Improve text of VST training. Update version.

jesseengel · Magenta Team · commit d1e9b555bf7e · 2022-04-22T10:39:19.000-07:00
PiperOrigin-RevId: 443703095
diff --git a/ddsp/colab/demos/Train_VST.ipynb b/ddsp/colab/demos/Train_VST.ipynb
@@ -23,45 +23,75 @@
         "#     limitations under the License.\n",
         "# ------------------------------------------------------------------------------\n",
         "\n",
-        "#@title Train your own DDSP-VST Model\n",
-        "#@markdown Just press the &#9654;&#65039; button!\n",
+        "#@markdown #  Train your own DDSP-VST Model \n",
+        "#@markdown &#127931;&#127930;&#127928;&#127925; [g.co/magenta/train-ddsp-vst](g.co/magenta/train-ddsp-vst)\n",
         "\n",
-        "#@markdown \u003cbr/\u003e Custom models can train on as little as 10 minutes of audio (`.wav` or `.mp3`). Best results from \"monophonic\" (only one note at a time) audio from a single recording session (same mic, same reverb).\n",
+        "#@markdown \u003cbr/\u003e \n",
         "\n",
-        "#@markdown Training typically takes ~2-3 hours with free Colab, and less than an hour with ColabPro.\n",
+        "#@markdown ## Instructions\n",
         "\n",
         "\n",
-        "#@markdown We recommend using Google Drive for training to load faster and save your model during training. Just create a folder on your drive with your audio files in it, and select the folder. If you don't use drive, you can still upload audio through the browser (slower) and download the final trained model.\n",
+        "#@markdown * Create a folder in Google Drive with your training audio (`.wav` or `.mp3`)\n",
         "\n",
-        "#@markdown Colab often kicks people off after ~12 hours, but hopefully that shouldn't be a problem.\n",
+        "Name = 'My Instrument' #@param {type:\"string\"}\n",
+        "Name = Name.replace(' ', '_')\n",
         "\n",
-        "#@markdown After training, it should automatically export and download your model as `{my_name}.tflite` that you can use by dropping in the VST custom models folder. If it doesn't automatically download, you can find the file in the `ddsp-training-{date-time}/export` folder either on this page (click the &#128193; icon on the left), or in the folder you selected from your drive.\n",
+        "\n",
+        "#@markdown * Press the &#9654;&#65039; button in the upper left!\n",
+        "\n",
+        "#@markdown * Login to your Google account when asked\n",
+        "\n",
+        "#@markdown *  Select your folder with the file chooser below when asked\n",
+        "\n",
+        "#@markdown *  Wait (with this window open) for training to finish and download the model\n",
+        "\n",
+        "#@markdown *  If something breaks, resume training by refreshing this page, press &#9654;&#65039;, and choose the same folder\n",
         "\n",
         "\n",
-        "#@markdown \u003cbr/\u003e \u003cbr/\u003e\n",
-        "#@markdown Name your model!\n",
-        "Name = 'MyInstrument' #@param {type:\"string\"}\n",
-        "Name = Name.replace(' ', '_')\n",
         "\n",
         "#@markdown \u003cbr/\u003e\n",
-        "#@markdown Use Google Drive for training?\n",
-        "Google_Drive = True #@param {type:\"boolean\"}\n",
+        "\n",
+        "#@markdown \u003cbr/\u003e\n",
+        "\n",
+        "#@markdown ## Data\n",
+        "#@markdown Custom models can train on as little as 10 minutes of audio (`.wav` or `.mp3`). You can get the best results from \"monophonic\" (only one note at a time) audio from a single recording session (same mic, same reverb). All of your data is private, used locally, and erased as soon as your colab session ends.\n",
+        "\n",
+        "\n",
+        "#@markdown ## Training\n",
