-
Notifications
You must be signed in to change notification settings - Fork 1
/
data_classifier.py
178 lines (156 loc) · 7.85 KB
/
data_classifier.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
166
167
168
169
170
171
172
173
174
175
176
177
178
from scipy import signal
from scipy.ndimage import gaussian_filter
from tensorflow.keras import layers, models
from const import *
import numpy as np
import tensorflow as tf
class DataClassifier:
def __init__(self, firebase_comm):
self.firebase_comm = firebase_comm
self.data = None
self.labels = None
self.model = None
self.count = 0
self.viable_data = True
self.model_fetched = False
# Fetches the number of files pertaining to left/right motion data
def initialize_file_counts(self):
self.left_motion_file_count = self.firebase_comm.get_left_motion_file_count()
self.right_motion_file_count = self.firebase_comm.get_right_motion_file_count()
# Checks if our data is balanced enough
def check_if_data_viable(self):
file_count_ratio = self.left_motion_file_count / self.right_motion_file_count
"""
if file_count_ratio < 0.8 or file_count_ratio > 1.25:
print("It is ill-advised to create a network with skewed data! Please record more data and try again.")
self.viable_data = False
"""
# Create an array that will hold all of our necessary data
def initialize_data_and_labels(self):
num_images = self.left_motion_file_count + self.right_motion_file_count
rows_per_img = STFT_F_SIZE
columns_per_img = STFT_T_SIZE
self.all_data = np.zeros((num_images, rows_per_img, columns_per_img, 1))
self.all_labels = np.zeros((num_images, 1))
# Hmmm maybe shuffle data later on? this isn't well encapsulated since random_indices is defined here but only used in another function
self.random_indices = np.arange(num_images)
np.random.shuffle(self.random_indices)
# Le butterworth filter. It filters out the non-important frequencies for us
def filter_data(self, data, fs=128):
nyq = 0.5 * fs
low = MIN_FREQ / nyq
high = MAX_FREQ / nyq
b, a = signal.butter(9, [low, high], btype="band") # Bandpass, which means we select a "band" of frequencies
# What are a and b? what do signal.butter and signal.lfilter do?
b_notch, a_notch = signal.iirnotch(30, 5, 128)
filtered = signal.filtfilt(b_notch, a_notch, data)
b_notch, a_notch = signal.iirnotch(31, 5, 128)
filtered = signal.filtfilt(b_notch, a_notch, data)
b_notch, a_notch = signal.iirnotch(32, 5, 128)
filtered = signal.filtfilt(b_notch, a_notch, filtered)
filtered = signal.lfilter(b, a, filtered)
return filtered
def preprocess_signal(self, data):
data = (data - np.min(data)) / (np.max(data) - np.min(data))
filtered_signal = self.filter_data(data)
f, t, signal_stft = signal.stft(filtered_signal, nperseg=98)
signal_stft = np.abs(signal_stft)
mean = np.mean(signal_stft)
var = np.std(signal_stft)
outlier = mean + 2*var
signal_stft[signal_stft > outlier] = mean
signal_stft = gaussian_filter(signal_stft, sigma=1)
return signal_stft
# Add data from the database to our classifier
def load_in_data(self, data, label):
self.count = 0
for signal in data:
signal_dict = signal.to_dict()
c3_signal = signal_dict['c3_data']
c4_signal = signal_dict['c4_data']
c3_signal_stft = self.preprocess_signal(c3_signal)
c4_signal_stft = self.preprocess_signal(c4_signal)
print(c3_signal_stft.shape)
for i in range(STFT_F_SIZE):
for j in range(STFT_T_SIZE):
self.all_data[self.random_indices[self.count]][i][j][0] = c3_signal_stft[i][j] #(c3_signal_stft[i][j]/c4_signal_stft[i][j])
self.all_labels[self.random_indices[self.count]] = label
self.count += 1
# Verifies, initializes and loads our data and labels
def build_data(self):
print("****************** CREATING NETWORK ******************")
self.initialize_file_counts()
self.check_if_data_viable()
if not self.viable_data:
return None, None
self.initialize_data_and_labels()
# Load in left motion data
left_motion_data = self.firebase_comm.left_motion_recordings.stream()
self.load_in_data(data=left_motion_data, label=0)
# Load in right motion data
right_motion_data = self.firebase_comm.right_motion_recordings.stream()
self.load_in_data(data=right_motion_data, label=1)
return self.all_data, self.all_labels
# Makes our keras model
def make_model(self, input_shape):
# What are all these settings??
self.model = models.Sequential()
self.model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=input_shape))
self.model.add(layers.MaxPooling2D((2, 2)))
self.model.add(layers.Conv2D(64, (3, 3), activation='relu'))
self.model.add(layers.MaxPooling2D((2, 2)))
self.model.add(layers.Conv2D(64, (3, 3), activation='relu'))
self.model.add(layers.Flatten())
self.model.add(layers.Dense(64, activation='relu'))
self.model.add(layers.Dense(2))
# Takes in the training data (STFT, as an image) and labels. Outputs a trained network
def train_network(self, images, labels):
num_images = len(images)
assert num_images == len(labels)
num_of_train_images = int(num_images * .2)
# Split 20% of the data into the test set and the rest into the training set
test_images = images[num_of_train_images:]
test_labels = labels[num_of_train_images:]
train_images = images[:num_of_train_images]
train_labels = labels[:num_of_train_images]
self.make_model(input_shape=train_images.shape[1:])
self.model.compile(optimizer='adam', loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True), metrics=['accuracy'])
self.model.fit(train_images, train_labels, epochs=10)
_, accuracy = self.model.evaluate(test_images, test_labels)
print(f'ACCURACY: {accuracy}')
self.firebase_comm.save_model(self.model)
# This is used to covert single signals streamed from the headset into a usuable format
def convert_signal(self, c3_data, c4_data):
converted_input = np.ones((1, STFT_F_SIZE, STFT_T_SIZE, 1))
c3_data = self.filter_data(c3_data)
c4_data = self.filter_data(c4_data)
f, t, c3_data_stft = signal.stft(c3_data, nperseg=196)
f, t, c4_data_stft = signal.stft(c4_data, nperseg=196)
c3_data_stft = np.abs(c3_data_stft)
c4_data_stft = np.abs(c4_data_stft)
for i in range(STFT_F_SIZE):
for j in range(STFT_T_SIZE):
converted_input[0][i][j][0] *= c3_data_stft[i][j]
#converted_input[0][i][j][0] /= c4_data_stft[i][j]
return converted_input
# Classifies the signal streamed in from the headset
def classify_input(self, c3_data, c4_data):
if not self.model_fetched:
self.firebase_comm.get_model_source()
self.model_fetched = True
# Load TFLite model and allocate tensors.
interpreter = tf.lite.Interpreter(model_path = 'arasi_model.tflite')
interpreter.allocate_tensors()
# Get input and output tensors.
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
# Get input
data = self.convert_signal(c3_data, c4_data)
# Test model on random input data.
input_data = np.array(data, dtype=np.float32)
interpreter.set_tensor(input_details[0]['index'], input_data)
interpreter.invoke()
# The function 'get_tensor()' returns a copy of the tensor data.
# Use 'tensor()' in order to get a pointer to the tensor.
output_data = interpreter.get_tensor(output_details[0]['index'])
return output_data