Skip to content

Latest commit

 

History

History
101 lines (80 loc) · 2.63 KB

FORMAT.md

File metadata and controls

101 lines (80 loc) · 2.63 KB
Raw

Format

Dataset folder structure

Dataset filder is prepared for each use cases. The folder has the following structure:

 - /dataset_folder --+-- /data/*.csv
                     |
                     +-- /dataset.yaml
                     |
                     +-- /class_labels.yaml
                     |
                     +-- /*.h5

Operation

  • Initial setup phase: You have to edit "dataset.yaml" manually
  • Data acquisition phase: oscilloscope.py saves csv files in /data
  • Training phase: Jupyter notebook generates "class_labels.yaml" and "*.h5"
  • Inference phase: add the h5 file in "dataset.yaml" as a model

Dataset format

Windows

A list of windows is automatically generated by using the following info:


         <--------------------- samples ----------------------->
         +-----------------------------------------------------+
         |                                                     |
         |                                                     |
       ^ +-----------------+                                   |
Cutoff | |                 |                                   |
       V |                 |                                   |
         +-----------------------------------------------------+
         <---- length ----->
         - stride->

One record in a dataset


         <------------------ samples (200) -------------------->
         +----------------+----------------+----------------+--+
         |                |                |                |  | ^
         |                |                |                |  | |
       ^ |                |                |                |  | |
Cutoff | |                |                |                |  | | Filters(64 or 40)
(13)   V |                |                |                |  | V
         +----------------+----------------+----------------+--+
         <---- length ---->
             (64 or 96)
         - stride->
         (32, 64, 48 or 96)

class_labels.yaml

Each index position in the list corresponds to each class identifier.

Example:

[blues_harp, classical_guitar, framenco_guitar, piano, silence, tin_whistle]

dataset.yaml

"dataset.yaml" is a config file with all the parameters for each use cases.

Example for MFSCs:

feature: mfsc
files: 100
training_files: 70
samples: 200
length: 64
filters: 40
cutoff: null
stride: 12
model: null

Example for MFCCs:

feature: mfcc
files: 100
training_files: 70
samples: 200
length: 64
filters: 40
cutoff: 13
stride: 12
model: null

"cutoff" is an upper bound of MFCC coefficients. For example, "13" means coefficients from 1st to 12th.