app.py

import json
import math
import os

import gdown
import streamlit as st


from demo.src.models import load_trained_model
from demo.src.utils import predict_to_image, render_predict_from_pose

st.set_page_config(page_title="DietNeRF")

with open("./demo/config.json") as f:
    cfg = json.loads(f.read())

MODEL_DIR = "models"

SCENES_LIST = ["Mic", "Chair", "Lego", "Drums", "Ship", "Hotdog"]
# random_index = random.randint(0, len(SCENES_LIST) - 1)


def select_model(obj_select):
    DIET_NERF_MODEL_NAME = cfg[obj_select]["DIET_NERF_MODEL_NAME"]
    DIET_NERF_FILE_ID = cfg[obj_select]["DIET_NERF_FILE_ID"]

    NERF_MODEL_NAME = cfg[obj_select]["NERF_MODEL_NAME"]
    NERF_FILE_ID = cfg[obj_select]["NERF_FILE_ID"]

    return DIET_NERF_MODEL_NAME, DIET_NERF_FILE_ID, NERF_MODEL_NAME, NERF_FILE_ID


pi = math.pi

st.title("DietNeRF")
st.sidebar.markdown(
    """
<style>
.aligncenter {
    text-align: center;
}
</style>
<p class="aligncenter">
    <img src="https://user-images.githubusercontent.com/77657524/126361638-4aad58e8-4efb-4fc5-bf78-f53d03799e1e.png" width="420" height="400"/>
</p>
""",
    unsafe_allow_html=True,
)
st.sidebar.markdown(
    """
<p style='text-align: center'>
<a href="https://github.com/codestella/putting-nerf-on-a-diet" target="_blank">GitHub</a> | <a href="https://www.notion.so/DietNeRF-Putting-NeRF-on-a-Diet-4aeddae95d054f1d91686f02bdb74745" target="_blank">Project Report</a>
</p>
    """,
    unsafe_allow_html=True,
)
st.sidebar.header("SELECT YOUR VIEW DIRECTION")
theta = st.sidebar.slider(
    "Rotation (Left to Right)",
    min_value=-pi,
    max_value=pi,
    step=0.5,
    value=0.0,
    help="Rotational angle in Vertical direction (Theta)",
)
phi = st.sidebar.slider(
    "Rotation (Bottom to Top)",
    min_value=0.0,
    max_value=0.5 * pi,
    step=0.1,
    value=1.0,
    help="Rotational angle in Horizontal direction (Phi)",
)
radius = st.sidebar.slider(
    "Distance (Close to Far)",
    min_value=2.0,
    max_value=6.0,
    step=1.0,
    value=3.0,
    help="Distance between object and the viewer (Radius)",
)
caption = (
    "`DietNeRF` achieves state-of-the-art few-shot learning capacity in 3D model reconstruction. "
    "Thanks to the 2D supervision by `CLIP (aka. Semantic Consisteny Loss)`, "
    "it can render novel and challenging views with `ONLY 8 training images`, "
    "**outperforming** original [NeRF](https://www.matthewtancik.com/nerf)!"
)
st.markdown(caption)
st.markdown(
    "> 📒 **NOTE**: To get a detailed comparison of differences in results between `DietNeRF` and `NeRF`, you can take a look at the "
    "[Experimental Results](https://www.notion.so/DietNeRF-Putting-NeRF-on-a-Diet-4aeddae95d054f1d91686f02bdb74745#0f6bc8f1008d4765b9b4635999626d4b) "
    "section in our project report."
)


obj_select = st.selectbox("Select a Scene", SCENES_LIST, index=0)
DIET_NERF_MODEL_NAME, DIET_NERF_FILE_ID, NERF_MODEL_NAME, NERF_FILE_ID = select_model(obj_select)


@st.cache(show_spinner=False)
def download_diet_nerf_model():
    os.makedirs(MODEL_DIR, exist_ok=True)
    diet_nerf_model_path = os.path.join(MODEL_DIR, DIET_NERF_MODEL_NAME)
    url = f"https://drive.google.com/uc?id={DIET_NERF_FILE_ID}"
    gdown.download(url, diet_nerf_model_path, quiet=False)
    print(f"Model downloaded from google drive: {diet_nerf_model_path}")


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


diet_nerf_model_path = os.path.join(MODEL_DIR, DIET_NERF_MODEL_NAME)
if not os.path.isfile(diet_nerf_model_path):
    download_diet_nerf_model()


diet_nerf_model, diet_nerf_state = fetch_diet_nerf_model()

st.markdown("")

with st.spinner("Rendering view..."):
    with st.spinner(
        ":information_source: **INFO**: It may take around 1-2 mins to render the view. "
        "In the meantime, why don't you take a look at our "
        "[project report](https://www.notion.so/DietNeRF-Putting-NeRF-on-a-Diet-4aeddae95d054f1d91686f02bdb74745), "
        "if you haven't already :slightly_smiling_face:"
    ):
        dn_pred_color, _ = render_predict_from_pose(diet_nerf_state, theta, phi, radius)
        dn_im = predict_to_image(dn_pred_color)
        dn_w, _ = dn_im.size
        dn_new_w = int(2 * dn_w)
        dn_im = dn_im.resize(size=(dn_new_w, dn_new_w))
        st.markdown(
            "> 📒 **NOTE**: The rendered view does not fully reflect the true quality of the view generated by the model "
            "because it has been downsampled to speedup the process."
        )
        st.markdown(f"""<h4 style='text-align: center'>Rendered view for {obj_select}</h4>""", unsafe_allow_html=True)
        st.image(dn_im, use_column_width=True)