review

[Lallapallooza]

No description provided.

[Lallapallooza]

зачем эти приседания с установкой куднн руками? можно взять готовый образ с куднн, например
nvcr.io/nvidia/tensorrt:22.08-py3

можно поискать более подходящий

[Lallapallooza]

$(pwd)

[Lallapallooza]

нужны комменты к каждому полю, вот пример как гпт4 сгенерил, поправьте чтоб получше было, т.е. оставь группировку

DATA OPTIONS

MODEL OPTIONS

TRAIN OPTIONS

AUGMENTATION OPTIONS

ну и убери избыточность в комментах

там где будут идти строчки в стиле

# Section
param1 : VeryLongType1 = "value1" # doc1
long_param_name : Type1 = "value1" # doc2

форматнуть как

# Section
param1          : VeryLongType1 = "value1" # doc1
long_param_name : Type1         = "value1" # doc2

@dataclass
class TrainConfig:
    # The computing platform for training: 'cuda' for NVIDIA GPUs or 'cpu' for CPU-based training.
    device: str = "cuda"

    # Directory path where the MUSDB18-HQ dataset is stored or to be downloaded.
    musdb_path: str = "musdb18hq"

    # Directory path for saving training metadata, like track names and lengths.
    metadata_train_path: str = "metadata"

    # Directory path for saving testing metadata.
    metadata_test_path: str = "metadata1"

    # Length (in seconds) of each audio segment used during training. Shorter segments can help in learning finer details.
    segment: int = 5

    # Batch size for training, determining how many samples are processed together. Affects memory usage and gradient updates.
    batch_size: int = 6

    # Whether to shuffle the training dataset at the beginning of each epoch. Shuffling can help in reducing overfitting.
    shuffle_train: bool = True

    # Whether to shuffle the validation dataset. Generally not needed as validation performance is independent of data order.
    shuffle_valid: bool = False

    # Whether to drop the last incomplete batch in case the dataset size isn't divisible by the batch size.
    drop_last: bool = True

    # Number of worker processes used for loading data. Higher numbers can speed up data loading.
    num_workers: int = 2

    # Strategy for monitoring metrics to save model checkpoints: 'min' for saving when a monitored metric is minimized.
    metric_monitor_mode: str = "min"

    # Number of best-performing model weights to save based on the monitored metric.
    save_top_k_model_weights: int = 1

    # PM_Unet model configurations
    # Specific sources to target in source separation, e.g., separating drums, bass, etc.
    model_source: tuple = ("drums", "bass", "other", "vocals")

    # The depth of the U-Net architecture, affecting its capacity and complexity.
    model_depth: int = 4

    # Number of initial channels in U-Net layers, influencing the model's learning capability.
    model_channel: int = 28

    # Indicates whether the input audio should be treated as mono (True) or stereo (False).
    is_mono: bool = False

    # Whether to utilize masking within the model for source separation.
    mask_mode: bool = False

    # Mode of skip connections in U-Net ('concat' for concatenation, 'add' for summation).
    skip_mode: str = "concat"

    # Number of bins used in Fast Fourier Transform (FFT) during audio processing.
    nfft: int = 4096

    # Determines whether to use LSTM layers as bottleneck in the U-Net architecture.
    bottlneck_lstm: bool = True

    # Number of LSTM layers if bottleneck is enabled, adding to the model's complexity and ability to capture temporal dependencies.
    layers: int = 2

    # Flag to decide whether to apply Short Time Fourier Transform (STFT) on input audio.
    stft_flag: bool = True

    # Augmentation parameters
    # Maximum number of samples to shift in time during data augmentation, helping the model generalize over temporal variations.
    shift: int = 8192

    # Probability of applying pitch shift augmentation to the audio, introducing pitch variability in training data.
    pitchshift_proba: float = 0.2

    # Range of semitone shifts for pitch shift augmentation applied specifically to vocal tracks.
    vocals_min_semitones: int = -5
    vocals_max_semitones: int = 5

    # Range of semitone shifts for pitch shift augmentation applied to non-vocal tracks.
    other_min_semitones: int = -2
    other_max_semitones: int = 2

