search

andabi / music-source-separation

Deep neural networks for separating singing voice from music written in TensorFlow
795 stars 150 forks source link

Help setting ModelConfig! #24

Closed lydo closed 5 years ago

lydo commented 5 years ago

Hi Andabi, I have downloaded 289,205 sample wav files from https://magenta.tensorflow.org/datasets/nsynth. Each sample file is 4s. Running to train.py fails to the to_spectrogram (possibly the len_frame and len_hop arguments are wrong)). The configuration I keep by default

class ModelConfig:      SR = 16000 # Sample Rate      L_FRAME = 1024 # default 1024      L_HOP = closest_power_of_two (L_FRAME / 4)      SEQ_LEN = 4      # For Melspectogram      N_MELS = 512      F_MIN = 0.0

Can you guide me how to set the parameters correctly? Thank you so much!