 之 后 油 的 吹 不 太 干 & hellip ; & hellip ; 难 不 成 我 昨 晚 发 膜 还 要 拿 洗 发 水 再 洗 一 遍 ？ 那 请 问 意 义 何 在 了 & hellip ; & hellip ; 实 在 是 很 尴 尬', 'IOB': ['O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'I-ASP', 'I-ASP', 'I-ASP', 'I-ASP', 'I-ASP', 'I-ASP', 'I-ASP', 'I-ASP', 'I-ASP', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', '[SEP]', 'O', 'O', 'O', 'O', 'O', 'O'], 'tokens': ['我', '就', '想', '问', '，', '这', '个', '真', '的', '用', '清', '水', '可', '以', '清', '洗', '的', '干', '净', '的', '吗', '？', '洗', '完', '之', '后', '油', '的', '吹', '不', '太', '干', '&', 'hellip', ';', '&', 'hellip', ';', '难', '不', '成', '我', '昨', '晚', '发', '膜', '还', '要', '拿', '洗', '发', '水', '再', '洗', '一', '遍', '？', '那', '请', '问', '意', '义', '何', '在', '了', '&', 'hellip', ';', '&', 'hellip', ';', '实', '在', '是', '很', '尴', '尬'], 'aspect': ['完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干', '完 之 后 油 的 吹 不 太 干'], 'position': [[23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31], [23, 24, 25, 26, 27, 28, 29, 30, 31]], 'sentiment': ['Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative', 'Negative']}]

checkpoint = 'model_garden/V0.8.8.0/Chinese/ATEPC/fast_lcf_atepc_Chinese_cdw_apcacc_96.69_apcf1_96.25_atef1_92.26' checkpoint = 'model_garden/V0.8.8.0/Chinese/ATEPC/fast_lcf_atepc_Multilingual_cdw_apcacc_79.61_apcf1_76.24_atef1_63.29.zip' 这两个模型都试了，Sentiment总是Positive，即便用很负向的表达。

我标注了一套自己的数据集，单独跑apc（用你们的APC模型跑的）任务指标正常。APC acc为91，f1为91.但是我跑多任务的时候，用你们的ATEPC的时候，ATE指标倒是正常，可是APC的指标很低，F1为37.37(max:37.37)。。。这是为啥。。。 是不是我想做多任务的时候只能先方面抽取ATE再情感极性分类？

Hi, yangheng! the project used to worked fine on my computer, but after installing the latest version of pyabsa, indexerror inccurs as below: yelp = "C:/Users/Li Wei/integrated_datasets/apc_datasets/SemEval/yelprestaurant" aspect_extractor = Trainer(config=config, dataset=yelp, checkpoint_save_mode=1, auto_device=True ).load_trained_model()

and indexerror related to above code is: `IndexError Traceback (most recent call last) in 1 yelp = "C:/Users/Li Wei/integrated_datasets/apc_datasets/SemEval/yelprestaurant" ----> 2 aspect_extractor = Trainer(config=config, 3 dataset=yelp, 4 checkpoint_save_mode=1, 5 auto_device=True

D:\Anaconda\lib\site-packages\pyabsa\functional\trainer\trainer.py in init(self, config, dataset, from_checkpoint, checkpoint_save_mode, auto_device) 71 72 """ ---> 73 config.ABSADatasetsVersion = query_local_version() 74 if isinstance(config, APCConfigManager): 75 self.train_func = train4apc

D:\Anaconda\lib\site-packages\pyabsa\utils\file_utils.py in query_local_version() 293 def query_local_version(): 294 fin = open(find_cwd_file(['init.py', 'integrated_datasets'])) --> 295 local_version = fin.read().split(''')[-2] 296 fin.close() 297 return local_version

IndexError: list index out of range`

I have run the "https://github.com/yangheng95/PyABSA/blob/release/examples/aspect_term_extraction/extract_aspects_chinese.py" on CPU device, and set "auto_device=False", but error message received that "Torch not compiled with CUDA enabled"。I have checked the class of "AspectExtractor" and the model class of "LCF_ATEPC", but no mistake were found。

Hi author @yangheng95 ,

I'm using the FAST-LCF-ATEPC model on my custom dataset and I have 4 questions on the ATEPC performance metrics and loss:

1. Whats the difference between these 2 Metric Visualizer (MV) tables? Is the validation set used to calculate these metrics?

2. As I understand from atepc_trainer.py , there are 3 types of losses which are loss_ate , loss_apc and lastly the combined loss that uses this formula loss = loss_ate + ate_loss_weight * loss_apc. I was wondering if you could explain it in simple terms how are each of the losses calculated from the expected output and the actual output?

3. In continuation to question 2, I want to check if the model overfits to my dataset and to do that I need to plot the training loss and validation loss. So does the `losses' list refer to the training loss? (see below) https://github.com/yangheng95/PyABSA/blob/964d7862da13ef8cc38cb56fe0e65086b343a9cd/pyabsa/core/atepc/training/atepc_trainer.py#L204

4. How can I retrieve the validation loss for ATE and APC separately so that I could plot them in a graph.

Kind regards, kerolzeeq

Dear @yangheng95,

Thanks for making and maintaining this repo, it's great!

I have some trouble to get the accuracy and F1 scores for the Restaurant Test data Gold. (Ideally I want to make a confusion matrix). What is the easiest way to get F1 scores for APC & ATE after running a checkpoint model on test data? Does the model store these metrics somewhere?

Alternatively, how do you compare your predictions to the TRUE test data (Restaurant Test data Gold annotated)? I can easily transform the models' predictions ('atepc_inference.result_json') to a pandas dataframe. But it is very hard to transform the test data stored in integrated datasets (from ABSAdatasets) (it is in IOB format) to that exact same format (pandas dataframe) in order to test performance. Do you have a script for that, or a certain function? I was not able to find it.

Btw: I used the multilingual checkpoint model (FAST-LCF-ATEPC) on the Restaurant14 Test data Gold (But, ultimately I want to use this model on Dutch data. That is why I want to know how to test performance).

Thanks a lot,

Karsten

Code:

import pyabsa as pyabsa

from pyabsa import available_checkpoints
# The results of available_checkpoints() depend on the PyABSA version
checkpoint_map = available_checkpoints()  # show available checkpoints of PyABSA of current version 

from pyabsa.functional import ABSADatasetList
from pyabsa.functional import ATEPCCheckpointManager
inference_source = ABSADatasetList.Restaurant14
aspect_extractor = ATEPCCheckpointManager.get_aspect_extractor(checkpoint='multilingual')
atepc_result = aspect_extractor.extract_aspect(inference_source=inference_source,
                                               save_result=True,
                                               print_result=True,  # print the result
                                               pred_sentiment=True,  # Predict the sentiment of extracted aspect terms
                                               )

import pandas as pd
df_restaurant_EN_test_pred = pd.read_json('atepc_inference.result_EN.json')

Hello can we update lCF-ATEPC to do Aspect term extraction and aspect category detection for SemEval dataset (instead of Aspect polarity classification) where replacing sentiment polarity(positive, negative, natural) with aspect categories (food, service, .....) Thanks in advance