modules/inference.py

from time import gmtime, strftime

print(f'{strftime("%Y-%m-%d %H:%M:%S", gmtime())} Preparing for inference...')  # noqa

from rudalle.pipelines import generate_images
from rudalle import get_rudalle_model, get_tokenizer, get_vae
from huggingface_hub import hf_hub_url, cached_download
import torch
from io import BytesIO
import base64

print(f"GPUs available: {torch.cuda.device_count()}")
print(f"GPU[0] memory: {int(torch.cuda.get_device_properties(0).total_memory / 1048576)}Mib")
print(f"GPU[0] memory reserved: {int(torch.cuda.memory_reserved(0) / 1048576)}Mib")
print(f"GPU[0] memory allocated: {int(torch.cuda.memory_allocated(0) / 1048576)}Mib")

device = "cuda" if torch.cuda.is_available() else "cpu"
fp16 = torch.cuda.is_available()

file_dir = "./models"
file_name = "pytorch_model.bin"
config_file_url = hf_hub_url(repo_id="minimaxir/ai-generated-pokemon-rudalle", filename=file_name)
cached_download(config_file_url, cache_dir=file_dir, force_filename=file_name)

model = get_rudalle_model('Malevich', pretrained=False, fp16=fp16, device=device)
model.load_state_dict(torch.load(f"{file_dir}/{file_name}", map_location=f"{'cuda:0' if torch.cuda.is_available() else 'cpu'}"))

vae = get_vae().to(device)
tokenizer = get_tokenizer()

print(f'{strftime("%Y-%m-%d %H:%M:%S", gmtime())} Ready for inference')


def english_to_russian(english_string):
    word_map = {
        "grass": "трава",
        "fire": "Пожар",
        "water": "вода",
        "lightning": "молния",
        "fighting": "борьба",
        "psychic": "психический",
        "colorless": "бесцветный",
        "darkness": "темнота",
        "metal": "металл",
        "dragon": "Дракон",
        "fairy": "сказочный"
    }

    return word_map[english_string.lower()]


def generate_image(prompt):
    if prompt.lower() in ['grass', 'fire', 'water', 'lightning', 'fighting', 'psychic', 'colorless', 'darkness',
                          'metal', 'dragon', 'fairy']:
        prompt = english_to_russian(prompt)

    result, _ = generate_images(prompt, tokenizer, model, vae, top_k=2048, images_num=1, top_p=0.995)

    buffer = BytesIO()
    result[0].save(buffer, format="PNG")
    base64_bytes = base64.b64encode(buffer.getvalue())
    base64_string = base64_bytes.decode("UTF-8")

    return "data:image/png;base64," + base64_string