{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "colab_type": "text",
    "id": "Be9yoh-ILfRr"
   },
   "source": [
    "# Hybrid Demucs\n",
    "\n",
    "Feel free to use the Colab version:\n",
    "https://colab.research.google.com/drive/1dC9nVxk3V_VPjUADsnFu8EiT-xnU1tGH?usp=sharing"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 139
    },
    "colab_type": "code",
    "executionInfo": {
     "elapsed": 12277,
     "status": "ok",
     "timestamp": 1583778134659,
     "user": {
      "displayName": "Marllus Lustosa",
      "photoUrl": "https://lh3.googleusercontent.com/a-/AOh14GgLl2RbW64ZyWz3Y8IBku0zhHCMnt7fz7fEl0LTdA=s64",
      "userId": "14811735256675200480"
     },
     "user_tz": 180
    },
    "id": "kOjIPLlzhPfn",
    "outputId": "c75f17ec-b576-4105-bc5b-c2ac9c1018a3"
   },
   "outputs": [],
   "source": [
    "!pip install -U demucs\n",
    "# or for local development, if you have a clone of Demucs\n",
    "# pip install -e ."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {},
    "colab_type": "code",
    "id": "5lYOzKKCKAbJ"
   },
   "outputs": [],
   "source": [
    "# You can use the `demucs` command line to separate tracks\n",
    "!demucs test.mp3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# You can also load directly the pretrained models,\n",
    "# for instance for the MDX 2021 winning model of Track A:\n",
    "from demucs import pretrained\n",
    "model = pretrained.get_model('mdx')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Because `model` is a bag of 4 models, you cannot directly call it on your data,\n",
    "# but the `apply_model` will know what to do of it.\n",
    "import torch\n",
    "from demucs.apply import apply_model\n",
    "x = torch.randn(1, 2, 44100 * 10)  # ten seconds of white noise for the demo\n",
    "out = apply_model(model, x)[0]     # shape is [S, C, T] with S the number of sources\n",
    "\n",
    "# So let see, where is all the white noise content is going ?\n",
    "for name, source in zip(model.sources, out):\n",
    "    print(name, source.std() / x.std())\n",
    "# The outputs are quite weird to be fair, not what I would have expected."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# now let's take a single model from the bag, and let's test it on a pure cosine\n",
    "freq = 440  # in Hz\n",
    "sr = model.samplerate\n",
    "t = torch.arange(10 * sr).float() / sr\n",
    "x = torch.cos(2 * 3.1416 * freq * t).expand(1, 2, -1)\n",
    "sub_model = model.models[3]\n",
    "out = sub_model(x)[0]\n",
    "\n",
    "# Same question where does it go?\n",
    "for name, source in zip(model.sources, out):\n",
    "    print(name, source.std() / x.std())\n",
    "    \n",
    "# Well now it makes much more sense, all the energy is going\n",
    "# in the `other` source.\n",
    "# Feel free to try lower pitch (try 80 Hz) to see what happens !"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# For training or more fun, refer to the Demucs README on our repo\n",
    "# https://github.com/facebookresearch/demucs/tree/main/demucs"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "authorship_tag": "ABX9TyM9xpVr1M86NRcjtQ7g9tCx",
   "collapsed_sections": [],
   "name": "Demucs.ipynb",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.8.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}