Migration to python3 and increase chance to make it run on your own machine

[benelot]

Migrate to python3 and pure keras, clean up and beautify the code, structure input and output data and added a requirement.txt and some nice READMEs.

I hope this serves some people to get this running.

Note: It seems that the compiled version in the repo is from some other state not included in the repository and has some very minor changes such as the bottom-left buttons for different instruments (as far as I could tell) (it is a rewrite in C++, check here: https://github.com/HackerPoet/Composer/issues/22).

[benelot]

I should also add python2 compatibility for print at least. But probably I used some other non python2 syntax, I am not sure.

[benelot]

I should add to the documentation how to install it.

[benelot]

I should add how to load the saved slider settings.

[benelot]

I should add timestamps to the files written to disk.

[HackerPoet]

How stable would you say this PR is right now? I haven't been able to test it myself, but if you were able to parse the data, train a network, and use the generator successfully, then I would feel comfortable merging this in.

[benelot]

I did some more stuff since then, but I think it is not in this branch. I will test it again and give you the good to merge. At the time I made this PR, I could properly load and parse the data, train the net and use the generator.

Just in case you are interested: In some continued work of mine, I also added pytorch training and composer support. Should I make you a separate PR or do you want to keep your project as is and just get the migration?

[ghost]

Hello,

I'm using Python 3.7.3; and I installed the appropriate libraries listed in the requirements.txt you included in your pull request.
However, there was a deviation in using "pip install" to get the library you listed for numpy.

I was trying to use pip install numpy==1.14.3 but since there were issues trying to get that version, I installed numpy 1.16.4 instead. Would this cause instability?

The main issue I'm encountering is being able to properly load songs. Is there more instructions on how to properly use the load_songs.py script?

I created a folder titled "Music" into the directory with all the other python files from your pull request, and I have 150 MIDI files in there. However, when I try running load_songs.py the command prompt shows that no samples are saved (see the image attachment).

Is there a reason this is happening?

[benelot]

So the problem of saving 0 samples is just that the code did not find your midi files. The folder must be called data/raw. It will then place the preprocessed samples into data/interim. The update to numpy==1.16.4 is correct and should not cause any instability. I will finish up the PR soon.

[ghost]

Awesome! Thanks for the fast response. That helped me to properly load the MIDI files. I was able to run some test trials of load_songs.py -> train.py -> live_edit.py and I ran across one more issue. I used a MIDI sample size of 10 files, as well as a sample size of 400 files.

However, after training and running both models (the 10-file size first, the 400-file size second), they both end up condensing all notes to the start of each measure. I did not alter the source code, so 2000 epochs in total during each training phase. Any suspicions on why this is happening?

[HackerPoet]

That's typical when all sliders are zeroed at the center. Does it still happen when you generate random songs with 'R' or adjust the sliders?

[ghost]

Sorry I should’ve specified. This issue is consistent even when changing the feature-sliders and the note-certainty generation level.

[HackerPoet]

Assuming you had WRITE_HISTORY on, did any of the earlier epochs produce music that had notes outside the first beat?

[ghost]

Unfortunately no, it seems that the first beat is where all generated notes are restricted to.
Here's a couple snapshots of what the 10th epoch looks like (trained with 400 songs).

[HackerPoet]

That's strange. Any other info would be helpful. What did your loss look like? Mine was around 0.003 and converged quickly. If you're getting a similar loss, then it may be a bug with generating the music. If not, then the network is having trouble converging. You may need to adjust the parameters of the network to compensate for your smaller dataset (mine was 10x larger at over 4000 midi files).

[ghost]

Here's the loss after training 2000 sessions with 400 MIDI files. I'll adjust the parameters, scrape for more files to increase the sample size and report back here with the results.

[HackerPoet]

Okay, that does looks like over-fitting to the smaller dataset. Try increasing the dropout rate DO_RATE to 0.25 or 0.5 to see if that helps. Also, just as a sanity check, can you confirm that the test song test.mid became nearly identical to a song from your sample set?

[ghost]

Finally had some time to do a few more tests.
I posted the results in image form, but all 4 of my training tests ultimately faced the same problem, where all notes generated are condensed into the first beat of each measure. The choice of parameters for testing are arbitrary, but I had limited time to try out a few different things:

• 100 epochs, 400 MIDI files for training, DO rate = 0.1
• 50 epochs, 400 MIDI files for training, DO rate = 0.5
• 100 epochs, 400 MIDI files for training, DO rate = 0.25
• 100 epochs, 3000 MIDI files for training, DO rate = 0.5

The last image I used Euphony (MIDI visual player) to try playing the test.mid file from the fourth trial (100 epochs // 3000 MIDI files // DO rate 0.5).
To clarify, those condensed notes seen in the last image occurs 16 times (notes are only appearing to be generated once per measure after training).

