New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Rave 2.3.1 doesn't work on my Ubuntu #273
Comments
You should roll back rave to a version that worked for you if you need to use it urgently !Le 20 déc. 2023 à 15:32, Federico8691 ***@***.***> a écrit :
Now I am not even able to preprocess the audiofiles.
These is my user path
image.png (view on web)
image.png (view on web)
Sorry for pressing you over all of this, but I am working with this tool and now I am stuck.
Many thanks,
Federico
—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>
|
Hi Antoine, Federico |
On the website it states: The original implementation of the RAVE model can be restored using git checkout v1 What odes it means? thanks Federico |
you can get back to a previous version using pip install acids-rave==2.1.1
Le mer. 20 déc. 2023 à 16:01, Federico8691 ***@***.***> a
écrit :
… On the website it states:
The original implementation of the RAVE model can be restored using
git checkout v1
What odes it means?
thanks
Federico
—
Reply to this email directly, view it on GitHub
<#273 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AH2A3QGQ4IDXF25MV7ETBWDYKL4VBAVCNFSM6AAAAABA3Q3DESVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNRUGYZDIMBYGY>
.
You are receiving this because you commented.Message ID:
***@***.***>
--
Antoine Caillon
0606927077
|
I do not understand. Everything seems ok. It was working until this morning. |
What I do not get is the (no such file or directory), but it is there, in front of my eyes. Maybe it is a problem with the installation of Rave inside miniconda3. A directory mismatch? |
Is there any way to get my hands on v1 on my linux machine? It was working so well. My best models have been done with v1. |
Any help, suggestion? |
You can follow the instructions at
https://github.com/acids-ircam/RAVE/tree/v1
Le mer. 20 déc. 2023 à 16:48, Federico8691 ***@***.***> a
écrit :
… Any help, suggestion?
—
Reply to this email directly, view it on GitHub
<#273 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AH2A3QEKAVENAE7LUW5CLALYKMCDPAVCNFSM6AAAAABA3Q3DESVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNRUG4YTMNJTGE>
.
You are receiving this because you commented.Message ID:
***@***.***>
--
Antoine Caillon
0606927077
|
Start by cloning RAVE v1 (git clone https://github.com/acids-ircam/RAVE.git
--branch v1)
Then follow the instructions at https://github.com/acids-ircam/RAVE/tree/v1
|
Using this in the terminal? git clone https://github.com/acids-ircam/RAVE and then .. You can now use python cli_helper.py to start a new training ! thanks! |
I am running this on Google Colab with the command below and using acids-rave==2.1.1: !rave train --config "v2" --db_path '/content/dataset' --name "testRave" --val_every 2500 It worked for me after I addressed an issue with the second --config flag, which is for the regularization method. By removing this flag, the training started normally. |
Hi, |
Obviously you have a problem with your data paths, no influence of RAVE version here. Double check the paths of your input folder (dragging / dropping into the console to retrieve that path maybe). Just a precision : since 2.3.1 RAVE is not constraining torch to be 1.13, in order to be compatible with new devices. I just advised you to update your libraries :) |
Axel wrote:
This is exactly what I do as you can see from this image: I have the miniconda3 folder (where Rave is then installed) and two additional folder; the Blippo_dataset contains the audio files and the other the preprocessed one called myDataset, plus the runs folder. rave preprocess --input_path 'Blippo_dataset..' --output_path 'myDataset' Then when I run it I get this:
I followed your suggestion as you wrote me. Maybe it is a problem where Rave is installed? Thanks for your kind help |
Federico, Please refer to the FAQ for the answers and let me know. PS : Regarding python environments, I think you wouldn't loose time (especially if you use it on a daily basis) reading basic python environment handling with Internet resources (https://realpython.com/python-virtual-environments-a-primer/). Environements can be painful even for advanced users, so I would strongly advise you to get in touch with familiar commands if you really need it. |
Hi Axel, thanks for your kind answer. |
There was actually a bug in 2.1 that prevented random cropping during training, pertaining overfitting problems. This problem was fixed. |
Getting the same error message (working on colab). Trying with v2.3.0 (installed with
|
Are you resuming a checkpoint? I cannot reproduce this bug from a fresh training |
also getting this same error when using 2.3.1 on fresh training install + preprocess worked fine then training caused I1221 20:56:26.757378 140430877622912 resource_reader.py:50] system_path_file_exists:v3.gin |
Can you please give
I tried with v2 and v3, and on GitHub Actions these configurations pass the test. There must be something wrong in your config or in your databaae |
