{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"4\"\n",
    "\n",
    "import torch\n",
    "import torchaudio\n",
    "\n",
    "import sys\n",
    "sys.path.insert(0, \"/home/christian/code/christian/scripts\")\n",
    "\n",
    "from train_refiner import HDemucsRefiner, corrupt\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "ckpt_path = \"/app/suno/christian/checkpoints/refiner-v1/2025-05-02_20-38-55_s2226/last_ckpt.pt\" \n",
    "\n",
    "\n",
    "\n",
    "ckpt = torch.load(ckpt_path)\n",
    "#model = Refiner(**ckpt[\"run_config\"][\"model\"])\n",
    "model = HDemucsRefiner()\n",
    "state_dict = ckpt[\"model\"]\n",
    "new_state_dict = {}\n",
    "for key, value in state_dict.items():\n",
    "    new_key = key.replace(\"module.\", \"\")\n",
    "    new_state_dict[new_key] = value\n",
    "model.load_state_dict(new_state_dict)\n",
    "model.eval()\n",
    "model.cuda()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# load audio and crop to 10s\n",
    "import IPython\n",
    "from suno_utils.utils.text import read_jsonl\n",
    "meta_filepath = \"/mnt/localdisk/tmp_cjs/v2-infill-data-v1/metas_val.jsonl\"\n",
    "metas = read_jsonl(meta_filepath)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pyloudnorm as pyln\n",
    "import numpy as np\n",
    "\n",
    "idx = np.random.randint(0, len(metas))\n",
    "meta = metas[idx]\n",
    "local_filepath = meta[\"input_filepath\"]\n",
    "target_filepath = meta[\"target_filepath\"]\n",
    "\n",
    "print(local_filepath, target_filepath)\n",
    "audio, sr = torchaudio.load(local_filepath)\n",
    "target_audio, sr = torchaudio.load(target_filepath)\n",
    "\n",
    "print(audio.shape, target_audio.shape)\n",
    "\n",
    "# crop to 10s\n",
    "min_len = min(audio.shape[1], target_audio.shape[1])\n",
    "start_idx = np.random.randint(0, min_len - 262144*2)\n",
    "end_idx = start_idx + 262144*2\n",
    "#start_idx = 262144*2\n",
    "#audio = audio[:, start_idx:end_idx]\n",
    "target_audio = target_audio[:, start_idx:end_idx]\n",
    "\n",
    "input_audio = corrupt(target_audio, sr)\n",
    "\n",
    "meter = pyln.Meter(sr)\n",
    "\n",
    "# loudness normalize both to -24\n",
    "#audio_loudness = meter.integrated_loudness(audio.permute(1, 0).numpy())\n",
    "#target_loudness = meter.integrated_loudness(target_audio.permute(1, 0).numpy())#\n",
    "#audio = audio * (10**((-24 - audio_loudness) / 20))\n",
    "#target_audio = target_audio * (10**((-24 - target_loudness) / 20))\n",
    "\n",
    "\n",
    "with torch.no_grad():\n",
    "    input_audio = input_audio.cuda()\n",
    "    input_audio = input_audio.unsqueeze(0)\n",
    "    output = torch.tanh(model(input_audio))\n",
    "    output = output.squeeze(0)\n",
    "    output = output.cpu()\n",
    "\n",
    "print(output.min(), output.max())\n",
    "\n",
    "IPython.display.display(IPython.display.Audio(target_audio.squeeze(0).cpu().numpy(), rate=sr, normalize=True))\n",
    "IPython.display.display(IPython.display.Audio(input_audio.squeeze(0).cpu().numpy(), rate=sr, normalize=True))\n",
    "IPython.display.display(IPython.display.Audio(output.cpu().numpy(), rate=sr, normalize=True))\n",
    "\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "suno_env",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}