-
Notifications
You must be signed in to change notification settings - Fork 5
/
music_main_nn_over_BS.py
226 lines (192 loc) · 7.55 KB
/
music_main_nn_over_BS.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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
# from visualize import save_matrices, save_dance
from multiprocessing import Pool
import sys
from PIL import Image
import numpy as np
import itertools
import argparse
import random
import scipy
import os
import time
from sklearn.decomposition import PCA
from sklearn.metrics.pairwise import cosine_similarity
import wave
import simpleaudio as sa
import torch
import torch.nn as nn
join = os.path.join
random.seed(123)
num_classes = 5
def get_mapped_note(note_idx):
mappings = {0: 60, 1: 62, 2: 64, 3: 67, 4: 69}
return mappings[note_idx]
def get_note_wave(note_val, duration=10.0):
wave_read = wave.open("music_files/" + str(note_val) + ".wav", 'rb')
wave_obj = sa.WaveObject.from_wave_read(wave_read)
return wave_obj
def take_action(cur_note, prev_note, prev_play_obj, note_waves):
if prev_note == -1 or cur_note != prev_note:
print("note changed")
if prev_note != -1:
prev_play_obj.stop()
play_obj = note_waves[cur_note].play()
return play_obj, 1
return prev_play_obj, 0
def preprocess_dance_matrix(dance_matrix):
l = len(dance_matrix)
if l != 60:
temp = np.zeros([60, 60])
temp[-l:,-l:] = dance_matrix
return np.expand_dims(temp, axis=0)
else:
return np.expand_dims(dance_matrix, axis=0)
def preprocess_past_music(past_music):
temp = [num_classes]*10 + list(past_music)
temp = temp[-10:]
new_past_music = np.array(temp)
return np.expand_dims(new_past_music, axis=0)
class Net(nn.Module):
def __init__(self):
# call init function of parent class
super(Net, self).__init__()
# define / initialize layers of network
self.embedding_dim = embedding_dim = 16
self.latent_dim_dance = latent_dim_dance = 32
self.latent_dim_music = latent_dim_music = 32
self.use_batch_norm = use_batch_norm = False
self.use_lstm = use_lstm = True
if use_batch_norm:
self.danceNet = nn.Sequential(
nn.Conv2d(in_channels=1, out_channels=64, kernel_size=3, stride=1),
nn.BatchNorm2d(64),
nn.ReLU(True),
nn.MaxPool2d(2, stride=2),
nn.Conv2d(64, 128, 3),
nn.BatchNorm2d(128),
nn.ReLU(True),
nn.MaxPool2d(2, stride=2),
nn.Conv2d(128, 256, 3),
nn.ReLU(True),
nn.Conv2d(256, 512, 3),
nn.ReLU(True),
nn.Conv2d(512, latent_dim_dance, 3)
)
else:
self.danceNet = nn.Sequential(
nn.Conv2d(in_channels=1, out_channels=64, kernel_size=3, stride=1),
nn.ReLU(True),
nn.MaxPool2d(2, stride=2),
nn.Conv2d(64, 128, 3),
nn.ReLU(True),
nn.Conv2d(128, 128, 3),
nn.ReLU(True),
nn.MaxPool2d(2, stride=2),
nn.Conv2d(128, 256, 3),
nn.ReLU(True),
nn.Conv2d(256, 512, 3),
nn.ReLU(True),
nn.Conv2d(512, latent_dim_dance, 3)
)
if use_lstm:
self.musicNet = nn.LSTM(self.embedding_dim, latent_dim_music, 2,bidirectional=True)
else:
self.musicNet = nn.Sequential(
nn.Linear(num_classes * self.embedding_dim, 256),
nn.ReLU(True),
nn.Linear(256, 512),
nn.ReLU(True),
nn.Linear(512, latent_dim_music)
)
self.danceMusicNet = nn.Sequential(
nn.Linear(latent_dim_dance + 2*latent_dim_music, 512),
# nn.Linear(512, 512),
nn.ReLU(True),
nn.Dropout(p=0.5),
nn.Linear(512, 256),
nn.ReLU(True),
nn.Linear(256, 128),
nn.ReLU(True),
nn.Linear(128, num_classes)
)
self.notes_embedding = nn.Embedding(6, self.embedding_dim)
# defines the computation performed at every call of the module
# this is not called directly, but via instance of Module
def forward(self, d, m):
dance_out = d
music_out = m
# how to pass input here
dance_out = self.danceNet(dance_out)
dance_out = torch.mean(dance_out, (2, 3))
# 32*512
# return dance_out
music_out = self.notes_embedding(music_out)
if self.use_lstm:
batch_size = d.size(0)
h0 = torch.zeros(4, batch_size, self.latent_dim_music)
c0 = torch.zeros(4, batch_size, self.latent_dim_music)
music_out = music_out.permute(1, 0, 2)
music_out, _ = self.musicNet(music_out, (h0, c0))
music_out = music_out[-1]
else:
music_out = music_out.reshape((music_out.shape[0], -1))
music_out = self.musicNet(music_out)
dance_music_concat = torch.cat((dance_out, music_out), 1)
dance_music_out = self.danceMusicNet(dance_music_concat)
return dance_music_out
class Music(object):
def __init__(self, nn_frame_start, music_freq, joint_idx=None, get_note=False):
self.nn_frame_start = nn_frame_start
self.music_freq = music_freq
self.joint_idx = joint_idx
note_mappings = {60: 'C4', 62: 'D4', 64: 'E4', 67: 'G4', 69: 'A4'}
note_waves = {}
for key in note_mappings:
note_waves[key] = get_note_wave(key)
self.note_waves = note_waves
self.prev_note = -1
self.prev_play_obj = None
self.net = Net()
print("loading the model")
self.net.load_state_dict(torch.load("models/models_32_32_16_500_0.0002", map_location=torch.device('cpu'))["state_dict"])
self.net.eval()
print("the model has been loaded")
self.joints = []
self.splits = []
self.get_note = get_note
self.past_music = []
self.music_history = 10
self.joints_history = 60
def add_pose(self, pose, frame_idx):
self.joints.append(pose)
print("the frame idx is ", frame_idx)
if frame_idx % self.music_freq == 0:
if len(self.past_music) == 0:
note_idx = 2
else:
if len(self.joints) >= self.joints_history:
self.joints = self.joints[-self.joints_history:]
joints_similarity = cosine_similarity(np.array(self.joints))
if len(self.past_music) < 10:
joint_matrix = joints_similarity
past_music = self.past_music
else:
joint_matrix = joints_similarity
past_music = self.past_music[-10:]
print("the shape of joint_matrix is ", joint_matrix.shape)
joint_matrix, past_music = preprocess_dance_matrix(joint_matrix), preprocess_past_music(past_music)
joint_matrix = torch.from_numpy(np.expand_dims(joint_matrix, 1)).float()
past_music = torch.from_numpy(past_music).long()
out = self.net(joint_matrix, past_music)
nn_note = out.cpu().data.numpy().argmax(axis=1)[0]
note_idx = nn_note
print("the music produced is ", note_idx)
state = get_mapped_note(note_idx)
self.past_music.append(note_idx)
cur_note = state
self.prev_play_obj, note_change = take_action(cur_note, self.prev_note, self.prev_play_obj, self.note_waves)
self.prev_note = cur_note
return note_change
return 0
if __name__ == "__main__":
music_obj = Music(50, 6, 9)