-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathsound_eval.py
183 lines (145 loc) · 6.26 KB
/
sound_eval.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
import numpy as np
import pandas as pd
import torch
import torchvision
import matplotlib.pyplot as plt
import imp
from pathlib import Path
from pippi.soundbuffer import SoundBuffer
from pippi import dsp,fx
import param_generation as pg
import _pickle as pickle
from IPython.display import Audio
from feature_extraction.mir_utils import *
import torch.utils.data as utils
import torchvision.transforms as transforms
from PIL import Image
import common_vars as comv
import csv
import helpers
imp.reload(helpers)
imp.reload(comv)
imp.reload(pg)
from feature_extraction import pytorch_models as tm
imp.reload(tm)
from helpers import *
from common_vars import SR
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
stack_size=3
BATCH_SIZE=1
NUM_BINS=100
classes=comv.classes
classes_ranked=comv.classes_ranked
cDict={v:i for i,v in enumerate(classes)}
drum_groups=['tom_low','snare','hihat_closed','rim','synth_noise','clap','kick','hihat_open','tom_mid']
drum_df=pd.DataFrame(columns=["name","fc","cnnlstm","env+freq","consensus","stack_size"])
import feature_extraction.pytorch_models as pm
import imp
imp.reload(pm)
input_dim = 120
seq_dim=100
output_size = 5
hidden_dim = 1000
n_layers = 1
#########DVN#########
#env model stuff
env_model_dvn = tm.env_Model(D_in=10,H1=10,H2=5,H3=10,H4=2,H5=10,device=device)
state='../feature_extraction/models/fc/env_dvn_95_35.pt'
env_model_dvn.load_state_dict(torch.load(state))
env_trans=tm.envTrans()
env_model_dvn.to(device)
# pitch model stuff
fc_spec_model_dvn=tm.getFCSpecModel(D_in=20*20,H1=20,H2=10,H3=4,D_out=2)
state='../feature_extraction/models/fc/spec_fc_dvn_96.pt'
fc_spec_model_dvn.load_state_dict(torch.load(state))
fc_spec_trans_dvn=tm.specTrans(20)
fc_spec_model_dvn.to(device)
#cnn lstm
cnn_model_dvn=pm.CNNLSTM_dvn()
state='../feature_extraction/models/cnn/clst_dvn_96_279.pt'
cnn_model_dvn.load_state_dict(torch.load(state))
cnn_model_dvn.to(device)
cnn_model_dvn.to(device)
#########DVD#########
#fc dvd
dvd_fc=pm.getFCSpecModel(D_in=20*20,H1=100,H2=50,H3=100,D_out=len(drum_groups))
state='../feature_extraction/models/fc/spec_fc_dvd_81.pt'
dvd_fc.load_state_dict(torch.load(state))
dvd_fc.to(device)
#cnn_dvd
dvd_cnn=pm.CNNLSTM_dvd(len_out=len(drum_groups))
state='../feature_extraction/models/cnn/clst_dvd_90_170.pt'
dvd_cnn.load_state_dict(torch.load(state))
dvd_cnn.to(device)
#freq+env model
freq_env_model_dvd = tm.env_freq_Model(D_in=60,H1=30,H2=10,H3=10,D_out=len(drum_groups))
state='../feature_extraction/models/fc/freq_env_dvd_86_351.pt'
freq_env_model_dvd.load_state_dict(torch.load(state))
freq_env_trans=tm.freq_and_env_Trans(feat_mels=50,env_mels=1)
freq_env_model_dvd.to(device)
def evalSound(a):
transform_input= {"signal":torch.tensor(a,dtype=torch.float),"label":torch.tensor(0)}
env_feats=env_trans.call(transform_input)["feats"].to(device)
freq_feats=freq_env_trans.call(transform_input)["feats"].to(device)
