This repository contains code to generate Hindi Shayaris. Please see this Medium article for more details.
- Make sure CUDA >= 10.0 is installed to enable GPU acceleration.
- Clone this repository and do
pip install -r requirements.txt
- Download pre-trained models from here.
python main.py [-h] --model_path MODEL_PATH --tokenizer_path TOKENIZER_PATH
[--length LENGTH] [--num_samples NUM_SAMPLES]
[--temperature TEMPERATURE] [--top_k TOP_K] [--top_p TOP_P]
[--no_cuda] [--seed SEED]
Required arguments:
--model_path MODEL_PATH
Path to pre-trained model
--tokenizer_path TOKENIZER_PATH
Path to pre-trained tokenizer
Optional Arguments:
-h, --help Show this help message and exit
--length LENGTH Maximum length of sample (default: 100)
--num_samples NUM_SAMPLES
Number of samples to generate for a prompt (default:
1)
--temperature TEMPERATURE
Softmax temperature, temperature=0 implies greedy
sampling (default: 1.0)
--top_k TOP_K Value of k for top-k sampling, for k=0 top-k sampling
is not used (default: 0)
--top_p TOP_P Value of p for nucleus sampling (default: 0.9)
--no_cuda Avoid using CUDA when available (default: False)
--seed SEED Random seed for initialization (default: 2020)
For example:
python main.py --model_path models/mle-model/ --tokenizer_path tokenizer/ --top_p 0.8 --temperature 0.9