app.py

import gradio as gr
import numpy as np
import onnxruntime as ort
import json
from cryptography.fernet import Fernet
import os
from dotenv import load_dotenv

load_dotenv()

# Model load
key = os.getenv("ONNX_KEY")
cipher = Fernet(key)

with open("species_bag.onnx.encrypted", "rb") as f:
    encrypted = f.read()
    decrypted = cipher.decrypt(encrypted)
    ort_session = ort.InferenceSession(decrypted)

# Initialize ONNX session
input_name = ort_session.get_inputs()[0].name
output_name = ort_session.get_outputs()[0].name

# Load mappings
with open("idx2spec.json", "r") as f:
    idx2spec = json.load(f)

with open("spec2idx.json", "r") as f:
    spec2idx = json.load(f)

with open("spec2key.json", "r") as f:
    spec2key = json.load(f)

baseurl = "https://www.gbif.org/species/"

def predict_species(selected_species, n_hits=10):
    if not selected_species:
        return "", ""
    
    # Convert species names to indices using spec2idx
    input_indices = [int(spec2idx[name]) for name in selected_species]
    
    # Model inference
    input_np = np.array(input_indices, dtype=np.int64).reshape(1, -1)
    output = ort_session.run([output_name], {input_name: input_np})[0][0]
    
    # Get top predictions
    top_indices = output.argsort()[-n_hits:][::-1]
    top_scores = output[top_indices]
    
    # Format selected species with links
    selected_html = ["<a href='{}{}' target='_blank'>{}</a>".format(
        baseurl, spec2key[species], species
    ) for species in selected_species]
    
    # Format predictions with species names, links and scores
    predictions_html = ["<a href='{}{}' target='_blank'>{}</a> ({:.1f}%)".format(
        baseurl, spec2key[idx2spec[str(idx)]], idx2spec[str(idx)], 100*score
    ) for idx, score in zip(top_indices, top_scores)]
    
    return "<br>".join(selected_html), "<br>".join(predictions_html)

# Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("## Danmarks planter - hvem mangler?")
    gr.Markdown("*Sammensæt et plantesamfund og få forslag til andre arter der passer ind. Vælg mellem 3199 danske terrestriske og akvatiske planter.*")

    with gr.Row():
        species_dropdown = gr.Dropdown(
            choices=sorted(spec2idx.keys()),
            multiselect=True,
            label="Find arter",
        )
    
    with gr.Row():
        with gr.Column(scale=5, min_width=200):
            selected_output = gr.HTML(
                label="Arter",
                show_label=True
            )
        with gr.Column(scale=5, min_width=200):
            
            predictions_output = gr.HTML(
                label="Top hits",
                show_label=True
            )
        with gr.Column(scale=1, min_width=100):
            n_hits = gr.Number(10, label="Antal hits", minimum=1, maximum=100)
            add_button = gr.Button("Tilføj top hit", scale=8)

    gr.Markdown("Forslag er baseret på et neuralt netværk trænet til at forudsige de mest sandsynlige arter som mangler i et plantesamfund. Trænet på stort datasæt af plantesamfund registreret i Danmark (**4.3 millioner registreringer af 3199 arter/slægter/varianter i mere end 180.000 undersøgelser**).")
    gr.Markdown("App og model af Kenneth Thorø Martinsen.")

    def add_top_prediction(selected_species):
        if not selected_species:
            return selected_species
        top_prediction = predict_species(selected_species)[1].split("<br>")[0].split(" (")[0]
        top_prediction = top_prediction.split(">")[1].split("<")[0]  # Extract species name from HTML
        if top_prediction not in selected_species:
            selected_species.append(top_prediction)
        return selected_species

    species_dropdown.change(
        predict_species,
        inputs=[species_dropdown, n_hits],
        outputs=[selected_output, predictions_output]
    )

    add_button.click(
        add_top_prediction,
        inputs=[species_dropdown],
        outputs=[species_dropdown]
    )

if __name__ == "__main__":
    demo.launch()