    # Flag to enable pitch shift augmentation on non-vocal sources.
    pitchshift_flag_other: bool = False

    # Probability of applying time stretching or compression to alter the speed of the audio without changing its pitch.
    time_change_proba: float = 0.2

    # Factors for time stretching/compression, defining the range and intensity of this augmentation.
    time_change_factors: tuple = (0.8, 0.85, 0.9, 0.95, 1.05, 1.1, 1.15, 1.2, 1.25, 1.3)

    # Probability of remixing audio tracks, useful for training the model to recognize and separate blended sounds.
    remix_proba: float = 1

    # Number of tracks to group together for the remix augmentation.
    remix_group_size: int = batch_size

    # Probability and range of scaling the amplitude of audio tracks, introducing dynamic range variability.
    scale_proba: float = 1
    scale_min: float = 0.25
    scale_max: float = 1.25

    # Probability of applying a fade effect to the audio, simulating natural starts and ends of sounds.
    fade_mask_proba: float = 0.1

    # Probability of doubling one channel's audio to both channels, simulating mono audio in a stereo setup.
    double_proba: float = 0.1

    # Probability of reversing a segment of the audio track, useful for capturing reversed audio patterns.
    reverse_proba: float = 0.2

    # Probability and depth of mixing tracks together to create mashups, aiding the model in learning from complex blends of sounds.
    mushap_proba: float = 0.0
    mushap_depth: int = 2

    # Loss function parameters
    # Multiplicative factors for different components of the loss function, affecting the emphasis on different aspects of the training objective.
    factor: int = 1
    c_factor: int = 1

    # Number of FFT bins for calculating loss, impacting the granularity of frequency-based loss computation.
    loss_nfft: tuple = (4096,)

    # Gamma parameter for adjusting the focus of the loss function on certain aspects of the audio spectrum.
    gamma: float = 0.3

    # Learning rate for the optimizer, a crucial parameter for training convergence.
    lr: float = 0.5 * 3e-3

    # Period of the cosine annealing schedule in learning rate adjustment, influencing the adaptation of learning rate over time.
    T_0: int = 40

    # Maximum number of training epochs, defining the total duration of the training process.
    max_epochs: int = 100

    # Precision of training computations, e.g., '16' for 16-bit floating-point precision, reducing memory usage and potentially speeding up training.
    precision: str = 16

    # Gradient clipping value to prevent exploding gradients, aiding in training stability.
    grad_clip: float = 0.5

[Lallapallooza]

поправить на

"""
A dataset class for audio source separation, compatible with WAV (or MP3) files.
This class allows training with arbitrary sources, where each audio track is represented by a separate folder within the specified root directory. Each folder should contain audio files for different sources, named as `{source}.{ext}`.

Args:
    root (Path or str): The root directory of the dataset where audio tracks are stored.
    metadata (dict): Metadata information generated by the `build_metadata` function. It contains details like track names and lengths.
    sources (list[str]): A list of source names to be separated, e.g., ['drums', 'vocals'].
    segment (Optional[float]): The length of each audio segment in seconds. If `None`, the entire track is used.
    shift (Optional[float]): The stride in seconds between samples. Determines the overlap between consecutive audio segments.
    normalize (bool): If True, normalizes the input audio based on the entire track's statistics, not just individual segments.
    samplerate (int): The target sample rate. Audio files with a different sample rate will be resampled to this rate.
    channels (int): The target number of audio channels. If different, audio will be converted accordingly.
    ext (str): The file extension of the audio files, default is '.wav'.

Note:
    The `samplerate` and `channels` parameters are used to ensure consistency across the dataset. They allow on-the-fly conversion of audio properties to match the target specifications.
"""

[Lallapallooza]

плохо хардкодить имена весов

[Lallapallooza]

хардкод имен весов плохо, передавай параметром

[Lallapallooza]

хардкод плохо, передавай параметром

[Lallapallooza]

хардкод плохо, передавай параметром

[d-a-yakovlev]

model_filename = f"{args.class_name}_outer_stft_{config.segment_duration:.1f}"
model_path = args.out_dir + '/' + model_filename

ok?

[Lallapallooza]

хардкод плохо, передавай параметром

[Lallapallooza]

название на нормальное, tf_lite_stream.py