app.py

from denoising_diffusion_pytorch import Unet, GaussianDiffusion
import streamlit as st
import torch


def get_model():
    unet = Unet(
        dim = 64,
        dim_mults = (1, 2, 4, 8)
    )
    model = GaussianDiffusion(
        unet,
        image_size = 64,
        timesteps = 1000,           # number of steps
        sampling_timesteps = 250,   # number of sampling timesteps (using ddim for faster inference [see citation for ddim paper])
        loss_type = 'l1' ,           # L1 or L2
        p2_loss_weight_gamma = 1.
    )
    model.load_state_dict(torch.load("./model-final.pt", map_location="cpu"))
    model.eval()
    return model

def scale_to_255(x):
    return ((x+1)/2*255).astype('uint8')

if __name__ == "__main__":
    st.title("Sushi Diffusion")
    st.text("The generation process takes about 10 mins.")
    st.text("If you don't want to wait, please visit: https://thissushidoesnotexist.com/")
    model = get_model()
    st.text('Press the button below to generate sushi!')
    if st.button('🍣'):
        bar = st.progress(0)
        img = torch.randn((1,3,64,64), device="cpu")
        for t in reversed(range(0, model.num_timesteps)):
            img, _ = model.p_sample(img, t, None)
            bar.progress((model.num_timesteps-t) / model.num_timesteps)
        img = scale_to_255(img.squeeze().numpy().transpose(1,2,0))
        st.image(img, caption='This sushi does not exist.')