ru_transformers

Russian GPT-2

Google colab notebook for finetuning.

https://colab.research.google.com/drive/1jwFks82BLyy8x3oxyKpiNdlL1PfKSQwW?usp=sharing

Google colab notebook for generating text corpus.

https://colab.research.google.com/drive/1Hsp2508TXMR0ihYOLjKYOzWm9byqg9ue

1. I just want to play with your models

You can try writing with the model here https://porfirevich.ru and with Telegram chat bot @PorfBot

You can try poetry with Telegram chat bot @NeuroPoetBot

2. What are results?

Your perplexity will be different, depending on the tokenizer, the vocab and the dataset. The better your tokenizer the worse your perplexity, actually.

Values in the table are perplexity on the validation set.

Huge dataset

Classics dataset. It's only 500Mb and GPT-2 overfits it pretty fast.

Poetry dataset

Pelevin dataset

I've trained the model using gradual unfreezing with '--unfreeze_level' parameter. The sequence was 0,1,2,7,-1 (as in the table with results). When loss don't improve for a day I switch to next value (like from 2 to 7). You can find my exact scripts in tpu/schedule_small.txt and tpu/schedule_medium.txt.

3. I'd like to download your models

The model that isn't fine-tuned on any author is here

pip install awscli
aws s3 sync --no-sign-request s3://models.dobro.ai/gpt2/ru/unfreeze_all gpt2

Folders with s_ prefix contain Small (124M) model, m_ - for Medium (355M) model.

To understand how to generate text you should start by looking at rest.py.

Also, you can download all fine-tuned models.

aws s3 sync --no-sign-request s3://models.dobro.ai/gpt2/ru all

The one with which you can play on the site is located in the Pelevin folder.

4. I've got a small Russian dataset and I want to finetune your model on it

Download the models (intructions above), choose the model and put it in your output folder. Use validation set and be careful with overfitting. On small dataset it will overfit very fast - 3-7 epoch. Follow instructions below, except you don't need to train you tokenization dictionary, because you already have one.

5. I've got a big dataset on my lang and I want to train GPT-2 on it

I'd suggest that if you don't have a bunch of GPU's you should consider renting a Google TPU. On my Nvidia Titan RTX an epoch takes 70 minutes and the same epoch takes 12.5 minutes on TPU v3-8. I've used fp16 on GPU, but I can't use bfloat16 on TPU, because it's training poorly on bfloat16 at the moment (it could have been 8 minutes if implemented properly).

You can ask for access to Google's TensorFlow Research Cloud and use TPUs for free for one month.

In the process, I've switched tokenization library from SentencePiece to YTTM. YTTM is better (10% smaller files) and much faster. If you for some reason want to use SentencePiece then the code is here, just change the tokenizer in the command line.

First, the GPT-2 model will learn Russian on a huge dataset (230 GB), and then it will learn good Russian on the Russian classical literature (500 MB). I use progressive layer unfreezing to use transfer training. Validation set is the correspondence between Leo Tolstoy with young Mahatma Gandhi.

5.1. Download a fb2 library

Main link

For finetuning first second Dostoyevskiy Tolstoy Pushkin Bulgakov Gogol Pelevin

5.2. Install dependencies

sudo xargs -a apt.txt apt install
conda env create -f environment.yml

5.3. Build and Install SentencePiece (skip if use YTTM)

Follow instructions here https://github.com/google/sentencepiece

5.4. Prepare the dataset files

Use corpus/corpus.ipynb on your dataset.

Or in google colab: https://colab.research.google.com/drive/1Hsp2508TXMR0ihYOLjKYOzWm9byqg9ue

5.5. Create vocabulary for the YTTM (and SentencePiece) tokenizer

You can skip this step if you want only to finetune the model with the existing vocab.

yttm bpe --data ./corpus/tmp/russian_corpus_for_vocab.txt --model bpe/yt.model --vocab_size 50257 --coverage 0.9999

# SentencePiece
spm_train --input=./corpus/tmp/russian_corpus_for_vocab.txt --model_prefix=bpe/m50 --vocab_size=50257 --user_defined_symbols=''

5.6. If you want to use Google TPU, go here https://github.com/mgrankin/ru_transformers/tree/master/tpu

5.7. Install fp16 support

Mixed precision training with opt_level O2 gives the exact same loss but much faster and with less memory. The downside - APEX with O2 doesnt work with DataParallel yet, see https://github.com/NVIDIA/apex/issues/227

5.7.1 Make sure to install proper bare metal cuda.

wget http://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run -O cuda.run
chmod +x cuda.run
sudo ./cuda.run

5.7.2 Apex

export CUDA_HOME=/usr/local/cuda-10.2
git clone https://github.com/NVIDIA/apex
cd apex
# fix setup.py if complains for version mismatch
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

5.8. Train your model!

…

5.9. Save trained model

aws s3 cp output_s/config.json s3://models.dobro.ai/gpt2/ru/small/
aws s3 cp output_s/encoder.model s3://models.dobro.ai/gpt2/ru/small/
aws s3 cp output_s/pytorch_model.bin s3://models.dobro.ai/gpt2/ru/small/

5.10. Deploy the model

git clone https://github.com/mgrankin/ru_transformers.git
cd ru_transformers
mkdir logs
aws s3 sync --no-sign-request s3://models.dobro.ai/gpt2/ru gpt2
cp -R gpt2/pelevin/m_checkpoint-3365357 gpt2/medium
cp -R gpt2/poetry/m_checkpoint-3397989 gpt2/medium/poetry
conda env create -f environment.yml
conda activate gpt
uvicorn rest:app --reload --host 0.0.0.0
# crontab  DEVICE="cuda:1"
# @reboot /bin/bash -c "cd ru_transformers; git pull; source ~/.bashrc; conda activate gpt; DEVICE="cuda:1" uvicorn rest:app --reload --host 0.0.0.0"

6. Additional scripts

evaluate_model.py - to evaluate your model using input file or prompt.

text_processing.py - to process your dataset.

to_token_convertor.py - to convert your string to tokens. In case if you curious.