-
Notifications
You must be signed in to change notification settings - Fork 2
/
default.conf
136 lines (92 loc) · 3.04 KB
/
default.conf
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
# Default configurations that are used as a fallback
#
# Do NOT delete or modify this file unless you know what you are doing
[Default]
# Path to the parent directory containing the program output
work_path = _workspace
# Path to the directory containing the dataset
dataset_path = _dataset
# Path to the directory containing extracted feature vectors
extraction_path = %(work_path)s/features
# Path to the directory containing saved training models
model_path = %(work_path)s/models
# Path to the directory containing log files
log_path = %(work_path)s/logs
# Path to the directory containing predictions
prediction_path = %(work_path)s/predictions
# Path to the directory containing results
result_path = %(work_path)s/predictions
[Extraction]
# Whether to recompute feature vectors if they already exist
recompute = False
[Extraction.Logmel]
# Target sample rate
sample_rate = 32000
# Length of the FFT window
n_fft = 1024
# Number of audio samples between frames
hop_length = 512
# Number of Mel bands
n_mels = 64
[Training]
# String identifying a particular training instance
training_id = default
# Neural network architecture
#
# Choices: vgg, densenet
model = vgg
# Mask for selecting a subset of the training set
#
# Format: key1=value1,key2=value2,...
# Example: mask = manually_verified=0,noisy_small!=1
#
# The keys should correspond to the column names of the metadata file.
# Currently, only integer values are accepted.
mask =
# Random seed used prior to training
seed = 1000
# Size of a block after partitioning the feature vectors
#
# If a feature vector has size TxF, it is paritioned into blocks of size
# BxF, where B is the block size. The BxF blocks are then used as inputs
# to train the neural network.
block_size = 128
# Number of examples in a mini-batch
batch_size = 128
# Number of epochs to train the network
n_epochs = 40
# Initial learning rate
lr = 0.0005
# Factor for learning rate decay
lr_decay = 0.90
# Frequency of learning rate decay in epochs
lr_decay_rate = 2
# Whether to use data augmentation
augment = False
# Whether to overwrite any previously-saved models
# Setting this to False means that training can be resumed
overwrite = False
[Training.Relabel]
# Whether relabeling should be enabled
relabel = False
# Threshold used to select out-of-distribution instances
relabel_threshold = 0.55
# Weight used to relabel out-of-distribution instances
relabel_weight = 0.5
# Path to CSV file containing pseudo-labels
pseudolabel_path = metadata/training_pseudo.csv
# Path to CSV file containing confidence estimates
# Leave blank to use pseudo-labels for confidence estimation
confidence_path =
[Prediction]
# Specification of which models (epochs) to select for prediction
#
# Either a list of epoch numbers (e.g. '1,2,3') or a string with format
# 'metric:n' specifying which metric to use to select the top n epochs.
#
# Valid metrics: val_loss, val_acc, val_mAP
epochs = val_mAP:3
# Whether to use ODIN when generating predictions
odin = False
# Whether to remove the model files that were not used for prediction
clean = False