[benelot]

Thanks for testing. I am currently on holiday, but if I am back I can give you some test to perform with a simple dataset (bach piano dataset works well for me). Having all notes at the beginning is usually if you did not train for long enough. We will see.

On Fri, Jun 14, 2019, 02:15 moonflowAeon [email protected] wrote:

Finally had some time to do a few more tests. I posted the results in image form, but all 4 of my training tests ultimately faced the same problem, where all notes generated are condensed into the first beat of each measure.

• 100 epochs, 400 MIDI files for training, DO rate = 0.1
• 50 epochs, 400 MIDI files for training, DO rate = 0.5
• 100 epochs, 400 MIDI files for training, DO rate = 0.25
• 100 epochs, 3000 MIDI files for training, DO rate = 0.5

The last image I used Euphony (MIDI visual player) to try playing the test.mid files from the fourth trial (100 epochs // 3000 MIDI files // DO rate 0.5). To clarify, those condensed notes seen in the last image occurs 16 times (notes are only appearing to be generated once per measure after training).

[image: 4 test results] https://user-images.githubusercontent.com/51399704/59474913-cd854a00-8e16-11e9-9fc1-df72f6e2979c.png [image: 16 consolidated notes test (from t4 epoch 100)] https://user-images.githubusercontent.com/51399704/59474920-d249fe00-8e16-11e9-8066-ffaa8d8fb366.png

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/HackerPoet/Composer/pull/21?email_source=notifications&email_token=AAXXXK3PF2U47NDZ2YBLVVLP2LPL7A5CNFSM4HLTBOD2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODXVLZ7A#issuecomment-501923068, or mute the thread https://github.com/notifications/unsubscribe-auth/AAXXXK4NWWKNTAT3JHOJ7L3P2LPL7ANCNFSM4HLTBODQ .

[ghost]

Sounds good benelot, I'll do my best to keep trying different variations and tests, this is quite a strange issue.

[benelot]

Ok, so finally I have some time to test this through with you @moonflowAeon. Sorry for the delay in answering. I will prepare a run pipeline jupyter notebook/script and include the bach dataset into the repository for some simple training (~1.4 Mb, so it should not inflate/clutter the repo).

[benelot]

I included some more commits I did a while ago and tested all stages with the bach dataset. I included the bach dataset now into the data/raw folder. Please try the whole training again with that dataset and tell me how this goes with your setup. I also updated the requirements.txt to represent the dependencies with which it works for me here. If you want to test the composer stage only. I included a link to a pretrained model in the updated Readme.md in the repo. Here is the excerpt from it:

Pretrained models

Below are some pretrained models for different datasets. To use the model, download the zip from the link below and extract it into the results/history folder of your project. The zip should contain a folder named e and some number, indicating the number of epochs the model was trained and some model.h5. Pay attention when extracting the model not to override one of your own trained, most beloved models!

• Bach dataset: https://drive.google.com/open?id=1P_hOF0v55m1Snzvri5OBVP5yy099Lzgx

[ahfriedman]

I'm not sure where to post this but I couldn't find a better place so I figured I'd ask here.

In midi_utils.py, I noticed midi_to_samples has an option for encode_length. If I read the comments correctly, it sounds like this should be a way of making it such that the notes are played for longer than one tick. However, when enabling it nothing seems to change. I guessed this was because of the end for the note already being set. But when I tried changing that, I broke it. Are notes with length not yet supported?

[benelot]

Good place to ask if you ask me. Are you using the code released in this PR? If yes, I can help you out on making this work. Can you train a network without note lengths? I did not actively work with the note length, but we can look into it.

On Fri, Jul 26, 2019, 17:28 AHFriedman [email protected] wrote:

I'm not sure where to post this but I couldn't find a better place so I figured I'd ask here.

