AdamW.amsgrad: false
AdamW.betas: !!python/tuple
- 0.9
- 0.999
AdamW.capturable: false
AdamW.differentiable: false
AdamW.eps: 1.0e-08
AdamW.lr: 0.0001
AdamW.maximize: false
AdamW.weight_decay: 0.01

LinearWarmupCosineDecay.eta_min: 1.0e-05
LinearWarmupCosineDecay.last_iter: -1
LinearWarmupCosineDecay.total_iters: 100000
LinearWarmupCosineDecay.warmup_iters: 10000

PrefixDecoderTransformerMultiOut.attn_dropout: 0.1
PrefixDecoderTransformerMultiOut.cross_attend: false
PrefixDecoderTransformerMultiOut.depth: 16
PrefixDecoderTransformerMultiOut.dim: 1024
PrefixDecoderTransformerMultiOut.ff_dropout: 0.1
PrefixDecoderTransformerMultiOut.heads: 16
PrefixDecoderTransformerMultiOut.input_delay: 0
PrefixDecoderTransformerMultiOut.input_emb_dim: 128
PrefixDecoderTransformerMultiOut.max_seq_len: 512
PrefixDecoderTransformerMultiOut.num_rvq_layers: 4
PrefixDecoderTransformerMultiOut.num_tokens: 1024
PrefixDecoderTransformerMultiOut.output_emb_dim: 1024
PrefixDecoderTransformerMultiOut.pad_value: 0
PrefixDecoderTransformerMultiOut.shared: false
PrefixDecoderTransformerMultiOut.use_delay_pattern: false

accelerator: auto

add_inst_tokens: false

args.debug: 0
args.load: configs/multiout/prefix_decoder_base_cluster.yml
args.save: null
args.unknown: []

batch_size: 32

cfg_drop_prob: 0.0

checkpoint_interval: 5000

checkpoint_metric: val/loss

checkpoint_mode: min

checkpoint_top_k: -1

compile: true

data_base_dir: /network/scratch/w/wu.yusong/precompute_audio_mixdown_10s

dataset_names:
- slakh2100

devices: auto

duration: 10

gradient_clip_val: 1.0

inst_tokens_as_pattern_token: true

limit_val_batches: 128

load_audio: 'false'

load_from_latest_checkpoint: true

log_every_n_steps: 1

max_duration: 10

max_log_examples: 8

num_nodes: 1

num_rvq_layers: 4

num_workers: 8

overfit_batches: 0

pattern: flatten

precision: bf16-mixed

rms_base_dir: /network/scratch/w/wu.yusong/rms_50hz

sample_interval: 10000

save_dir: /network/scratch/w/wu.yusong/online_stem_gen_logs/prefix_decoder_base

seed: 42

shuffle: false

split: train

strategy: auto

train_steps: 100000

use_fabric: false

val_interval: 1000

wandb_project: online-stem-gen

weights:
- 1