This repo contains a PyTorch implementation of a pretrained BERT model for chinese text classification.
At the root of the project, you will see:
├── pybert
| └── callback
| | └── lrscheduler.py
| | └── trainingmonitor.py
| | └── ...
| └── config
| | └── base.py #a configuration file for storing model parameters
| └── dataset
| └── io
| | └── bert_processor.py
| └── model
| | └── nn
| | └── pretrain
| └── output #save the ouput of model
| └── preprocessing #text preprocessing
| └── train #used for training a model
| | └── trainer.py
| | └── ...
| └── utils # a set of utility functions
├── run_bert.py
- csv
- tqdm
- numpy
- pickle
- scikit-learn
- PyTorch 1.0
- matplotlib
- pytorch_transformers=1.1.0
you need download pretrained chinese bert model
- Download the Bert pretrained model from s3
- Download the Bert config file from s3
- Download the Bert vocab file from s3
- modify
bert-base-chinese-pytorch_model.bin
topytorch_model.bin
,bert-base-chinese-config.json
toconfig.json
,bert-base-chinese-vocab.txt
tovocab.txt
- place
model
,config
andvocab
file into the/pybert/pretrain/bert/base-uncased
directory. pip install pytorch-transformers
from github.- Prepare BaiduNet{password:ruxu}, you can modify the
io.bert_processor.py
to adapt your data. - Modify configuration information in
pybert/config/base.py
(the path of data,...). - Run
python run_bert.py --do_data
to preprocess data. - Run
python run_bert.py --do_train --save_best
to fine tuning bert model. - Run
run_bert.py --do_test --do_lower_case
to predict new data.
Epoch: 3 - loss: 0.0222 acc: 0.9939 - f1: 0.9911 val_loss: 0.0785 - val_acc: 0.9799 - val_f1: 0.9800
label | precision | recall | f1-score | support |
---|---|---|---|---|
财经 | 0.97 | 0.96 | 0.96 | 1500 |
体育 | 1.00 | 1.00 | 1.00 | 1500 |
娱乐 | 0.99 | 0.99 | 0.99 | 1500 |
家居 | 0.99 | 0.99 | 0.99 | 1500 |
房产 | 0.96 | 0.97 | 0.96 | 1500 |
教育 | 0.98 | 0.97 | 0.97 | 1500 |
时尚 | 0.99 | 0.98 | 0.99 | 1500 |
时政 | 0.97 | 0.98 | 0.98 | 1500 |
游戏 | 1.00 | 0.99 | 0.99 | 1500 |
科技 | 0.96 | 0.97 | 0.97 | 1500 |
avg / total | 0.98 | 0.98 | 0.98 | 15000 |
- When converting the tensorflow checkpoint into the pytorch, it's expected to choice the "bert_model.ckpt", instead of "bert_model.ckpt.index", as the input file. Otherwise, you will see that the model can learn nothing and give almost same random outputs for any inputs. This means, in fact, you have not loaded the true ckpt for your model
- When using multiple GPUs, the non-tensor calculations, such as accuracy and f1_score, are not supported by DataParallel instance
- As recommanded by Jocob in his paper https://arxiv.org/pdf/1810.04805.pdf, in fine-tuning tasks, the hyperparameters are expected to set as following: Batch_size: 16 or 32, learning_rate: 5e-5 or 2e-5 or 3e-5, num_train_epoch: 3 or 4
- The pretrained model has a limit for the sentence of input that its length should is not larger than 512, the max position embedding dim. The data flows into the model as: Raw_data -> WordPieces -> Model. Note that the length of wordPieces is generally larger than that of raw_data, so a safe max length of raw_data is at ~128 - 256
- Upon testing, we found that fine-tuning all layers could get much better results than those of only fine-tuning the last classfier layer. The latter is actually a feature-based way