from transformers import GPT2LMHeadModel, AutoTokenizer
import demo_util
import numpy as np
import torch
from PIL import Image
import os

torch.backends.cuda.matmul.allow_tf32 = True
torch.manual_seed(0)

device = "cuda:1"
dtype = torch.float16

config = demo_util.get_config("configs/titok_l32.yaml")
titok_tokenizer = demo_util.get_titok_tokenizer(config)
titok_tokenizer = titok_tokenizer.to(device)
tokenizer = AutoTokenizer.from_pretrained("./image_tokenizer")
model = GPT2LMHeadModel.from_pretrained("./checkpoint-20000").to(device).to(dtype).eval()

def detokenize(tokens):
    encoded_tokens = torch.from_numpy(np.array(tokens).astype(np.int64)).view(1, 1, -1).to(device)
    reconstructed_image = titok_tokenizer.decode_tokens(encoded_tokens)
    reconstructed_image = torch.clamp(reconstructed_image, 0.0, 1.0)
    reconstructed_image = (reconstructed_image * 255.0).permute(0, 2, 3, 1).to("cpu", dtype=torch.uint8).numpy()[0]
    return Image.fromarray(reconstructed_image)

prompt = ""

inputs = tokenizer(f"{text}<|startofimage|>", return_tensors="pt").to(device)
input_ids = inputs["input_ids"]
init = input_ids.shape[-1]
soi_token = tokenizer.encode("<|image:0|>")[0]

for _ in range(33):
    logits = model(input_ids).logits[:, -1, :]
    probas = torch.nn.functional.softmax(logits, dim=-1)
    pred_idx = torch.argmax(probas, dim=-1, keepdim=True)
    input_ids = torch.cat((input_ids, pred_idx), dim=-1)
tokenizer.decode(input_ids[0])
tokens = input_ids[:, init:-1].detach().cpu().squeeze().numpy() - soi_token

if np.any(tokens < 0) or np.any(tokens >= 4096):
    print("Illegal Image Tokens")

else:
    img = detokenize(tokens)
    img.save(f"./out.png")