-
Notifications
You must be signed in to change notification settings - Fork 1
/
data_processing.py
165 lines (123 loc) · 4.87 KB
/
data_processing.py
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
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
import os
import sys
import time
import copy
import gc
import numpy as np
import random
import pandas as pd
import librosa
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
import torch
torch.backends.cudnn.benchmark=True
torch.manual_seed(0)
import torch.utils.data as utils
sys.path.append('audio_tagging_functions')
from models import *
from create_birds_dataset import FINAL_LABELS_PATH
from mp3towav import SAMPLE_RATE
DEVICE = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
### Data processing functions ###
def load_df2array(df, sample_rate, audio_duration):
"""
Loads and stores arrays, and also returns a dict of missing files.
"""
nonValidDict = {}
waveforms_list = []
ConstantShape = sample_rate * audio_duration # Zero-padding
for idx in range(len(df)):
filename = df['fname'][idx]
length = df['length'][idx]
label = df['label'][idx]
if os.path.isfile(filename):
waveform = load_waveform2numpy(filename, length, sample_rate=sample_rate, audio_duration=audio_duration)
length_waveform = len(waveform)
# Zero-padding
if length_waveform != ConstantShape:
waveform = np.pad(waveform, (0, ConstantShape - length_waveform), 'constant')
waveforms_list.append([waveform, label])
else:
iD = df['iD'][idx]
nonValidDict[iD] = filename
del df
gc.collect()
return (waveforms_list, nonValidDict)
def load_waveform2numpy(filename, length, sample_rate, audio_duration):
"""
Loads from filename path and returns numpy array.
"""
# Random crop of a 10 sec segment
offset = random.randint(0, length-audio_duration)
waveform, _ = librosa.core.load(filename, sr=sample_rate, mono=True, offset=offset, duration=audio_duration)
return waveform
def get_dataloaders(x, y, batch_size):
"""
Converts numpy arrays to Pyrtoch dataloader.
"""
tensor_x = torch.from_numpy(x).float().to('cpu')
tensor_y = torch.from_numpy(y).long().to(DEVICE)
tensordataset = utils.TensorDataset(tensor_x, tensor_y)
dataloader_length = len(tensordataset)
dataloader = utils.DataLoader(tensordataset, batch_size=batch_size, shuffle=True)
# Free some memory spaces
del x, y, tensor_x, tensor_y, tensordataset
gc.collect()
return (dataloader, dataloader_length)
def plot_distribution(do_plot, y_list):
"""
Plot distribution for a dataset.
"""
if do_plot:
nb_y = len(y_list)
labels = ["trainset", "validationset", "testset"]
fig, axes = plt.subplots(1,nb_y,figsize=(8*nb_y,4))
axes = axes.ravel()
for idx, ax in enumerate(axes):
y_array = y_list[idx]
label = labels[idx]
len_y = len(y_array)
ax.hist(y_array, bins=np.arange(min(y_array), max(y_array) + 0.5, 0.5), align='left')
ax.set_title(f"Distribution of the samples for {label} (n={len_y})")
ax.set_xlabel('label')
ax.set_ylabel('Number of samples')
plt.tight_layout()
plt.show()
def process_data(df, batch_size, sample_rate, audio_duration, random_state, do_plot=False):
"""
Process data function, returns the dataloader of the dataframe df.
"""
start_time = time.time()
waveforms_list, nonValidDict = load_df2array(df, sample_rate, audio_duration)
print(f"Valid files: {len(waveforms_list)}\nUnvalid files: {len(nonValidDict)}")
print("Vstacking data...")
waveforms_arrays = [x[0] for x in waveforms_list]
X_all = np.vstack(waveforms_arrays)
y_all = np.array([x[1] for x in waveforms_list])
del waveforms_list, waveforms_arrays
gc.collect()
print("Attributing arrays to dataloader...")
dataloader, dataloader_length = get_dataloaders(X_all, y_all, batch_size)
del X_all, y_all
gc.collect()
now = time.time()-start_time
print(f"Data processing duration: {int(now//60)}min {int(now%60)}s\n")
return [dataloader, dataloader_length]
if __name__ == "__main__":
### Main parameters ###
# Path
LABELS_PATH = FINAL_LABELS_PATH
# Audio parameters
SR = SAMPLE_RATE # Sample Rate
AUDIO_DURATION = 10 # 10 seconds duration window for all audios
# Model parameters
BATCH_SIZE = 8
# Misc parameters
RANDOM_STATE = 42
random.seed(RANDOM_STATE)
### Data processing ###
labels_df = pd.read_csv(LABELS_PATH)
print('Testing the process_data function with 10% of the dataset')
alldataloader, dataloader_length = process_data(df=labels_df.take(np.arange(int(len(labels_df)*0.1))), batch_size=BATCH_SIZE,
sample_rate=SR, audio_duration=AUDIO_DURATION,
random_state=RANDOM_STATE, do_plot=True)