metadata
language:
- ru
tags:
- keywords2poetry
- Pushkin
- Mayakovsky
- Esenin
- Blok
- Tyutchev
This repo contains the fune-tuned version of ai-forever/rugpt3medium_based_on_gpt2, which can generate poetry from keywords in style of Pushkin, Mayakovsky, Esenin, Blok and Tyutchev.
To use this model, you can do this:
from transformers import AutoTokenizer, AutoModelForCausalLM
def generate_poetry(input: str, model, num_beams=3):
input = input if len(input) > 0 else tokenizer.bos_token
input_ids = tokenizer.encode(input, return_tensors="pt").to(device)
# Create an attention mask
attention_mask = (input_ids != tokenizer.pad_token_id).float()
# Set the pad_token_id
tokenizer.pad_token_id = tokenizer.eos_token_id
with torch.no_grad():
out = model.generate(input_ids,
do_sample=True,
num_beams=num_beams,
temperature=2.0,
top_p=0.9,
max_length = 200,
eos_token_id=tokenizer.eos_token_id,
bos_token_id=tokenizer.bos_token_id,
attention_mask=attention_mask
).to(device)
return tokenizer.batch_decode(out, skip_special_tokens=True)[0]
path = 'AnyaSchen/rugpt3-medium-keywords2poetry'
tokenizer = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path).to(device)
inp = 'Автор: Маяковский\nКлючевые слова:<write your keywords>'
print(generate_poetry(inp, model))