server.py

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 torch import nn
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("minDALL-E/1.3B")  # This will automatically download the pretrained model.
model.to(device=device)

# -----------------------------------------------------------
state_dict_ = torch.load('last.ckpt', map_location='cpu')
vqgan_stage_dict = model.stage1.state_dict()

for name, param in state_dict_['state_dict'].items():
    if name not in model.stage1.state_dict().keys():
        continue
    if isinstance(param, nn.parameter.Parameter):
        param = param.data
    vqgan_stage_dict[name].copy_(param)

model.stage1.load_state_dict(vqgan_stage_dict)
# ---------------------------------------------------------
# state_dict_dalle = torch.load('dalle_last.ckpt', map_location='cpu')
# dalle_stage_dict = model.stage2.state_dict()
#
# for name, param in state_dict_dalle['state_dict'].items():
#     if name[6:] not in model.stage2.state_dict().keys():
#         print(name)
#         continue
#     if isinstance(param, nn.parameter.Parameter):
#         param = param.data
#     dalle_stage_dict[name[6:]].copy_(param)
#
# model.stage2.load_state_dict(dalle_stage_dict)

# 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())