dayyass / language-modeling

Pipeline for training Language Models using PyTorch.

About

Pipeline for training Language Models using PyTorch.
Inspired by Yandex Data School NLP Course (week03: Language Modeling)

Usage

First, install dependencies:

# clone repo
git clone https://github.com/dayyass/language_modeling.git

# install dependencies
cd language_modeling
pip install -r requirements.txt

Data Format

Prepared text file with space separated words on each line.
More about it here.

Statistical Language Modeling

Training

Script for training statistical language models:

python statistical_lm/train.py --path_to_data "data/arxiv_train.txt" --n 3 --path_to_save "models/3_gram_language_model.pkl"

Required arguments:

• --path_to_data - path to train data
• --n - n-gram order

Optional arguments:

• --smoothing - smoothing method (available: None, "add-k") (default: None)
• --delta - smoothing additive parameter (only for add-k smoothing) (default: 1.0)
• --path_to_save - path to save model (default: "models/language_model.pkl")
• --verbose - verbose (default: True)

Validation

Script for validation statistical language models using perplexity:

python statistical_lm/validate.py --path_to_data "data/arxiv_test.txt" --path_to_model "models/3_gram_language_model.pkl"

Required arguments:

• --path_to_data - path to validation data
• --path_to_model - path to language model

Optional arguments:

• --verbose - verbose (default: True)

Inference

Script for generation new sequences using statistical language models:

python statistical_lm/inference.py --path_to_model "models/3_gram_language_model.pkl" --prefix "artificial" --temperature 0.5

Required arguments:

• --path_to_model - path to language model

Optional arguments:

• --prefix - prefix before sequence generation (default: "")
• --strategy - decoding strategy (available: "sampling", "top-k-uniform", "top-k", "top-p-uniform", "top-p" and "beam search") (default: "sampling")
• --temperature - sampling temperature, if temperature == 0.0, always takes most likely token - greedy decoding (only for "sampling" decoding strategy) (default: 0.0)
• --k - top-k parameter (only for "top-k-uniform" and "top-k" decoding strategy) (default: 10)
• --p - top-p parameter (only for "top-p-uniform" and "top-p" decoding strategy) (default: 0.9)
• --max_length - max number of generated words (default: 100)
• --seed - random seed (default: 42)

Command output with 3-gram language model trained on arxiv.txt with prefix "artificial" and temperature 0.5:

artificial neural network ( cnn ) architectures on h2o platform for real - world applications . <EOS>

RNN Language Modeling

Training

Script for training RNN language models:

python rnn_lm/train.py --path_to_data "data/arxiv_train.txt" --path_to_save_folder "models/rnn_language_model" --n_epoch 5 --max_length 512 --batch_size 128 --embedding_dim 64 --rnn_hidden_size 256

Required arguments:

• --path_to_data - path to train data
• --n_epoch - number of epochs
• --batch_size - dataloader batch_size
• --embedding_dim - embedding dimension
• --rnn_hidden_size - LSTM hidden size

Optional arguments:

• --path_to_save_folder - path to save folder (default: "models/rnn_language_model")
• --max_length - max sentence length (chars) (default: None)
• --shuffle - dataloader shuffle (default: True)
• --rnn_num_layers - number of LSTM layers (default: 1)
• --rnn_dropout - LSTM dropout (default: 0.0)
• --train_eval_freq - evaluation frequency (number of batches) (default: 50)
• --clip_grad_norm - max_norm parameter in clip_grad_norm (default: 1.0)
• --seed - random seed (default: 42)
• --device - torch device (available: "cpu", "cuda") (default: "cuda")
• --verbose - verbose (default: True)

Validation

Script for validation RNN language models using perplexity:

python rnn_lm/validate.py --path_to_data "data/arxiv_test.txt" --path_to_model_folder "models/rnn_language_model" --max_length 512

Required arguments:

• --path_to_data - path to validation data
• --path_to_model - path to language model

Optional arguments:

• --max_length - max sentence length (chars) (default: None)
• --seed - random seed (default: 42)
• --device - torch device (available: "cpu", "cuda") (default: "cuda")
• --verbose - verbose (default: True)

Inference

Script for generation new sequences using RNN language models:

python rnn_lm/inference.py --path_to_model_folder "models/rnn_language_model" --prefix "artificial" --temperature 0.5

Required arguments:

• --path_to_model_folder - path to language model folder

Optional arguments:

• --prefix - prefix before sequence generation (default: "")
• --temperature - sampling temperature, if temperature == 0.0, always takes most likely token - greedy decoding (default: 0.0)
• --max_length - max number of generated tokens (chars) (default: 100)
• --seed - random seed (default: 42)
• --device - torch device (available: "cpu", "cuda") (default: "cuda")

Command output with RNN language model trained on arxiv.txt with prefix "artificial" and temperature 0.5:

artificial visual information of the number , using an intervidence for detection for order to the recognition

Models

List of implemented models:

• N-gram Language Model
• RNN Language Model (char-based)
• GPT Language Model

Decoding Strategy

• greedy
• temperature sampling
• top-k-uniform
• top-k
• top-p-uniform
• top-p
• beam search

Smoothing (only for N-gram Language Models)

• no smoothing
• add-k / Laplace smoothing
• interpolation smoothing
• back-off / Katz smoothing
• Kneser-Ney smoothing

Models Comparison

Generation comparison available here.

Statistical Language Modeling

perplexity (train / test)	none	add-k / Laplace	interpolation	back-off / Katz	Kneser-Ney
1-gram	881.27 / 1832.23	882.63 / 1838.22	-	-	-
2-gram	95.32 / 8.57e+7	1106.79 / 1292.02	-	-	-
3-gram	12.78 / 6.2e+22	7032.91 / 10499.24	-	-	-