Hi domkirke My system is an RTX 3090 24GB/AMD EPYC 7551P/32GB ram/Torch = 2.1.2/Torchaudio = 2.1.2 Steps from a completely clean system install
You mentioned "if a folder is created, the .gin inside summarizing the architecture of the training". I'm not sure where this folder would be located. If you're referring to the dataset folder from step 4, the resulting folder did not have a .gin file inside it. If you need anymore information please let me know. Thanks in advance. |
Ok, got that ; it is a problem with the default amount of audio channels. Meanwhile, Please add --channels X (X being your number of channels: 1 for mono, 2 for stereo, etc). Will add that to the README.md, and fix that in the next version. Does it work? Thanks! |
Unfortunately still getting an error when adding this flag. It's a different error this time though. Maybe I need to remake the dataset with this channel flag? rave train --config v3 --db_path dataset/ --out_path model/ --name brute --channels 2 --val_every 2500 Traceback (most recent call last): |
There is a log error I also corrected, but it seems that yes you preprocessed your dataset as a mono dataset and asked for a stereo. Also add --channels 2 to your preprocess step, and --channels 2 in the train step |
Got it. Not sure if this is a my system issue but for some reason when I try and preprocess the dataset with channels 2 flag I get a threading error thats seemingly freezes the process. I tried again using the original command (which I assume defaults to mono) and it worked though for some reason. Is there a flag to limit the number of threads or something like that? Appreciate your help with this. rave preprocess --input_path audio/ --output_path dataset/ --channels 2
BUT rave preprocess --input_path audio/ --output_path dataset/
|
Works fine in my configuration, both mac and linux. What gives |
ulimit -u Hmm. Not sure why this would be the case. Think I may try a different machine. Have been using a cloud machine for training. Edit 2: Redid steps on a new machine. Same issue !rave preprocess --input_path audio/ --output_path dataset/ --channels 2 0it [00:00, ?it/s]Exception in thread Thread-5 (accepter): |
Yes, I was. Still getting the same error. However, I if Istart a fresh training with v2.3.0 using the same dataset with these configurations I can get it running (resuming works, too).
Howver when trying to export
Moreover, using v2.3.1 with these configuartions seems to work fine, too, for training and resuming training.
But when trying to export the model I get this error which I don't know the reason for. Did I maybe not training long enough?
|
@Federico8691 Can you add |
@chrizzlemadizzle v3.2.0 had trouble with wasserstein and spherical configs, leading to global fixes and v3.2.1. So, if I summarize well : you are not able to export a Wasserstein model trained with v3.2.1? Could you make another issue, as it is not the topic of the current one? Thanks :-) |
Wanted to make a quick update to this. Using a clean install on the exact same machines I rolled back to v2.3 and recreated the exact steps (ie doing --channels 2 and the process described previously) and RAVE worked flawlessly and is now training. Leads me to believe something may have happened in 2.3.1 with the threading? EDIT: In attempting to export the model from this version/setup I'm getting a "size mismatch" error rave export --run model/brute_caeb149cc6 --streaming --channels 2 --output modelfinal/ root@C.7946089:/workspace$ rave export --run model/brute_caeb149cc6 --streaming --channels 2 --output modelfinal/ |
Out of topic Axel, |
Hi to everyone still nothing new. I made all the steps Axel requested. I am still getting same problem. I created a new directory, and I put an old batch of audiofiles I used for some training a while ago without problems. here is the outcome: the funny thing is that in the --output_path directory Rave creates these two files: I tried both with 2.1.1 and with 2.3.1 I have no idea about what is happening here. Keep in mind that I wanted to present some of my research at March Forum 2024. (I will be in Paris). So I am stuck. Looking forward to your answer (I am working even if it is Christmas :-D) |
I made this check myself, running ulimit -u I got 256513 What does it means? |
Dear Axel, I was able to solve the preprocess, I just updated ffmpeg encoder. |
There was a missing developer library called libsox-dev. |
Dear everyone,
I am sorry for bothering everyone again but I really do not know where to hit my head in order to solve these kinds of questions being no computer scientist.
Since I updated Rave to its latest version (2.3.1) it doesn't work anymore.
Here is a screenshot of the kind of errors I get when I try to run the training over an already preprocessed dataset.
Any possible help would be welcomed and thanks for your time.
best,
Federico
The text was updated successfully, but these errors were encountered: