Valid bug fix

configs/datamodule/dummy.yaml

@@ -0,0 +1,36 @@
+_target_: src.datamodules.musdb18hq_datamodule.Musdb18hqDataModule
+
+data_dir: ${data_dir} # data_dir is specified in config.yaml
+target_name: # <- No default! Mandatory!
+n_fft: 4096
+hop_length: 128
+dim_c: 4
+dim_f: 2048
+dim_t: 256
+sampling_rate: 44100
+
+batch_size: 8
+num_workers: 0
+pin_memory: False
+
+external_datasets:
+augmentation: True
+
+validation_set:
+  - Actions - One Minute Smile
+  - Clara Berry And Wooldog - Waltz For My Victims
+  - Johnny Lokke - Promises & Lies
+  - Patrick Talbot - A Reason To Leave
+  - Triviul - Angelsaint
+  - Alexander Ross - Goodbye Bolero
+  - Fergessen - Nos Palpitants
+  - Leaf - Summerghost
+  - Skelpolu - Human Mistakes
+  - Young Griffo - Pennies
+  - ANiMAL - Rockshow
+  - James May - On The Line
+  - Meaxic - Take A Step
+  - Traffic Experiment - Sirens
+
+
+gpus: 1

configs/datamodule/musdb18_hq.yaml

@@ -7,6 +7,7 @@ hop_length: ${model.hop_length}
 dim_c: ${model.dim_c}
 dim_f: ${model.dim_f}
 dim_t: ${model.dim_t}
+gpus: ${trainer.gpus}
 sampling_rate: 44100
 
 batch_size: 8

configs/eval_config/debug.yaml

@@ -0,0 +1,55 @@
+model:
+  # model
+  _target_: src.models.mdxnet.ConvTDFNet
+  target_name: ${target_name}
+
+  # model configuration
+  num_blocks: 9
+  l: 5
+  g: 32
+  k: 3
+  bn: 8
+  bias: False
+
+  # optimizer
+  lr: 0.0001
+  optimizer: rmsprop
+
+  # stft
+  n_fft: 6144
+  dim_f: 256
+  dim_t: 128
+  dim_c: 4
+  hop_length: 1024
+
+##########################
+
+sources:
+  - 'vocals'
+
+
+vocals:
+  # model configuration
+  _target_: src.models.mdxnet.ConvTDFNet
+  target_name: vocals
+
+  # model configuration
+  num_blocks: 5
+  l: 5
+  g: 32
+  k: 3
+  bn: 8
+  bias: False
+
+  # optimizer
+  lr: 0.0001
+  optimizer: rmsprop
+
+  # stft
+  n_fft: 6144
+  dim_f: 2048
+  dim_t: 128
+  dim_c: 4
+  hop_length: 1024
+
+gpus: 1

configs/eval_config/default.yaml

@@ -0,0 +1,52 @@
+model:
+  # model
+  _target_: src.models.mdxnet.ConvTDFNet
+  target_name: ${target_name}
+
+  # model configuration
+  num_blocks: 9
+  l: 5
+  g: 32
+  k: 3
+  bn: 8
+  bias: False
+
+  # optimizer
+  lr: 0.0001
+  optimizer: rmsprop
+
+  # stft
+  n_fft: 6144
+  dim_f: 256
+  dim_t: 128
+  dim_c: 4
+  hop_length: 1024
+
+##########################
+
+sources:
+  - 'vocals'
+
+vocals:
+  # model
+  _target_: src.models.mdxnet.ConvTDFNet
+  target_name: vocals
+
+  # model configuration
+  num_blocks: 9
+  l: 5
+  g: 32
+  k: 3
+  bn: 8
+  bias: False
+
+  # optimizer
+  lr: 0.0001
+  optimizer: rmsprop
+
+  # stft
+  n_fft: 6144
+  dim_f: 2048
+  dim_t: 128
+  dim_c: 4
+  hop_length: 1024

configs/eval_config/mskim.yaml

@@ -0,0 +1,52 @@
+model:
+  # model
+  _target_: src.models.mdxnet.ConvTDFNet
+  target_name: ${target_name}
+
+  # model configuration
+  num_blocks: 9
+  l: 5
+  g: 32
+  k: 3
+  bn: 8
+  bias: False
+
+  # optimizer
+  lr: 0.0001
+  optimizer: rmsprop
+
+  # stft
+  n_fft: 6144
+  dim_f: 256
+  dim_t: 128
+  dim_c: 4
+  hop_length: 1024
+
+##########################
+
+sources:
+  - 'vocals'
+
+vocals:
+  # model
+  _target_: src.models.mdxnet.ConvTDFNet
+  target_name: vocals
+
+  # model configuration
+  num_blocks: 9
+  l: 3
+  g: 32
+  k: 3
+  bn: 8
+  bias: False
+
+  # optimizer
+  lr: 0.0001
+  optimizer: rmsprop
+
+  # stft
+  n_fft: 6144
+  dim_f: 2048
+  dim_t: 128
+  dim_c: 4
+  hop_length: 1024

configs/evaluation.yaml

@@ -0,0 +1,29 @@
+# @package _global_
+
+# specify here default training configuration
+defaults:
+  - datamodule: dummy.yaml
+  - eval_config: default.yaml
+  - logger: wandb.yaml
+  # enable color logging
+  - override hydra/hydra_logging: colorlog
+  - override hydra/job_logging: colorlog
+
+# dummy
+target_name: vocals
+work_dir: ${hydra:runtime.cwd}
+# dummy
+
+# default
+data_dir: ${oc.env:data_dir}
+onnx_dir: ${oc.env:onnx_dir}
+
+split: 'valid'
+
+wandb_api_key: ${oc.env:wandb_api_key}
+
+batch_size: 8
+
+logger:
+  wandb:
+    project: mdx_eval_${split}

configs/experiment/debug.yaml

@@ -0,0 +1,87 @@
+# @package _global_
+
+# to execute this experiment run:
+# python run.py experiment=example_simple.yaml
+
+defaults:
+  - override /trainer: minimal.yaml
+  - override /callbacks: none.yaml
+  - override /logger: none.yaml
+
+# Training Environment
+target_name: vocals
+batch_size: 4
+gpus: 1
+lr: 0.0003
+check_val_every_n_epoch: 1
+seed: 2021
+deterministic: True
+num_workers: 2 # num of cpus
+pin_memory: False # True then faster
+
+# TODO: check effects
+sync_batchnorm: False
+resume_from_checkpoint:
+
+# STFT
+n_fft: 6144
+dim_f: 2048
+dim_t: 128
+dim_c: 4
+hop_length: 1024
+
+# data augmentation
+external_datasets:
+augmentation: False
+
+model:
+  # model
+  _target_: src.models.mdxnet.ConvTDFNet
+  target_name: ${target_name}
+
+  # model configuration
+  num_blocks: 3
+  l: 5
+  g: 8
+  k: 3
+  bn: 8
+  bias: False
+
+  # optimizer
+  lr: ${lr}
+  optimizer: rmsprop
+
+  # stft
+  n_fft: ${n_fft}
+  dim_f: ${dim_f}
+  dim_t: ${dim_t}
+  dim_c: ${dim_c}
+  hop_length: ${hop_length}
+
+trainer:
+  max_epochs: 3
+  min_epochs: 1
+  sync_batchnorm: ${sync_batchnorm}
+  precision: 16
+  resume_from_checkpoint: ${resume_from_checkpoint}
+  auto_lr_find: False
+  amp_backend: "native"
+  amp_level: "O2"
+
+  deterministic: ${deterministic}
+  gpus: 1
+  check_val_every_n_epoch: ${check_val_every_n_epoch}
+
+# auto script
+datamodule:
+  num_workers: ${num_workers} # num of cpus
+  pin_memory: ${pin_memory} # True then faster
+  external_datasets: ${external_datasets}
+  augmentation: ${augmentation}
+  batch_size: ${batch_size}
+  target_name: ${target_name}
+  n_fft: ${n_fft}
+  hop_length: ${hop_length}
+  dim_c: ${dim_c}
+  dim_f: ${dim_f}
+  dim_t: ${dim_t}

configs/knou/final_other.yaml

@@ -11,7 +11,7 @@ target_name: other
 batch_size: 3
 gpus: 4
 sync_batchnorm: False
-lr: 0.00005
+lr: 0.00015855604910877148
 check_val_every_n_epoch: 10
 seed: 2021
 

configs/knou/final_other_sweep.yaml

@@ -2,7 +2,8 @@
 
 # to execute this experiment run:
 # python run.py experiment=example_simple.yaml
-
+# python run.py -m hparams_search=knou_12 knou=final_other_sweep
+# {'model.lr': 0.00015855604910877148}
 defaults:
   - final_other.yaml
   - override /logger: tensorboard.yaml

configs/knou/final_vocals_debug.yaml

@@ -0,0 +1,50 @@
+# @package _global_
+
+# to execute this experiment run:
+# python run.py experiment=example_simple.yaml
+
+defaults:
+  - multigpu_default.yaml
+
+# Training Environment
+target_name: vocals
+batch_size: 3
+gpus: 4
+sync_batchnorm: False
+lr: 0.00005
+check_val_every_n_epoch: 1
+seed: 2021
+
+# STFT
+n_fft: 6144
+dim_f: 2048
+dim_t: 128
+dim_c: 4
+hop_length: 1024
+
+# data augmentation
+external_datasets:
+augmentation: False
+
+model:
+  # model configuration
+  # model configuration
+  _target_: src.models.mdxnet.ConvTDFNet
+  target_name: vocals
+
+  # model configuration
+  num_blocks: 5
+  l: 5
+  g: 32
+  k: 3
+  bn: 8
+  bias: False
+
+  # optimizer
+  lr: 0.0001
+  optimizer: rmsprop
+
+
+trainer:
+  max_epochs: 5
+  check_val_every_n_epoch: 1