mirror of
https://github.com/storytold/storyteller-ml.git
synced 2026-10-17 00:23:14 +00:00
Brandon Thomas
184 lines
3.9 KiB
YAML
184 lines
3.9 KiB
YAML
name: &name "QuartzNet15x5"
|
|
sample_rate: &sample_rate 16000
|
|
repeat: &repeat 1
|
|
dropout: &dropout 0.0
|
|
separable: &separable true
|
|
labels: &labels [" ", "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m",
|
|
"n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "'"]
|
|
|
|
model:
|
|
train_ds:
|
|
manifest_filepath: ???
|
|
sample_rate: 16000
|
|
labels: *labels
|
|
batch_size: 32
|
|
trim_silence: True
|
|
max_duration: 16.7
|
|
shuffle: True
|
|
is_tarred: False
|
|
tarred_audio_filepaths: null
|
|
tarred_shard_strategy: "scatter"
|
|
|
|
validation_ds:
|
|
manifest_filepath: ???
|
|
sample_rate: 16000
|
|
labels: *labels
|
|
batch_size: 32
|
|
shuffle: False
|
|
|
|
preprocessor:
|
|
_target_: nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor
|
|
normalize: "per_feature"
|
|
window_size: 0.02
|
|
sample_rate: *sample_rate
|
|
window_stride: 0.01
|
|
window: "hann"
|
|
features: &n_mels 64
|
|
n_fft: 512
|
|
frame_splicing: 1
|
|
dither: 0.00001
|
|
stft_conv: false
|
|
|
|
spec_augment:
|
|
_target_: nemo.collections.asr.modules.SpectrogramAugmentation
|
|
rect_freq: 50
|
|
rect_masks: 5
|
|
rect_time: 120
|
|
|
|
encoder:
|
|
_target_: nemo.collections.asr.modules.ConvASREncoder
|
|
feat_in: *n_mels
|
|
activation: relu
|
|
conv_mask: true
|
|
|
|
jasper:
|
|
- filters: 128
|
|
repeat: 1
|
|
kernel: [11]
|
|
stride: [1]
|
|
dilation: [1]
|
|
dropout: *dropout
|
|
residual: true
|
|
separable: *separable
|
|
se: true
|
|
se_context_size: -1
|
|
|
|
- filters: 256
|
|
repeat: *repeat
|
|
kernel: [13]
|
|
stride: [1]
|
|
dilation: [1]
|
|
dropout: *dropout
|
|
residual: true
|
|
separable: *separable
|
|
se: true
|
|
se_context_size: -1
|
|
|
|
- filters: 256
|
|
repeat: *repeat
|
|
kernel: [15]
|
|
stride: [1]
|
|
dilation: [1]
|
|
dropout: *dropout
|
|
residual: true
|
|
separable: *separable
|
|
se: true
|
|
se_context_size: -1
|
|
|
|
- filters: 256
|
|
repeat: *repeat
|
|
kernel: [17]
|
|
stride: [1]
|
|
dilation: [1]
|
|
dropout: *dropout
|
|
residual: true
|
|
separable: *separable
|
|
se: true
|
|
se_context_size: -1
|
|
|
|
- filters: 256
|
|
repeat: *repeat
|
|
kernel: [19]
|
|
stride: [1]
|
|
dilation: [1]
|
|
dropout: *dropout
|
|
residual: true
|
|
separable: *separable
|
|
se: true
|
|
se_context_size: -1
|
|
|
|
- filters: 256
|
|
repeat: 1
|
|
kernel: [21]
|
|
stride: [1]
|
|
dilation: [1]
|
|
dropout: 0.0
|
|
residual: false
|
|
separable: *separable
|
|
se: true
|
|
se_context_size: -1
|
|
|
|
- filters: &enc_feat_out 1024
|
|
repeat: 1
|
|
kernel: [1]
|
|
stride: [1]
|
|
dilation: [1]
|
|
dropout: 0.0
|
|
residual: false
|
|
separable: *separable
|
|
se: true
|
|
se_context_size: -1
|
|
|
|
decoder:
|
|
_target_: nemo.collections.asr.modules.ConvASRDecoder
|
|
feat_in: 1024
|
|
num_classes: 28
|
|
vocabulary: *labels
|
|
|
|
optim:
|
|
name: novograd
|
|
# _target_: nemo.core.optim.optimizers.Novograd
|
|
lr: .01
|
|
# optimizer arguments
|
|
betas: [0.8, 0.5]
|
|
weight_decay: 0.001
|
|
|
|
# scheduler setup
|
|
sched:
|
|
name: CosineAnnealing
|
|
|
|
# pytorch lightning args
|
|
monitor: val_loss
|
|
reduce_on_plateau: false
|
|
|
|
# Scheduler params
|
|
warmup_steps: null
|
|
warmup_ratio: null
|
|
min_lr: 0.0
|
|
last_epoch: -1
|
|
|
|
trainer:
|
|
gpus: 0 # number of gpus
|
|
max_epochs: 5
|
|
max_steps: null # computed at runtime if not set
|
|
num_nodes: 1
|
|
accelerator: ddp
|
|
accumulate_grad_batches: 1
|
|
checkpoint_callback: False # Provided by exp_manager
|
|
logger: False # Provided by exp_manager
|
|
log_every_n_steps: 1 # Interval of logging.
|
|
val_check_interval: 1.0 # check once per epoch .25 for 4 times per epoch
|
|
|
|
exp_manager:
|
|
exp_dir: null
|
|
name: *name
|
|
create_tensorboard_logger: True
|
|
create_checkpoint_callback: True
|
|
|
|
hydra:
|
|
run:
|
|
dir: .
|
|
job_logging:
|
|
root:
|
|
handlers: null
|