app.py

# All the datasets will use the same format: a collection of HDF5 files with data cubes
# in t0_fields: scalar fields, like density, pressure, energy
#  the data is of shape (n_trajectories, n_time_steps, x, y)
# in t1_fields: vector fields, like velocity (size=2 => vx, vy)
#  the data is of shape (n_trajectories, n_time_steps, x, y, vx/vy)
# in t2_fields: tensor fields, like ???
#  the data is of shape (n_trajectories, n_time_steps, x, y, d1, d2), with d1, d2 in [0, 1]
#   ie, instead of 1 additional dimension for velocity: a (2,2) matrix where each component
#   (0,0),(1,0),(0,1),(1,1) can be plotted
# Size:
# - n_trajectories: 8 to 256
# - n_time_steps: 101
# - x: 128 to 512
# - y: 128 to 512
# - physical fields: 2 to 8 (density, pressure, energy, velocity…)
from functools import lru_cache

import gradio as gr
import h5py
import numpy as np
from fsspec import url_to_fs
from matplotlib import cm
from PIL import Image

repo_id = "lhoestq/turbulent_radiative_layer_tcool_demo"
set_path = f"hf://datasets/{repo_id}/**/*.hdf5"
fs, _ = url_to_fs(set_path)
paths = fs.glob(set_path)
files = {path: h5py.File(fs.open(path, "rb", cache_type="none"), "r") for path in paths}

def get_scalar_fields(path: str) -> list[str]:
    # TODO: support t1_fields (vector) and t2_fields (tensor)
    return list(files[path]["t0_fields"].keys())

def get_trajectories(path: str, field: str) -> list[int]:
    # The first dimension is the trajectory (8 to 256)
    return list(range(len(files[path]["t0_fields"][field])))

@lru_cache(maxsize=4)
def get_images(path: str, scalar_field: str, trajectory: int) -> list[Image.Image]:
    # The data is of shape (n_trajectories, n_time_steps, x, y)
    out = files[path]["t0_fields"][scalar_field][trajectory]
    out = np.log(out) # not sure why
    out = (out - out.min()) / (out.max() - out.min())
    out = np.uint8(cm.RdBu_r(out) * 255)
    return [Image.fromarray(img) for img in out]

default_scalar_fields = get_scalar_fields(paths[0])
default_trajectories = get_trajectories(paths[0], default_scalar_fields[0])
default_images = get_images(paths[0], default_scalar_fields[0], default_trajectories[0])

with gr.Blocks() as demo:
    gr.Markdown(f"# 💠 HDF5 Viewer for the [{repo_id}](https://huggingface.co/datasets/{repo_id}) Dataset 🌊")
    gr.Markdown(f"Showing files at `{set_path}`")
    with gr.Row():
        files_dropdown = gr.Dropdown(choices=paths, value=paths[0], label="File", scale=4)
        scalar_fields_dropdown = gr.Dropdown(choices=default_scalar_fields, value=default_scalar_fields[0], label="Physical field")
        trajectory_dropdown = gr.Dropdown(choices=default_trajectories, value=default_trajectories[0], label="Trajectory")
    gallery = gr.Gallery(default_images, preview=True, selected_index=len(default_images) // 2)
    gr.Markdown("_Tip: click on the image to go forward or backwards_")

    @files_dropdown.select(inputs=[files_dropdown], outputs=[scalar_fields_dropdown, trajectory_dropdown, gallery])
    def _update_file(path: str):
        scalar_fields = get_scalar_fields(path)
        trajectories = get_trajectories(path, scalar_fields[0])
        images = get_images(path, scalar_fields[0], trajectories[0])
        yield {
            scalar_fields_dropdown: gr.Dropdown(choices=scalar_fields, value=scalar_fields[0]),
            trajectory_dropdown: gr.Dropdown(choices=trajectories, value=trajectories[0]),
            gallery: gr.Gallery(images)
        }
        yield {gallery: gr.Gallery(selected_index=len(default_images) // 2)}
    
    @scalar_fields_dropdown.select(inputs=[files_dropdown, scalar_fields_dropdown], outputs=[trajectory_dropdown, gallery])
    def _update_scalar_field(path: str, scalar_field: str):
        trajectories = get_trajectories(path, scalar_field)
        images = get_images(path, scalar_field, trajectories[0])
        yield {
            trajectory_dropdown: gr.Dropdown(choices=trajectories, value=trajectories[0]),
            gallery: gr.Gallery(images)
        }
        yield {gallery: gr.Gallery(selected_index=len(default_images) // 2)}

    @trajectory_dropdown.select(inputs=[files_dropdown, scalar_fields_dropdown, trajectory_dropdown], outputs=[gallery])
    def _update_trajectory(path: str, scalar_field: str, trajectory: int):
        images = get_images(path, scalar_field, trajectory)
        yield {gallery: gr.Gallery(images)}
        yield {gallery: gr.Gallery(selected_index=len(default_images) // 2)}

demo.launch()