Based on magenta's GitHub README, set up magenta's development environment without using Docker. I would like to generate a song with Basic RNN.
Download and use the trained model.
If you want to use Docker easily, please refer to here. Try making a song with TensorFlow's art / music generation project "magenta".
To prepare the development environment for Magenta -Magenta's GitHub repository ・ Bazel ・ TensorFlow You need three.
Also, Magenta seems to support only Python 2.7 at the moment, so I created a Python 2.7 environment with virtualenv on a Mac and set it up.
git clone https://github.com/tensorflow/magenta.git
https://www.bazel.io/versions/master/docs/install.html
$ brew install bazel
TensorFlow https://www.tensorflow.org/versions/master/get_started/os_setup.html
When installing TensorFlow, please check whether your PC is CPU only or GPU can be used. It seems that TensorFlow supports GPU called CUDA, so if you have a Mac, you may use that. Reference: OSX GPU is supported by Tensorflow
I proceeded with CPU only.
$ export TF_BINARY_URL=https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-0.10.0-py2-none-any.whl
$ pip install --upgrade $TF_BINARY_URL
Let's run the test after setting. Go to / magenta / under git clone and hit the following command.
bazel test //magenta/...
If the following message appears, it is successful and the setting is complete.
.........
INFO: Found 65 targets and 27 test targets...
INFO: Elapsed time: 43.359s, Critical Path: 35.97s
//magenta/common:concurrency_test PASSED in 5.4s
//magenta/interfaces/midi:midi_hub_test PASSED in 14.8s
(Omission)
Executed 27 out of 27 tests: 27 tests pass.
There were tests whose specified size is too big. Use the --test_verbose_timeout_warnings command line option to see which ones these are.
Magenta has three types of learning models: Basic RNN, LookBack RNN, and Attention RNN. This time, I will try the method of Basic RNN.
This model is a melody that produces only one note at the same timing. I won't do it in this post, but if you put in MIDI and learn by yourself, it seems that only one sound is considered even if there are sounds that sound at the same timing with that MIDI.
You can download it from here.
After downloading, please place it anywhere.
Set the path of the trained model downloaded above with "Absolute path".
BUNDLE_PATH=<absolute path of basic_rnn.mag>
It can be generated with the following command.
bazel run //magenta/models/basic_rnn:basic_rnn_generate -- \
--bundle_file=${BUNDLE_PATH} \
--output_dir=/tmp/basic_rnn/generated \
--num_outputs=10 \
--num_steps=128 \
--primer_melody="[60]"
The options seem to have the following meanings:
--output_dir=/tmp/basic_rnn/generated \
You can decide where to generate the song.
--num_outputs=10 \
You can decide how many songs to make.
--num_steps=128 \
Set how many notes and how many measures to make a song. A song of 8 bars is made with 128. I think that 16 sounds can be inserted per bar. (The definition of words such as "bar" is ambiguous ...)
--primer_melody="[60]"
It's what sound the song starts with.
As an example
--primer_melody="[60, -2, 60, -2, 67, -2, 67, -2]”
This stands for "Dodo Soso" and will now generate songs that start with a glittering star intro.
(By the way, since the above is 8 notes, it means that half of the first bar is set.)
The numbers are
-2：no event
-1：note-off event
It is set like this, and I do not understand it in words, but
In the case of [60, -2, 60, -2, 67, -2, 67, -2], "Dodo Soso" and the sound before it continue, whereas
In the case of [60, -1, 60, -1, 67, -1, 67, -1], the sound was interrupted as "do do so so".
Is it the difference between "doing nothing" and "turning off the sound"?
60: de 67: So I think that the number of keys on the piano indicates the sound.
When executed, a song will be generated.
The following songs have been made. It started with the intro of "Do Do So So" and made a 16-bar song. Sample
Here is the first one with "Dodo Soso". Sample
The atmosphere changes quite a bit, and you can feel that the intro settings have an effect on the subsequent melody.
Thank you very much.
Recommended Posts