Way to generate more melodies sequences

[valientegaston]

Thanks for the previous answer. I have tried and it effectively lengthens the duration of the theme, but repeating the same thing over and over again. Is there a way to generate more sequences of different melodies to have a more extensive musical track but not for that reason repetitive?

[jacbz]

I'm afraid that's not possible without retraining the model. You can adjust the parameters before training.

https://github.com/jacbz/Lofi/blob/a379facc79972ed573cb317640abbb0e600f5586/model/constants.py#L34

[valientegaston]

Thanks for the reply. I am trying to retrain the model. I have changed the parameter that you have indicated from 50 to 200. If with 50 I obtain 1:30 minutes, I expect to obtain 4 times more with 200. I have followed the instructions in the model/dataset readme file and I have been able to obtain the "processed" folder with 15 files, but it says "Result: 0 samples", is that correct? Or am I doing something wrong? Thank you so much

gaston@iMac dataset % python3 preprocessor.py
preprocessor.py:70: SyntaxWarning: "is" with a literal. Did you mean "=="?
  if len(json_data["tracks"]["melody"]) is 0 or all(
preprocessor.py:75: SyntaxWarning: "is" with a literal. Did you mean "=="?
  if len(json_data["tracks"]["chord"]) is 0 or all(
Result: 0 samples
gaston@iMac dataset % cd .. 
gaston@iMac model % ls
README.md		embeddings.py		lyrics2lofi_model.py
__init__.py		lofi2lofi_dataset.py	lyrics2lofi_train.py
constants.py		lofi2lofi_model.py	train.py
dataset			lofi2lofi_train.py
dataset.py		lyrics2lofi_dataset.py
gaston@iMac model % python3 lofi2lofi_train.py 
Traceback (most recent call last):
  File "lofi2lofi_train.py", line 3, in <module>
    from model.lofi2lofi_dataset import Lofi2LofiDataset
ModuleNotFoundError: No module named 'model'
gaston@iMac model % 

These are the generated files

gaston@iMac model % cd dataset gaston@iMac dataset % ls README.md preprocessor.py spotify_client_id hooktheory processed spotify_client_secret gaston@iMac dataset % cd processed gaston@iMac processed % ls X Ambassadors - Unsteady - intro.json Xi - Freedom Dive - intro.json Xi - Freedom Dive - verse.json Xilent - Animation - chorus-lead-out.json Xilent - Animation - chorus.json Xilent - Choose Me - chorus.json Xilent - Choose Me - intro-and-verse.json Xilent - Disconnect - instrumental.json Xilent - Synthony - chorus.json Xilent - The Place - chorus.json Xilent - The Place - instrumental.json Xxanaxx - Give U The World - chorus.json Xxanaxx - Got U Under Spisek Jednego Remix - verse-and-pre-chorus.json Xxanaxx - Story - intro-and-verse.json Xxanaxx - Story - pre-chorus-and-chorus.json gaston@iMac processed %

[jacbz]

The dataset should contain more than 10,000 files. I'm not quite sure what went wrong on your machine.

[valientegaston]

Where should those 15 thousand files come from? From Spotify?

[valientegaston]

Should the hooktheory folder look like this?

[jacbz]

It seems you downloaded the repo directly, which only contains an excerpt of the whole dataset. The repo links to a bigger 4.9 GB dataset in its readme.

[valientegaston]

Thank you very much for your patience. I already have more than 17 thousand files inside the processed folder. After that I tried to run lofi2lofi_train and it threw an error that it couldn't find the folder 'dataset/processed-spotify-all' I renamed the processed folder to that folder I was looking for and started with the following...