In midi_utils.py, I noticed midi_to_samples has an option for encode_length. If I read the comments correctly, it sounds like this should be a way of making it such that the notes are played for longer than one tick. However, when enabling it nothing seems to change. I guessed this was because of the end for the note already being set. But when I tried changing that, I broke it. Are notes with length not yet supported?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/HackerPoet/Composer/pull/21?email_source=notifications&email_token=AAXXXK6RXMGJTPJBLFJ4NIDQBMJZNA5CNFSM4HLTBOD2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD24544I#issuecomment-515497585, or mute the thread https://github.com/notifications/unsubscribe-auth/AAXXXK6VC4CV2IHK757IIQDQBMJZNANCNFSM4HLTBODQ .

[ahfriedman]

I am using the same code and am able to get the network to train. When I forced encode_length to be True and re-ran preprocess_songs and then train.py, gt.mid still only had the notes without duration. When I looked at midi_utils.py I wondered if the notes were being cut short because of:

for note in all_notes:
        for start_end in all_notes[note]:
            if len(start_end) == 1:
                start_end.append(start_end[0] + 1)

(starting on line 86 in midi_utils.py) and when it goes to encode the length on line 107, it sees the notes all ending one tick after they start. The program only threw an error when I removed this section to see if the part of the code that encodes the length would make up for it.

[benelot]

Cool! Thanks for testing the code. I will see if I can make it work and give you some updates! The original code never uses that, but it could be fun to run on many more midis to learn the length as well.

P.S. Just to make this PR ready for merging: Anything special you had to run to make the network train? Any specific dataset you train on that works well?

On Sat, Jul 27, 2019, 17:07 AHFriedman [email protected] wrote:

I am using the same code and am able to get the network to train. When I forced encode_length to be True and re-ran preprocess_songs and then train.py, gt.mid still only had the notes without duration. When I looked at midi_utils.py I wondered if the notes were being cut short because of: for note in all_notes: for start_end in all_notes[note]: if len(start_end) == 1: start_end.append(start_end[0] + 1) (starting on line 86 in midi_utils.py) and when it goes to encode the length on line 107, it sees the notes all ending one tick after they start. The program only threw an error when I removed this section to see if the part of the code that encodes the length would make up for it.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/HackerPoet/Composer/pull/21?email_source=notifications&email_token=AAXXXK4MJJ3IYOX3KOBUCGDQBRQE5A5CNFSM4HLTBOD2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD26NEKA#issuecomment-515691048, or mute the thread https://github.com/notifications/unsubscribe-auth/AAXXXK6LHGKIEXZSOFFLACTQBRQE5ANCNFSM4HLTBODQ .

[ahfriedman]

Thanks! Even if it just ends up putting a note every tick between the start and end of the actual note and then has to convert back to notes with length, I think the results could be cool. The only change I had to make to the code to get to run was installing pyaudio by doing:

pip install pipwin
pipwin install pyaudio

But I guess I kind of asked for issues because I didn’t switch over to Linux while running this. In requirements.txt pyaduio is also listed twice. I didn’t try this on any specific dataset, just some midi files of songs I found (so I also increased the generated song length).

Thank you for putting time into this. I couldn’t get the original version to work on the GPU (and just adding pip install tensorflow-gpu worked for this). Its been really interesting to mess with. Just for fun and in part because I just took some songs to train it with, I’ve been working on a java program that attempts to split the bass/rhythm/drums or whatever into a second instrument. I did use a comically small sample size for training, so at times it uses small parts of songs. But I deserve and was asking for that by using a comically small sample size. And it still does make interesting results with that.

[ghost]

Sorry for the delay @benelot! The updated migration you posted has been stable from the training tests that I’ve run on it.

Here are five models that I recorded data on:

Model 1 parameters (4.5 hours processing time) • 2000 epochs • Bach repository

Model 2 • 100 epochs • Bach repository

Model 3 • 100 epochs • Random MIDI repository (168 songs)

Model 4 • 100 epochs • Random MIDI repository (936 songs)

---

Model 5 (14 hours processing time) • 4000 epochs • Bach repository

[benelot]

So it looks like it works!

[benelot]

@HackerPoet It looks like we are ready to merge.

[HackerPoet]

@AHFriedman The note length feature was something I was originally planning to do, but later decided against it to make the training more uniform, but the code was left there still in an unfinished state. The problem was that some of the MIDIs in my database contained zero-length notes, some had small fixed durations, and some had dynamic lengths. I was worried the network would get confused by this with my limited dataset and it would have complicated the synth and midi code, so I just scraped it.

@benelot @moonflowAeon @AHFriedman It's great that you all have it working now! I'll merge this PR after a quick review. I'd love to see any samples or models you produce from other datasets, I'm really curious how it compares to the video game one that I used.