import gradio as gr
import fasttext
import html
import numpy as np
import types
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from transformers import AutoTokenizer

# Projektspezifische Module
from lib.bert_regressor import BertMultiHeadRegressor
from lib.bert_regressor_utils import (
    load_model_and_tokenizer,
    predict_flavours,
    #predict_is_review,
    TARGET_COLUMNS,
    ICONS
)

### Stettings ####################################################################

##################################################################################

# offizielles Mirror-Repo mit lid.176.*
lid_path = hf_hub_download(
    repo_id="julien-c/fasttext-language-id",
    filename="lid.176.ftz"
)

lid_model = fasttext.load_model(lid_path)

# robustes predict mit NumPy-2-Fix + Fallback, falls fastText nur Labels liefert
def _predict_np2_compat(self, text, k=1, threshold=0.0, on_unicode_error='strict'):
    out = self.f.predict(text, k, threshold, on_unicode_error)
    # Fälle:
    # 1) (labels, probs)
    # 2) labels-only (einige Builds/SWIG-Versionen)
    if isinstance(out, tuple) and len(out) == 2:
        labels, probs = out
    else:
        labels = out
        # sinnvolle Defaults, falls keine Wahrscheinlichkeiten vorliegen
        if isinstance(labels, (list, tuple)):
            probs = [1.0] * len(labels)
        else:
            labels = [labels]
            probs = [1.0]
    return labels, np.asarray(probs)  # np.asarray statt np.array(copy=False)

# Instanz patchen
lid_model.predict = types.MethodType(_predict_np2_compat, lid_model)

### Check if lang is english #####################################################

def is_eng(review: str):
    lang_labels, lang_probs = lid_model.predict(review)

    print(lang_labels, lang_probs)

    if not lang_labels:  # kein Label zurückgegeben
        return False, 0.0
    
    lang_label = lang_labels[0]
    lang_prob = float(lang_probs[0])

    return lang_label[1] == "__label__en", lang_prob

### Do actual prediction #########################################################

def predict(review: str, mode: str):

    review = (review or "").strip()

    review_is_eng, review_is_eng_prob = is_eng(review)

    if not review:
        # immer zwei Outputs zurückgeben
        return "<i>Please enter a review.</i>", {}

    if mode == "table":
        html_out = f"<b>{html.escape(review)}</b> | <i>{mode}</i>"
        json_out = {}  # leer lassen
        return html_out, json_out
    else:  # "JSON"
        html_out = ""  # leer lassen
        json_out = {
            "review": review,
            "mode": mode,
            "is_en": {
                "is": review_is_eng,
                "prob": review_is_eng_prob
            }
        }
        return html_out, json_out
    
### Create Form interface with Gradio Framework ##################################

iface = gr.Interface(
    fn=predict,
    inputs=[
        gr.Textbox(
            label="Whisky Review", 
            lines=8, 
            placeholder="Enter whisky review",
            value="Honey roasted figs, cloves and sticky toffee pudding with a suggestive smokiness evocative of barbecued bananas."
        ),
        gr.Dropdown(
            label="Display mode",
            choices=[("Table", "table"), ("JSON", "json")],
            value="table"
        )
    ],
    outputs=[
        gr.HTML(label="Table"),
        gr.JSON(label="JSON"),
    ],
    title="Submit Whisky Review for Classification",
    # description="Paste an English whisky review, choose display mode, then submit.",
)

iface.launch()