pitch_feats=fc_spec_trans_dvn.call(transform_input)["feats"].to(device)
spec_flat=torch.flatten(pitch_feats, start_dim=1)
outputEnvFreq=freq_env_model_dvd(freq_feats)
outputCNNDVD=dvd_cnn(pitch_feats.reshape(-1,1,20,20))
outputFCDVD=dvd_fc(spec_flat)
outputE=env_model_dvn(env_feats)
outputSPECFC=fc_spec_model_dvn(spec_flat)
outputCNN= cnn_model_dvn(pitch_feats.reshape(-1,1,20,20))
o_e,o_fc_spec,o_cnn_spec=outputE[0].item(), outputSPECFC[0][0].item() ,outputCNN[0][0].item()
tote=outputE[0].item()+outputSPECFC[0][0].item()
gfc=drum_groups[torch.argmax(outputFCDVD).item()]
gcnn=drum_groups[torch.argmax(outputCNNDVD).item()]
genvfreq=drum_groups[torch.argmax(outputEnvFreq).item()]
cat_consensus=drum_groups[torch.argmax(outputFCDVD+outputCNNDVD+outputEnvFreq)]
spec_score=(o_fc_spec+o_cnn_spec)/2
# print("env score",o_e)
# print("fc_spec",o_fc_spec)
# print("o_cnn_spec",o_cnn_spec)
# print("fc category",gfc,"\ncnn category",gcnn,"\nenv+freq",genvfreq)
# print("total consensus",cat_consensus)
# return(a,env_feats,freq_feats,pitch_feats,stack_size,"found")
return o_e
# out,params= stackMaker(stack_size)
def findDrum(stack_size=1,iteration=0):
a= memToAud(out)
return a,evalSound(a)
#used by GA for classsification
class drum_classifier():
def __init__(self,device="cpu"):
self.device=device
def transformInput(self,a):
self.transform_input= {"signal":torch.tensor(a,dtype=torch.float),"label":torch.tensor(0)}
def getEnvDVN(self):
#return prob of being drum
self.env_feats=env_trans.call(self.transform_input)["feats"].to(self.device)
self.outputE=env_model_dvn(self.env_feats)
return self.outputE[0].item()
def getEnvFreqDVD(self):
self.freq_feats=freq_env_trans.call(self.transform_input)["feats"].to(self.device)
self.outputEnvFreq=freq_env_model_dvd(self.freq_feats)
self.envfreq_cat=drum_groups[torch.argmax(self.outputEnvFreq).item()]
return dict(zip(drum_groups,self.outputEnvFreq.tolist()))
def drumTypeEstimation(self,predictions):
#predictions is a array of prediction values, which are mapped to drum types
#returns most likely and least likely
likely_drum_index=torch.argmax(predictions).item()
unlikely_drum_index=torch.argmin(predictions).item()
return {"likely":[predictions[likely_drum_index],drum_groups[likely_drum_index]],
"unlike":[predictions[unlikely_drum_index],drum_groups[unlikely_drum_index]]}
def setAllFeats():
self.pitch_feats=fc_spec_trans_dvn.call(self.transform_input)["feats"].to(self.device)
self.spec_flat=torch.flatten(self.pitch_feats, start_dim=1)
def setAllOutputs():
self.outputCNNDVD=dvd_cnn(self.pitch_feats.reshape(-1,1,20,20))
self.outputCNN= cnn_model_dvn(self.pitch_feats.reshape(-1,1,20,20))
self.outputFCDVD=dvd_fc(self.spec_flat)
self.outputSPECFC=fc_spec_model_dvn(self.spec_flat)
o_fc_spec,o_cnn_spec=outputSPECFC[0][0].item() ,outputCNN[0][0].item()
tote=outputE[0].item()+outputSPECFC[0][0].item()
gfc=drum_groups[torch.argmax(outputFCDVD).item()]
gcnn=drum_groups[torch.argmax(outputCNNDVD).item()]
cat_consensus=drum_groups[torch.argmax(outputFCDVD+outputCNNDVD+self.outputEnvFreq)]