README.md

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))