+        "#@markdown Training typically takes ~2-3 hours with free Colab, and less than an hour with ColabPro+. Free colab can sometimes disconnects before models finish training, but there are some unofficial [ways around this](https://stackoverflow.com/questions/57113226/how-to-prevent-google-colab-from-disconnecting). If you do get disconnected, don't worry, just press play again and choose the same folder. The training will resume where it left off.\n",
+        "\n",
+        "\n",
+        "\n",
+        "#@markdown ## Export\n",
+        "\n",
+        "#@markdown After training, it should automatically export and download your model as `{my_name}.tflite` that you can use by dropping in the VST custom models folder, which you can access by pushing the 'Models' button within the plugin itself. If it doesn't automatically download, you can find the file in the `ddsp-training-{date-time}/export` folder, either on this page (click the &#128193; icon on the left), or in the folder you selected from your drive.\n",
+        "\n",
+        "##@markdown We recommend using Google Drive to load data faster and save your model during training. Just create a folder on your drive with your audio files in it, and select the folder. If you don't use drive, you can still upload audio through the browser (slower) and download the final trained model.\n",
         "\n",
         "\n",
         "\n",
         "#@markdown \u003cbr/\u003e \u003cbr/\u003e\n",
-        "#@markdown ### Advanced Options\n",
+        "#@markdown ## Advanced Options\n",
         "\n",
-        "#@markdown \u003ca href=\"https://colab.research.google.com/github/magenta/ddsp/blob/main/ddsp/colab/demos/Train_VST.ipynb\" target=\"_parent\"\u003e\u003cimg src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/\u003e\u003c/a\u003e\n",
+        "##@markdown \u003ca href=\"https://colab.research.google.com/github/magenta/ddsp/blob/main/ddsp/colab/demos/Train_VST.ipynb\" target=\"_parent\"\u003e\u003cimg src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/\u003e\u003c/a\u003e\n",
+        "\n",
+        "#@markdown \u003cbr/\u003e\n",
+        "#@markdown  Usually we will produce good results after training between 30k-50k steps, but your results may vary depending on your audio files/instrument. Too few steps will often have the model sound bland/generic, too many steps can often lead to more \"sputtering\" and big volume fluctuations\n",
         "\n",
         "Training_Steps = 30000 #@param {type:\"integer\"}\n",
         "\n",
         "#@markdown \u003cbr/\u003e\n",
-        "#@markdown Ignore previous checkpoints and start a fresh run\n",
+        "#@markdown Ignore previous checkpoints in the folder and start a fresh run from step 0\n",
         "\n",
         "Ignore_Previous = False #@param {type:\"boolean\"}\n",
         "\n",
         "\n",
+        "#@markdown \u003cbr/\u003e\n",
+        "#@markdown Use Google Drive for training? Otherwise, loads audio from browser, which is much slower\n",
+        "Google_Drive = True #@param {type:\"boolean\"}\n",
+        "\n",
+        "\n",
+        "\n",
         "# Sample_Rate = '16kHz'  #@param ['16kHz', '32kHz', '48kHz']\n",
         "# Sample_Rate = {'16kHz': 16000, '32kHz': 32000, '48kHz': 48000}[Sample_Rate]\n",
         "# Model_Gin_File = 'models/vst/vst.gin'\n",
@@ -74,7 +104,7 @@
         "# ------------------------------------------------------------------------------\n",
         "print('Installing DDSP...')\n",
         "print('This should take about 2 minutes...')\n",
-        "!pip install -U ddsp[data_preparation]==3.3.4 \u0026\u003e /dev/null\n",
+        "!pip install -U ddsp[data_preparation]==3.4.0 \u0026\u003e /dev/null\n",
         "!pip install ipyfilechooser \u0026\u003e /dev/null\n",
         "\n",
         "\n",
@@ -304,6 +334,7 @@
     "accelerator": "GPU",
     "colab": {
       "collapsed_sections": [],
+      "last_runtime": {},
       "name": "Train_VST.ipynb",
       "private_outputs": true,
       "provenance": []
