import os
import sys
import base64
from io import BytesIO
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

import torch
from fastapi import FastAPI
import numpy as np
from PIL import Image

import clip
from dalle.models import Dalle
from dalle.utils.utils import clip_score, download

print("Loading models...")
app = FastAPI()


url = "https://arena.kakaocdn.net/brainrepo/models/minDALL-E/57b008f02ceaa02b779c8b7463143315/1.3B.tar.gz"
root = os.path.expanduser("~/.cache/minDALLE")
filename = os.path.basename(url)
pathname = filename[: -len(".tar.gz")]
download_target = os.path.join(root, filename)
result_path = os.path.join(root, pathname)
if not os.path.exists(result_path):
    result_path = download(url, root)


device = "cuda" if torch.cuda.is_available() else "cpu"
model = Dalle.from_pretrained(result_path)  # This will automatically download the pretrained model.
model.to(device=device)
model_clip, preprocess_clip = clip.load("ViT-B/32", device=device)
model_clip.to(device=device)

print("Models loaded !")


@app.get("/")
def read_root():
    return {"minDALL-E!"}


@app.get("/{generate}")
def generate(prompt):
    images = sample(prompt)
    images = [to_base64(image) for image in images]
    return {"images": images}


def sample(prompt):
    # Sampling
    images = (
        model.sampling(prompt=prompt, top_k=96, top_p=None, softmax_temperature=1.0, num_candidates=9, device=device)
        .cpu()
        .numpy()
    )
    images = np.transpose(images, (0, 2, 3, 1))

    # CLIP Re-ranking
    rank = clip_score(
        prompt=prompt, images=images, model_clip=model_clip, preprocess_clip=preprocess_clip, device=device
    )
    images = images[rank]
    
    pil_images = []
    for i in range(len(images)):
        im = Image.fromarray((images[i] * 255).astype(np.uint8))
        pil_images.append(im)
    
    return pil_images


def to_base64(pil_image):
    buffered = BytesIO()
    pil_image.save(buffered, format="JPEG")
    return base64.b64encode(buffered.getvalue())