%sudo python3 lofi2lofi_train.py Using cpu device Starting training: lofi2lofi == Epoch 0 == Scheduled sampling rate: C 0, M 0 Batch 0: Loss 5.503 (C: 2.198 + KL: 0.016 + M: 2.704 + T: 0.069 + K: 0.084 + Mo: 0.200 + V: 0.123 + E: 0.108) Batch 1: Loss 5.445 (C: 2.186 + KL: 0.017 + M: 2.683 + T: 0.065 + K: 0.082 + Mo: 0.194 + V: 0.112 + E: 0.107) Batch 2: Loss 5.386 (C: 2.172 + KL: 0.027 + M: 2.668 + T: 0.057 + K: 0.083 + Mo: 0.196 + V: 0.092 + E: 0.091) Batch 3: Loss 5.325 (C: 2.152 + KL: 0.040 + M: 2.651 + T: 0.050 + K: 0.081 + Mo: 0.191 + V: 0.082 + E: 0.077) Batch 4: Loss 5.315 (C: 2.147 + KL: 0.054 + M: 2.647 + T: 0.054 + K: 0.081 + Mo: 0.183 + V: 0.074 + E: 0.074) Batch 5: Loss 5.275 (C: 2.152 + KL: 0.058 + M: 2.626 + T: 0.049 + K: 0.082 + Mo: 0.181 + V: 0.069 + E: 0.059) Batch 6: Loss 5.192 (C: 2.128 + KL: 0.055 + M: 2.584 + T: 0.044 + K: 0.084 + Mo: 0.177 + V: 0.058 + E: 0.061) Batch 7: Loss 5.133 (C: 2.091 + KL: 0.053 + M: 2.579 + T: 0.047 + K: 0.082 + Mo: 0.173 + V: 0.056 + E: 0.052) Batch 8: Loss 5.084 (C: 2.075 + KL: 0.050 + M: 2.551 + T: 0.047 + K: 0.083 + Mo: 0.168 + V: 0.058 + E: 0.052) Batch 9: Loss 5.087 (C: 2.062 + KL: 0.049 + M: 2.570 + T: 0.045 + K: 0.082 + Mo: 0.166 + V: 0.059 + E: 0.055) Batch 10: Loss 5.045 (C: 2.065 + KL: 0.046 + M: 2.539 + T: 0.044 + K: 0.083 + Mo: 0.163 + V: 0.057 + E: 0.049) Batch 11: Loss 5.003 (C: 2.032 + KL: 0.043 + M: 2.528 + T: 0.046 + K: 0.083 + Mo: 0.160 + V: 0.060 + E: 0.052) Batch 12: Loss 5.017 (C: 2.080 + KL: 0.040 + M: 2.503 + T: 0.044 + K: 0.083 + Mo: 0.158 + V: 0.061 + E: 0.049) Batch 13: Loss 4.968 (C: 2.029 + KL: 0.037 + M: 2.507 + T: 0.044 + K: 0.082 + Mo: 0.158 + V: 0.061 + E: 0.050) Batch 14: Loss 5.052 (C: 2.092 + KL: 0.035 + M: 2.535 + T: 0.046 + K: 0.082 + Mo: 0.158 + V: 0.056 + E: 0.049) Batch 15: Loss 4.964 (C: 2.045 + KL: 0.037 + M: 2.505 + T: 0.042 + K: 0.082 + Mo: 0.149 + V: 0.053 + E: 0.052) Batch 16: Loss 4.989 (C: 2.031 + KL: 0.033 + M: 2.529 + T: 0.044 + K: 0.081 + Mo: 0.153 + V: 0.063 + E: 0.054) Batch 17: Loss 5.016 (C: 2.070 + KL: 0.034 + M: 2.539 + T: 0.043 + K: 0.083 + Mo: 0.144 + V: 0.051 + E: 0.051)

Is this what this is expected to do? What files should that output generate? Is it okay to rename the processed folder to processed-spotify-all? How many batch processes should this run? About 17 thousand? Thank you very much for helping me move forward.

[valientegaston]

I've changed MAX_CHORD_LENGTH = 400 in constants.py, I've retrained the model, but I'm still getting songs of 1:30 min.... do I need to change something else for a longer duration? I have tried changing MELODY_PREDICTION_LENGTH, to *4 , but it gives compile error. Of the files generated after the retraining, I only copy the lofi2lofi_decoder.pth to the checkpoints folder, is that correct? Can anybody help me? Thank you so much.

[valientegaston]

I've finally been able to retrain the AI. I've made changes to MAX_CHORD_LENGTH and MELODY_PREDICTION_LENGTH hoping to get longer songs but to no avail. In the early stages of training, epoch0 does generate songs up to 20 minutes long, but then as the AI ​​training progresses, it falls to around 1:30 minutes every time it is generated. Perhaps it is due to the input dataset, where the AI ​​learns that the songs should last that minute and a half, and for that reason it does not generate longer songs, and simply the 20-minute ones obtained at the beginning is bad learning or ignorance of the AI . I don't know what else I can change, so I leave the topic here until I see if someone can achieve it, or if it is directly not possible with this model. Thank you so much.

[ManHinnn0509]

Hey man, thank you so much for the detailed follow up with your training progress. I just got everything to work but I don't understand when I should stop training the model? At around how many epoch? A figure were shown after each epoch but I don't really understand it.

Looking for help on this, thanks!

[tesh254]

@valientegaston edit lofi2lofi_train.py line 8 to dataset_folder = "dataset/processed"

[AryanNanda17]

Heyy How did you lengthen the music... although it is repeating still!