# -*- coding: UTF-8 -*-
import gradio as gr
import torch, torchaudio
from timeit import default_timer as timer
from data_setups import audio_preprocess, resample

device = "cuda" if torch.cuda.is_available() else "cpu"

SAMPLE_RATE = 44100
AUDIO_LEN = 2.90

model = torch.load("models/torch_efficientnet_fold2_CNN.pth", map_location=torch.device('cpu'))

CHINESE_LABELS = [
    "大提琴", "單簧管", "長笛", "民謠吉他", "電吉他", "風琴", "鋼琴", "薩克斯風", "喇叭", "小提琴", "人聲"
]

LABELS = [
    "Cello", "Clarinet", "Flute", "Acoustic Guitar", "Electric Guitar", "Organ", "Piano", "Saxophone", "Trumpet", "Violin", "Voice"
]

example_list = [
    "samples/guitar_acoustic.wav",
    "samples/piano.wav",
    "samples/violin.wav",
    "samples/flute.wav"
]

def predict(audio_path):
    start_time = timer()
    wavform, sample_rate = torchaudio.load(audio_path)
    wav = resample(wavform, sample_rate, SAMPLE_RATE)
    if len(wav) > int(AUDIO_LEN * SAMPLE_RATE):
        wav = wav[:int(AUDIO_LEN * SAMPLE_RATE)]
    else:
        print(f"input length {len(wav)} too small!, need over {int(AUDIO_LEN * SAMPLE_RATE)}")
        return
    # input Preprocessing
    img = audio_preprocess(wav, SAMPLE_RATE).unsqueeze(0)
    model.eval()
    with torch.inference_mode():
        pred_probs = torch.softmax(model(img), dim=1)
    pred_labels_and_probs = {CHINESE_LABELS[i]: float(pred_probs[0][i]) for i in range(len(CHINESE_LABELS))}
    pred_time = round(timer() - start_time, 5)
    return pred_labels_and_probs, pred_time


title = "樂器辨識🎺🎸🎹🎻"
description = "使用IRMAS資料集訓練的深度學習模型，可辨識11種不同樂器，包含「大提琴, 單簧管, 長笛, 民謠吉他, 電吉他, 風琴, 鋼琴, 薩克斯風, 喇叭, 小提琴, 人聲」"
article = ""

demo = gr.Interface(fn=predict,
                    inputs=gr.Audio(type="filepath"),
                    outputs=[gr.Label(num_top_classes=11, label="Predictions"), 
                             gr.Number(label="Prediction time (s)")],
                    examples=example_list, 
                    title=title,
                    description=description,
                    article=article)

demo.launch(debug=False)