tfkit 0.3.38

Transformers kit - NLP library for different downstream tasks, built on huggingface project

TFKit lets everyone make use of transformer architecture on many tasks and models in small change of config.
At the same time, it can do multi-task multi-model learning, and can introduce its own data sets and tasks through simple modifications.

Feature

• One-click replacement of different pre-trained models
• Support multi-model and multi-task
• Classifier with multiple labels and multiple classifications
• Unify input formats for different tasks
• Separation of data reading and model architecture
• Support various loss function and indicators

Supplement

• Model list: Support Bert/GPT/GPT2/XLM/XLNet/RoBERTa/CTRL/ALBert/...
• NLPrep: download and preprocessing data in one line
• nlp2go: create demo api as quickly as possible.

Quick Start

Installing via pip

pip install tfkit

Running TFKit to train a ner model

download dataset using nlprep

nlprep --dataset tag_clner  --outdir ./clner_row --util s2t

train model with albert

tfkit-train --batch 10 \
--epoch 3 \
--lr 5e-6 \
--train ./clner_row/train \
--test ./clner_row/test \
--maxlen 512 \
--model tagRow \
--savedir ./albert_ner \
--config voidful/albert_chinese_small

eval model

tfkit-eval --model ./albert_ner/3.pt --valid ./clner_row/validation --metric clas

host prediction service

nlp2go --model ./albert_ner/3.pt --api_path ner

Overview

Train

$ tfkit-train
Run training

arguments:
  --train TRAIN [TRAIN ...]     train dataset path
  --test TEST [TEST ...]        test dataset path
  --config CONFIG               distilbert-base-multilingual-cased/bert-base-multilingual-cased/voidful/albert_chinese_small
  --model {once,twice,onebyone,clas,tagRow,tagCol,qa,onebyone-neg,onebyone-pos,onebyone-both} [{once,twice,onebyone,clas,tagRow,tagCol,qa,onebyone-neg,onebyone-pos,onebyone-both} ...]
                                model task
  --savedir SAVEDIR     model saving dir, default /checkpoints
optional arguments:
  -h, --help            show this help message and exit
  --batch BATCH         batch size, default 20
  --lr LR [LR ...]      learning rate, default 5e-5
  --epoch EPOCH         epoch, default 10
  --maxlen MAXLEN       max tokenized sequence length, default 368
  --lossdrop            loss dropping for text generation
  --tag TAG [TAG ...]   tag to identity task in multi-task
  --seed SEED           random seed, default 609
  --worker WORKER       number of worker on pre-processing, default 8
  --grad_accum          gradient accumulation, default 1
  --tensorboard         Turn on tensorboard graphing
  --resume RESUME       resume training
  --cache               cache training data

Eval

$ tfkit-eval
Run evaluation on different benchmark
arguments:
  --model MODEL             model path
  --metric {emf1,nlg,clas}  evaluate metric
  --valid VALID             evaluate data path

optional arguments:
  -h, --help            show this help message and exit
  --print               print each pair of evaluate data
  --enable_arg_panel    enable panel to input argument

Contributing

Thanks for your interest.There are many ways to contribute to this project. Get started here.

License

• License

Icons reference

Icons modify from Freepik from www.flaticon.com
Icons modify from Nikita Golubev from www.flaticon.com