signal-dataset-demo.md

signal-dataset-demo

This is mainly for understanding what is happening under the hood. If you do not need to know these details, or do not care, then skip to advanced-example in the README for more usage.

Create your metadata dataframe. Test signals are in /data of this repo so you can validate easily.

# NOTE: Make sure you are executing from /tf_dataset top-level dir to find these relpaths!
import os
import pandas as pd
import tensorflow as tf
from tf_dataset import *
d = {
'filepath': [
    'data/1561965276.252_1561965582.106_24346_passenger.wav',
    'data/1561970414.69_1561970702.051_24346_passenger.wav',
    'data/1561971147.625_1561971202.077_24346_cargo.wav',
    'data/1561971697.731_1561971718.283_24346_tug.wav'],
'class': ['passenger', 'passenger', 'cargo', 'tug']
}
df = pd.DataFrame.from_dict(d)
df['filepath'] = df['filepath'].apply(lambda x: os.path.abspath(x))
print(df.head())
>>>                                             filepath      class
>>> 0  C:\internal\sam\tf_dataset\data\1561965276.252...  passenger
>>> 1  C:\internal\sam\tf_dataset\data\1561970414.69_...  passenger
>>> 2  C:\internal\sam\tf_dataset\data\1561971147.625...      cargo
>>> 3  C:\internal\sam\tf_dataset\data\1561971697.731...        tug

write_signal_dataset_precursor_database() is where the sqlite database is created. Resulting path will link to the time and process_id stamped sqlite database, ending in .db

path = write_signal_dataset_precursor_database(df, overwrite=True)
print(path)
>>> C:\internal\tf_dataset\signal_dataset_precursor_1444_10_19_20_15-35-28_.sqlite

Grab the SQL query we want to execute. This query will grab all relevant information in the metadata sql table, which has been stored in the sqlite database at path.

deterministic = False
query = "SELECT * FROM `metadata` ORDER BY " +\
    ("`_ROWID_`" if deterministic else "RANDOM()")

Create a precursor dataset (tf.data.Dataset object) containing a minimum of [filepaths, labels] entries. No signal data is actually loaded at this point, only the filepaths, labels, and potentially other metadata are housed in the tf.data.Dataset.

print("\nCreating signal dataset precursor...\n".upper())
print("`tf.data.Dataset` now contains all metadata from padas `DataFrame`")
ds = signal_dataset_precursor(path, query)
for x in ds.take(1):
    print(x)
>>> {'class': <tf.Tensor: shape=(), dtype=string, numpy=b'tug'>,
>>>  'filepath': <tf.Tensor: shape=(), dtype=string,
>>>     numpy=b'C:\\internal\\tf_dataset\\data\\1561971697.731_1561971718.283_24346_tug.wav'>,
>>>  'index': <tf.Tensor: shape=(), dtype=int64, numpy=3>,
>>>  'num_classes': <tf.Tensor: shape=(), dtype=int64, numpy=4>,
>>>  'target': <tf.Tensor: shape=(), dtype=int64, numpy=2>}

The default behavior in process_db_entry() is to load signals as int32 1D tensors using tf.io. You may, however, pass your own process_db_entry() callable that processes the data to your liking. (See user-defined-maps)

# Example code for `process_db_entry()`.
# This function is mapped over every element in the dataset
keep_filepath = True
def process_db_entry(x):
    sig = tf.io.read_file(x['filepath'])
    sig = tf.io.decode_raw(sig, tf.int32)
    x['signal'] = sig
    if not keep_filepath:
        x.pop('filepath')
    return x

Now, to return a usable dataset with loaded signal data (lazily evaluated), we must map over the precursor dataset using process_db_entry() defined above.

num_parallel_calls = 4
ds = ds.map(process_db_entry,
        num_parallel_calls=tf.convert_to_tensor(num_parallel_calls, tf.int64))
for x in ds.take(1):
    print(x)
>>> {'class': <tf.Tensor: shape=(), dtype=string, numpy=b'cargo'>,
>>>  'filepath': <tf.Tensor: shape=(), dtype=string,
>>>     numpy=b'C:\\internal\\tf_dataset\\data\\1561971147.625_1561971202.077_24346_cargo.wav'>,
>>>  'index': <tf.Tensor: shape=(), dtype=int64, numpy=2>,
>>>  'num_classes': <tf.Tensor: shape=(), dtype=int64, numpy=4>,
>>>  'signal': <tf.Tensor: shape=(2613707,), dtype=int32,
>>>       numpy=array([1179011410, 1163280727, ..., 1526755493, 7365376])>,
>>>  'target': <tf.Tensor: shape=(), dtype=int64, numpy=1>}

NOTE: If your data are not homogeneously shaped, then it is strongly advised to pass the resultant dataset from signal_dataset() to dataset_signal_slice_windows() in order to have uniform shapes across batches.

ds = dataset_signal_slice_windows(ds, win_len=32768)