app.py

import os
import builtins
import math
import streamlit as st
import gdown
#from google_drive_downloader import GoogleDriveDownloader as gdd

from demo.src.models import load_trained_model
from demo.src.utils import render_predict_from_pose, predict_to_image
#from demo.src.config import MODEL_DIR, MODEL_NAME, FILE_ID

st.set_page_config(page_title="DietNeRF")

def select_model():
    obj_select = st.selectbox("Select an object to render", ('Chair', 'Lego','Ship','Hotdog'))
    if obj_select == 'Chair':
        FILE_ID = "17dj0pQieo94TozFv-noSBkXebduij1aM"
        MODEL_DIR = 'models'
        MODEL_NAME = 'diet_nerf_chair'
    elif obj_select == 'Lego':
        FILE_ID = "1D9I-qIVMPaxuCHfUWPWMHaoLYtAmCjwI"
        MODEL_DIR = 'models'
        MODEL_NAME = 'diet_nerf_lego'
    elif obj_select == 'Ship':
        FILE_ID = "14ZeJ86ETQr8dtu6CFoxU-ifvniHKo_Dt"
        MODEL_DIR = 'models'
        MODEL_NAME = 'diet_nerf_ship'
    elif obj_select == 'Hotdog':
        FILE_ID = "11vNlR4lMvV_AVFgVjZmKMrMWGVG7qhNu"
        MODEL_DIR = 'models'
        MODEL_NAME = 'diet_nerf_hotdog'
    return MODEL_DIR,MODEL_NAME,FILE_ID

MODEL_DIR,MODEL_NAME,FILE_ID = select_model()

@st.cache
def download_model():
    os.makedirs(MODEL_DIR, exist_ok=True)
    _model_path = os.path.join(MODEL_DIR, MODEL_NAME)
    # gdd.download_file_from_google_drive(file_id=FILE_ID,
    #                                     dest_path=_model_path,
    #                                     unzip=True)
    url = f'https://drive.google.com/uc?id={FILE_ID}'
    gdown.download(url, _model_path, quiet=False)
    print(f'Model downloaded from google drive: {_model_path}')


@st.cache(show_spinner=False, allow_output_mutation=True)
def fetch_model():
    model, state = load_trained_model(MODEL_DIR, MODEL_NAME)
    return model, state


model_path = os.path.join(MODEL_DIR, MODEL_NAME)
if not os.path.isfile(model_path):
    download_model()

model, state = fetch_model()
pi = math.pi
st.sidebar.image("images/diet-nerf.png", width=310)
st.sidebar.header('SELECT YOUR VIEW DIRECTION')
theta = st.sidebar.slider("Theta", min_value=-pi, max_value=pi,
                          step=0.5, value=0.)
phi = st.sidebar.slider("Phi", min_value=0., max_value=0.5*pi,
                        step=0.1, value=1.)
radius = st.sidebar.slider("Radius", min_value=2., max_value=6.,
                           step=1., value=3.)

caption = "Diet-NeRF achieves SoTA few-shot learning capacity in 3D model reconstruction. " \
          "Thanks to the 2D supervision by CLIP (aka semantic loss), " \
          "it can render novel and challenging views with ONLY 8 training images, " \
          "outperforming original NeRF!"
st.markdown(f""" <h4> {caption} </h4> """,
            unsafe_allow_html=True)
st.markdown("")

with st.spinner("Rendering Image, it may take 2-3 mins. So, why don't you read our report in the meantime"):
    pred_color, _ = render_predict_from_pose(state, theta, phi, radius)
    im = predict_to_image(pred_color)
    w, _ = im.size
    new_w = int(2*w)
    im = im.resize(size=(new_w, new_w))
    st.image(im, use_column_width=True)