config-reference.md

Configuration Reference

Sections

• AudioConfig
• TrainingConfig
• InferenceConfig
• MiscConfig
• BaseConfig

AudioConfig

Field	Type	Default	Description
spec_duration	float	5.0	Spectrogram duration in seconds
spec_height	int	128	Spectrogram height in pixels
spec_width	int	480	Spectrogram width in pixels (must be divisible by 32)
win_length	float	0.055	Window length is specified in seconds, to retain temporal and frequency resolution when max_freq and sampling rate are changed
n_fft	Union[int, NoneType]	None	If none, set n_fft = win_length_samples
max_freq	int	8000	Maximum frequency in Hz
min_freq	int	100	Minimum frequency in Hz
sampling_rate	int	18000	A little more than 2 * max_freq
freq_scale	str	'mel'	"linear", "log" or "mel"
power	float	1.0	Use 1.0 for magnitude and 2.0 for power spectrograms
decibels	bool	False	Use decibel amplitude scale?
top_db	float	80	Threshold below max amplitude in dB; lower values are clipped
db_power	float	1.0	Raise to this exponent after convert to decibels
log_freq_gain	float	0.6	Boost loudness of higher frequencies with log scale
mel_norm	Union[str, NoneType]	None	Mel filterbank normalization: None or "slaney"
choose_channel	bool	False	Use heuristic to pick the cleanest audio channel?
check_seconds	float	6.0	Check this many seconds to pick channel
use_spec_cache	bool	False	When use_spec_cache=False, each spectrogram is generated separately. When use_spec_cache=True, a single spectrogram is generated for the recording, by concatenating chunks. Then that big spectrogram is divided as needed, which saves time when there is a lot of overlap.
chunks_per_spec	int	3	chunks_per_spec defines the size of the individual spectrograms that are concatenated to create the cache, where the duration is chunks_per_spec * spec_duration.

TrainingConfig

Field	Type	Default	Description
model_type	str	'effnet.2'	Use timm.x for timm model "x"
head_type	Union[str, NoneType]	None	If None, use backbone's default
hidden_channels	int	256	Hidden channels in classifier head
lse_temp	float	0.5	LSE temperature for temporal_sed head
two_way	bool	True	Bi-directional or unidirectional temporal_sed head
pretrained	bool	False	Use pretrained weights (applies to timm models)
load_ckpt_path	Union[str, NoneType]	None	For transfer learning or fine-tuning
freeze_backbone	bool	False	Option when transfer learning
multi_label	bool	True	True for multi-label, False for multi-class classification
deterministic	bool	False	Enable deterministic training for reproducibility
seed	Union[int, NoneType]	None	Random seed for reproducibility; None uses random seed
learning_rate	float	0.001	Base learning rate
batch_size	int	64	Mini-batch size
shuffle	bool	True	Shuffle data during training?
num_epochs	int	10	Number of epochs
warmup_fraction	float	0.0	Learning rate warmup fraction
save_last_n	int	3	Save checkpoints for this many last epochs
num_folds	int	1	For k-fold cross-validation
val_portion	float	0	Used only if num_folds = 1
train_db	str	'data/training.db'	Path to training database
train_pickle	str	'data/training.pkl'	Path to training pickle file
test_pickle	Union[str, NoneType]	None	Path to test pickle file
frame_label_pickle	Union[str, NoneType]	None	Path to frame-label pickle used in SED training
num_workers	int	3	Number of trainer worker threads
compile	bool	False	Compile the model?
mixed_precision	bool	False	Use mixed precision?
use_class_weights	bool	False	Should loss function weight classes by spec count?
weight_exponent	float	0.5	Exponent to soften the class weights
pos_label_smoothing	float	0.08	Positive side of asymmetric label smoothing
neg_label_smoothing	float	0.01	Negative side of asymmetric label smoothing
optimizer	str	'radam'	Any timm optimizer
opt_weight_decay	Union[float, NoneType]	None	Weight decay option (L2 regularization)
opt_beta1	Union[float, NoneType]	None	Adam/RAdam beta1 (exponential decay rate for first moment)
opt_beta2	Union[float, NoneType]	None	Adam/RAdam beta2 (exponential decay rate for second moment)
drop_rate	Union[float, NoneType]	None	Standard dropout
drop_path_rate	Union[float, NoneType]	None	Stochastic depth dropout
sed_fps	int	4	Frames per second from SED heads
frame_loss_weight	float	0.5	Segment_loss_weight = 1 - frame_loss_weight
offpeak_weight	float	0	Weight for penalizing predictions outside peak regions
max_per_recording	Union[int, NoneType]	None	Per-recording sampling: if set, randomly select this many specs per recording per epoch
val_max_per_recording	Union[int, NoneType]	None	Per-recording limit for validation: if set, take the first N specs per recording
augment	bool	True	Use data augmentation?
max_augmentations	int	1	Up to this many per spectrogram
noise_class_name	str	'Noise'	Augmentation treats noise specially
prob_simple_merge	float	0.32	Prob of simple merge
prob_mixup	float	0.0	Prob of traditional mixup (mutually exclusive with simple merge)
prob_cutmix	float	0.0	Prob of CutMix (mutually exclusive with simple merge and mixup)
mixup_alpha	float	0.4	Beta distribution parameter for mixup/cutmix lambda
prob_fade1	float	0.5	Prob of fading after augmentation
min_fade1	float	0.1	Min factor for fading
max_fade1	float	1.0	Max factor for fading
augmentations	list	<factory >	Detailed augmentation settings

InferenceConfig

Field	Type	Default	Description
segment_len	Union[float, NoneType]	None	For models with SED heads, if segment_len is None, output tags of variable lengths that match the sounds detected, otherwise output tags of length segment_len seconds. For non-SED models, segment_len is defined by the model.
overlap	float	0.0	Number of seconds overlap for adjacent spectrograms
min_score	float	0.8	Only generate labels when score is at least this
num_threads	int	3	More threads = faster but more VRAM
max_models	Union[int, NoneType]	None	If specified, limit ensemble size accordingly
autocast	bool	True	Faster and less VRAM but less precision
audio_power	float	0.7	Audio power parameter during inference
scaling_coefficient	float	1.0	Platt scaling coefficient, to align predictions with probabilities
scaling_intercept	float	0.0	Platt scaling intercept, to align predictions with probabilities
label_field	str	'codes'	"names", "codes", "alt_names" or "alt_codes"
block_size	int	200	Do this many spectrograms at a time to avoid running out of GPU memory
openvino_block_size	int	100	Block size when OpenVINO is used (do not change after creating onnx files)
initial_offsets	Union[list, NoneType]	None	If specified, analyze command calls get_overlapping_scores instead of get_recording_scores and passes this array.

MiscConfig

Field	Type	Default	Description
force_cpu	bool	False	If true, use CPU (for performance comparisons)
ckpt_folder	str	'data/ckpt'	Use an ensemble of all checkpoints in this folder for inference
search_ckpt_path	Union[str, NoneType]	None	Folder with one or more checkpoints for embeddings and search
exclude_list	str	'data/exclude.txt'	Classes listed in this file are excluded from inference output
source_regexes	Union[list, NoneType]	<factory >	Sample regexes to map recording names to source names
map_codes	Union[dict, NoneType]	None	Dict mapping old to new class codes for checkpoint compatibility

BaseConfig

Field	Type	Default	Description
audio	AudioConfig
train	TrainingConfig
infer	InferenceConfig
misc	MiscConfig