Permalink
Switch branches/tags
Nothing to show
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
59 lines (47 sloc) 6.17 KB

Hosted Checkpoints

Short link: https://goo.gl/magenta/js-checkpoints

You can load the pre-trained checkpoints below in your app directly from our server with the links provided. If you would like to download the checkpoint to use locally or host yourself, pass the link to our checkpoint downloader script. For example, to download the basic_rnn checkpoint, you would run:

python ./scripts/checkpoint_downloader https://storage.googleapis.com/magentadata/js/checkpoints/music_rnn/basic_rnn ./my-checkpoints/

JSON Index

A JSON index of available checkpoints is at https://goo.gl/magenta/js-checkpoints-json, formatted as a list of entries with the following interface:

interface Checkpoint {
  id: string;  // A unique id for this checkpoint.
  model: 'MusicRNN'|'MusicVAE';  // The model class.
  sizeMb: number;  // The size of the weights in megabytes.
  description: string;  // A short human-readable description of the trained model.
  url: string;  // Path to the checkpoint directory.
}

While we do not plan to remove any of the current checkpoints, we will be adding more in the future.

If your application has a high QPS, you must mirror these files on your own server.

Table

ID Model Description Size MB URL
drums_2bar_lokl_small MusicVAE A 2-bar, 9-class onehot drum model with a strong prior (low KL divergence), which is better for sampling. Less accurate, but smaller in size than full model. 18.5 Right Click to Copy
drums_2bar_hikl_small MusicVAE A 2-bar, 9-class onehot drum model with a weak prior (higher KL divergence), which is better for reconstructions and interpolations. Less accurate, but smaller in size than full model. 18.5 Right Click to Copy
drums_2bar_nade_9_q2 MusicVAE A 2-bar, 9-class multilabel drum model with a NADE decoder. Quantized to 2-byte weights. 27.6 Right Click to Copy
mel_2bar_small MusicVAE A 2-bar, 90-class onehot melody model. Less accurate, but smaller in size than full model. 17.7 Right Click to Copy
mel_4bar_small_q2 MusicVAE A 4-bar, 90-class onehot melody model. Less accurate, but smaller in size than full model. Quantized to 2-byte weights. 26.5 Right Click to Copy
mel_chords MusicVAE A 2-bar, 90-class onehot melody model with chord conditioning. Quantized to 2-byte weights. 17.6 Right Click to Copy
mel_16bar_small_q2 MusicVAE A 16-bar, 90-class onehot melody model with a 16-step conductor level. Less accurate, but smaller in size than full model. Quantized to 2-byte weights. 23.5 Right Click to Copy
trio_4bar_lokl_small_q1 MusicVAE A 4-bar, 'trio' model for melody, bass, and drums, with a 4-step conductor level. Trained with a strong prior (low KL divergence), which is better for sampling. Less accurate, but smaller in size than full model. Quantized to 1-byte weights. 17.6 Right Click to Copy
multitrack MusicVAE A 1-bar multitrack model, trained with 64 free bits. Quantized to 1-byte weights. 26.4 Right Click to Copy
multitrack_med MusicVAE A larger 1-bar multitrack model, trained with 64 free bits. Quantized to 1-byte weights. 95.9 Right Click to Copy
multitrack_med_fb256 MusicVAE A larger 1-bar multitrack model, trained with 256 free bits. Quantized to 1-byte weights. 95.9 Right Click to Copy
multitrack_chords MusicVAE A 1-bar chord-conditioned multitrack model, trained with 64 free bits. Quantized to 1-byte weights. 26.9 Right Click to Copy
multitrack_med_chords MusicVAE A larger 1-bar chord-conditioned multitrack model, trained with 64 free bits. Quantized to 1-byte weights. 96.9 Right Click to Copy
multitrack_med_chords_fb256 MusicVAE A larger 1-bar chord-conditioned multitrack model, trained with 256 free bits. Quantized to 1-byte weights. 96.9 Right Click to Copy
basic_rnn MusicRNN A 36-class onehot MelodyRNN model. Converted from http://download.magenta.tensorflow.org/models/basic_rnn.mag. 13.0 Right Click to Copy
drum_kit_rnn MusicRNN A 9-class onehot DrumsRNN model. Converted from http://download.magenta.tensorflow.org/models/drum_kit_rnn.mag. 11.9 Right Click to Copy
chord_pitches_improv MusicRNN A 36-class onehot melody ImprovRNN model conditioned on chords as described at https://github.com/tensorflow/magenta/tree/master/magenta/models/improv_rnn#chord-pitches-improv. Converted from http://download.magenta.tensorflow.org/models/chord_pitches_improv.mag. 5.6 Right Click to Copy
onsets_frames_uni OnsetsAndFrames A unidirectional piano transcription model. 60 Right Click to Copy
onsets_frames_uni_q2 OnsetsAndFrames A unidirectional piano transcription model. Quantized to 2-byte weights. 30 Right Click to Copy