diff --git a/README.md b/README.md index b9b9746..73f60fc 100644 --- a/README.md +++ b/README.md @@ -41,6 +41,7 @@ This will help me alot to keep this project alive and improve it. * Type `.venv\Scripts\activate` and press enter. * You should see now a `(.venv)` in front of your console line. * Install the requirements with `pip install -r requirements.txt`. +* Install gpu requirements `pip3 install torch==2.0.1+cu117 torchvision==0.15.2+cu117 torchaudio==2.0.2+cu117 --index-url https://download.pytorch.org/whl/cu117` * Now you can use the UltraSinger source code with `py UltraSinger.py [opt] [mode] [transcription] [pitcher] [extra]`. See [How to use](#how-to-use) for more information. For more information about Python environments look [here](https://code.visualstudio.com/docs/python/environments#_global-virtual-or-conda-environments). @@ -51,6 +52,7 @@ As copy: py -m venv .venv .venv\Scripts\activate pip install -r requirements.txt +pip3 install torch==2.0.1+cu117 torchvision==0.15.2+cu117 torchaudio==2.0.2+cu117 --index-url https://download.pytorch.org/whl/cu117 ``` @@ -192,3 +194,38 @@ for low male voices. Accurate is the real tone specified in the txt. I had txt f singable by humans, but you could still reach the 10k points in the game. The accuracy is important here, because from this MIDI and sheet are created. And you also want to have accurate files + +### Use GPU + +With an GPU you can speed up the process and also the quality of the transcription and pitching is better. + +You need a cuda device for this to work. +For best use you need to install the cuda driver for your gpu see [here](https://developer.nvidia.com/cuda-downloads). +Install torch with cuda separately in your `venv`. See this [link](https://pytorch.org/get-started/locally/). + +Command for `pip`: +``` +pip3 install torch~=2.0.1+cu117 torchvision~=0.15.2+cu117 torchaudio~=2.0.2+cu117 --index-url https://download.pytorch.org/whl/cu117 +``` + +When you want to use `conda` instead you need a different installation command. See this [link](https://pytorch.org/get-started/locally/). + +#### Info + +If somthing crash because of low VRAM than use a smaller model. +Whisper needs more than 8GB VRAM in the `large` model! + +But you can force cpu usage with the extra options `--force_whisper_cpu` and `--force_separation_cpu`. + +### Use GPU + +With an GPU you can speed up the process and also the quality of the transcription and pitching is better. +You need a cuda device for this to work. +If you use MAC than sorry, there is no cuda device for MAC machinges. + +#### Info + +If somthing crash because of low VRAM than use a smaller model. +Whisper needs more than 8GB VRAM in the `large` model! + +But you can force cpu usage with the extra options `--force_whisper_cpu` and `--force_separation_cpu`. diff --git a/Settings.py b/Settings.py index 7557ca4..09c271a 100644 --- a/Settings.py +++ b/Settings.py @@ -26,4 +26,7 @@ class Settings: crepe_step_size = 10 # Device - device = "cpu" # cpu|cuda|gpu + device = 'cpu' # cpu|cuda + force_whisper_cpu = False + force_separation_cpu = False + diff --git a/UltraSinger.py b/UltraSinger.py index f4ac493..5f4ece7 100644 --- a/UltraSinger.py +++ b/UltraSinger.py @@ -437,9 +437,9 @@ def get_unused_song_output_dir(path): def transcribe_audio(transcribed_data): """Docstring""" if settings.transcriber == "whisper": + device = "cpu" if settings.force_whisper_cpu else settings.device transcribed_data, language = transcribe_with_whisper( - settings.mono_audio_path, settings.whisper_model, settings.device - ) + settings.mono_audio_path, settings.whisper_model, device) else: # vosk transcribed_data = transcribe_with_vosk( settings.mono_audio_path, settings.vosk_model_path @@ -450,14 +450,15 @@ def transcribe_audio(transcribed_data): def separate_vocal_from_audio( - basename_without_ext, cache_path, ultrastar_audio_input_path + basename_without_ext, cache_path, ultrastar_audio_input_path ): """Docstring""" audio_separation_path = os.path.join( cache_path, "separated", "htdemucs", basename_without_ext ) + device = "cpu" if settings.force_separation_cpu else settings.device if settings.use_separated_vocal or settings.create_karaoke: - separate_audio(ultrastar_audio_input_path, cache_path) + separate_audio(ultrastar_audio_input_path, cache_path, device) if settings.use_separated_vocal: vocals_path = os.path.join(audio_separation_path, "vocals.wav") convert_audio_to_mono_wav(vocals_path, settings.mono_audio_path) @@ -761,6 +762,8 @@ def init_settings(argv): "disable_separation=", "disable_karaoke=", "create_audio_chunks=", + "force_whisper_cpu=", + "force_separation_cpu=" ] opts, args = getopt.getopt(argv, short, long) if len(opts) == 0: @@ -790,6 +793,10 @@ def init_settings(argv): settings.create_karaoke = not arg elif opt in ("--create_audio_chunks"): settings.create_audio_chunks = arg + elif opt in ("--force_whisper_cpu"): + settings.force_whisper_cpu = arg + elif opt in ("--force_separation_cpu"): + settings.force_separation_cpu = arg if settings.output_file_path == "": if settings.input_file_path.startswith("https:"): dirname = os.getcwd() diff --git a/modules/Audio/separation.py b/modules/Audio/separation.py index 4797ad1..d34241e 100644 --- a/modules/Audio/separation.py +++ b/modules/Audio/separation.py @@ -19,9 +19,9 @@ def separate_audio(input_file_path: str, output_file: str, device="cpu") -> None # Model selection? # -n mdx_q # -n htdemucs_ft - # todo: -d cpu otherwise it automatically uses gpu + device = "cpu" if device != "cuda" else "cuda" subprocess.run( - ["demucs", "--two-stems=vocals", input_file_path], check=True + ["demucs", "-d", device, "--two-stems=vocals", input_file_path] ) separated_folder = path_join(current_executor_path(), "separated") - move(separated_folder, output_file) + move(separated_folder, output_file) \ No newline at end of file diff --git a/modules/DeviceDetection/device_detection.py b/modules/DeviceDetection/device_detection.py index 6ba0cec..a4b97e2 100644 --- a/modules/DeviceDetection/device_detection.py +++ b/modules/DeviceDetection/device_detection.py @@ -1,26 +1,21 @@ """Docstring""" -import tensorflow as tf +import torch from modules.console_colors import ULTRASINGER_HEAD, red_highlighted def get_available_device(): """Docstring""" - gpus = tf.config.list_physical_devices("GPU") + isCuda = torch.cuda.is_available() - if not gpus: + if not isCuda: print( - f"{ULTRASINGER_HEAD} There are no GPUs available. Using {red_highlighted('cpu')}." + f"{ULTRASINGER_HEAD} There are no {red_highlighted('cuda')} devices available. -> Using {red_highlighted('cpu')}." ) + return "cpu" + else: + print(f"{ULTRASINGER_HEAD} Using {red_highlighted('cuda')} GPU.") - print(f"{ULTRASINGER_HEAD} Found available GPUs:") - for gpu in gpus: - print(f"Name: {gpu.name}, Type: {gpu.device_type}") - - # Todo: Finish this - print( - f"{ULTRASINGER_HEAD} GPU usage are currently in development. Using {red_highlighted('cpu')}" - ) - return "cpu" + return "cuda" \ No newline at end of file diff --git a/requirements.txt b/requirements.txt index f6f4116..50083d0 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,11 +1,13 @@ crepe==0.0.13 librosa==0.9.2 -pretty_midi==0.2.9 +pretty_midi~=0.2.10 pydub==0.25.1 -Levenshtein==0.20.9 -scipy==1.10.0 +Levenshtein~=0.21.0 +scipy~=1.10.1 vosk==0.3.45 tensorflow~=2.12.0 +#tensorflow-directml-plugin~=0.4.0.dev230202 +#cudatoolkit==11.8.0 matplotlib~=3.7.1 yt_dlp~=2023.3.4 numpy~=1.23.5