-
Notifications
You must be signed in to change notification settings - Fork 0
/
time_of_flight_net.py
190 lines (171 loc) · 6.94 KB
/
time_of_flight_net.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
179
180
181
182
183
184
185
186
187
188
189
190
import torch
import torch.nn as nn
import torch.fft as fft
import numpy as np
from utils import is_power_of_2
from assert_eq import assert_eq
from reshape_layer import Reshape
from signals_and_geometry import time_of_flight_crop
class TimeOfFlightNet(nn.Module):
def __init__(
self,
speed_of_sound,
sampling_frequency,
recording_length_samples,
crop_length_samples,
emitter_location,
receiver_locations,
use_convolutions,
hidden_features,
kernel_size,
use_fourier_transform,
no_amplitude_compensation,
):
super(TimeOfFlightNet, self).__init__()
assert isinstance(speed_of_sound, float)
assert isinstance(sampling_frequency, float)
assert isinstance(recording_length_samples, int)
assert is_power_of_2(recording_length_samples)
assert isinstance(crop_length_samples, int)
assert is_power_of_2(crop_length_samples)
assert isinstance(emitter_location, np.ndarray)
assert_eq(emitter_location.shape, (3,))
assert_eq(emitter_location.dtype, np.float32)
assert isinstance(receiver_locations, np.ndarray)
assert receiver_locations.dtype == np.float32
assert receiver_locations.shape[1:] == (3,)
assert isinstance(use_convolutions, bool)
assert isinstance(hidden_features, int)
assert isinstance(kernel_size, int)
assert isinstance(use_fourier_transform, bool)
assert isinstance(no_amplitude_compensation, bool)
self.speed_of_sound = speed_of_sound
self.sampling_frequency = sampling_frequency
self.recording_length_samples = recording_length_samples
self.crop_length_samples = crop_length_samples
self.emitter_location = nn.parameter.Parameter(
data=torch.tensor(emitter_location, dtype=torch.float32),
requires_grad=False,
)
num_receivers = receiver_locations.shape[0]
self.num_receivers = num_receivers
receiver_locations_tensor = torch.tensor(
receiver_locations, dtype=torch.float32
)
assert receiver_locations_tensor.shape == (num_receivers, 3)
self.receiver_locations = nn.parameter.Parameter(
data=receiver_locations_tensor,
requires_grad=False,
)
self.no_amplitude_compensation = no_amplitude_compensation
self.use_fourier_transform = use_fourier_transform
if self.use_fourier_transform:
self.input_length = (crop_length_samples // 2) + 1
self.channels_per_receiver = 2
else:
self.input_length = crop_length_samples
self.channels_per_receiver = 1
def activation_function():
# return nn.ReLU()
return nn.LeakyReLU(0.1)
if use_convolutions:
final_length = self.input_length // 8 + (
1 if self.use_fourier_transform else 0
)
final_features = 1024 // final_length
conv_padding = (kernel_size - 1) // 2
self.model = nn.Sequential(
nn.BatchNorm1d(num_features=num_receivers * self.channels_per_receiver),
nn.Conv1d(
in_channels=num_receivers * self.channels_per_receiver,
out_channels=hidden_features,
kernel_size=kernel_size,
stride=2,
padding=conv_padding,
),
activation_function(),
nn.BatchNorm1d(num_features=hidden_features),
nn.Conv1d(
in_channels=hidden_features,
out_channels=hidden_features,
kernel_size=kernel_size,
stride=2,
padding=conv_padding,
),
activation_function(),
nn.BatchNorm1d(num_features=hidden_features),
nn.Conv1d(
in_channels=hidden_features,
out_channels=final_features,
kernel_size=kernel_size,
stride=2,
padding=conv_padding,
),
activation_function(),
Reshape(
(final_features, final_length), (final_features * final_length,)
),
nn.Linear(in_features=(final_features * final_length), out_features=1),
)
else:
# Simple 3-layer fully-connected model
self.model = nn.Sequential(
nn.BatchNorm1d(num_features=num_receivers * self.channels_per_receiver),
Reshape(
(num_receivers * self.channels_per_receiver, self.input_length),
(num_receivers * self.channels_per_receiver * self.input_length,),
),
nn.Linear(
in_features=num_receivers
* self.channels_per_receiver
* self.input_length,
out_features=hidden_features,
),
activation_function(),
nn.Linear(
in_features=hidden_features,
out_features=hidden_features,
),
activation_function(),
nn.Linear(in_features=hidden_features, out_features=1),
)
def forward(self, recordings, sample_locations):
recordings_cropped = time_of_flight_crop(
recordings=recordings,
sample_locations=sample_locations,
emitter_location=self.emitter_location,
receiver_locations=self.receiver_locations,
speed_of_sound=self.speed_of_sound,
sampling_frequency=self.sampling_frequency,
crop_length_samples=self.crop_length_samples,
apply_amplitude_correction=(not self.no_amplitude_compensation),
)
B1, B2 = sample_locations.shape[:2]
inputs = recordings_cropped.reshape(
(B1 * B2), self.num_receivers, self.crop_length_samples
)
if self.use_fourier_transform:
inputs = fft.rfft(
inputs, n=self.crop_length_samples, dim=2, norm="backward"
)
assert_eq(
inputs.shape,
(B1 * B2, self.num_receivers, (self.crop_length_samples // 2) + 1),
)
inputs = torch.cat([torch.real(inputs), torch.imag(inputs)], dim=1)
assert_eq(
inputs.shape,
(B1 * B2, 2 * self.num_receivers, (self.crop_length_samples // 2) + 1),
)
assert_eq(
inputs.shape,
(
B1 * B2,
self.num_receivers * self.channels_per_receiver,
self.input_length,
),
)
predictions = self.model(inputs)
assert_eq(predictions.shape, ((B1 * B2), 1))
predictions = predictions.reshape(B1, B2)
return predictions