/
knn_main.py
212 lines (178 loc) · 7.41 KB
/
knn_main.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
from argparse import ArgumentParser
import numpy as np
import librosa
import pickle
import torch
import os
from utils import SDR, get_mir_scores, load_pkl
def parse_arguments():
parser = ArgumentParser()
parser.add_argument("-p", "--is_proj", action='store_true',
help="kNN on projections")
parser.add_argument("--use_stft", action='store_true',
help="Using STFT features")
parser.add_argument("--use_mfcc", action='store_true',
help="Using MFCC features")
parser.add_argument("--use_log_mel", action='store_true',
help="Using log Mel features")
parser.add_argument("--is_closed", action='store_true',
help="Open (Test) / Closed (Val)")
parser.add_argument("--load_model", type=str, default=None,
help="Trained projections")
parser.add_argument("-n", "--n_proj", type=int, default=None,
help = "Number of projections")
parser.add_argument("--use_perc", type=float, default=1.0,
help = "Random sample %% of training set")
parser.add_argument("-k", "--K", type=int, default=10,
help = "Number of neighbors")
parser.add_argument("--seed", type=int, default=42,
help = "Seed for random sampling from dictionary")
parser.add_argument("--gpu_id", type=int, default=-1,
help = "GPU ID. -1 for ")
parser.add_argument("--print_every", type=int, default=100,
help = "Option for printing frequency")
return parser.parse_args()
def main():
args = parse_arguments()
if args.gpu_id != -1:
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"]=str(args.gpu_id)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# Decide features to work on (STFT, Mel (M=128), or MFCC)
if args.use_stft:
Xtr_load_nm = "Xtr_STFT.npy"
Xva_load_nm = "Xva_STFT.pkl"
Xte_load_nm = "Xte_STFT.pkl"
elif args.use_mfcc:
Xtr_load_nm = "Xtr_MFCC.npy"
Xva_load_nm = "Xva_MFCC.pkl"
Xte_load_nm = "Xte_MFCC.pkl"
elif args.use_log_mel:
Xtr_load_nm = "Xtr_log_Mel.npy"
Xva_load_nm = "Xva_log_Mel.pkl"
Xte_load_nm = "Xte_log_Mel.pkl"
# Name for saving results
feature_or_model = Xtr_load_nm.split('.')[0]
if args.load_model:
feature_or_model = args.load_model.split('/')[-1]
save_nm = "{}/results_Prj[{}|{}]_Fom[{}]_See[{}]_Per[{}]_Cld[{}]".format(
"Results_Checking", args.is_proj, args.n_proj, feature_or_model,
args.seed, int(args.use_perc*100), args.is_closed)
print ("Starting script on GPU {}...".format(args.gpu_id))
print ("Job: {}".format(save_nm))
# Load training dictionary
Xtr = np.load(Xtr_load_nm)
Ytr = np.load("IBM_STFT.npy")
# Random sampling
np.random.seed(args.seed)
perm_idx = np.random.permutation(len(Xtr))[:int(len(Xtr) * args.use_perc)]
Xtr = Xtr[perm_idx]
Ytr = Ytr[perm_idx]
# Load testing data
if args.is_closed:
print ("Loading closed set...")
Xva = load_pkl(Xva_load_nm)
vaX = load_pkl("vaX_STFT.pkl")
with open('vasnx_wavefiles.pkl', 'rb') as handle:
val_waves_dict = pickle.load(handle)
else:
print ("Loading open set...")
Xva = load_pkl(Xte_load_nm)
vaX = load_pkl("teX_STFT.pkl")
with open('tesnx_wavefiles.pkl', 'rb') as handle:
val_waves_dict = pickle.load(handle)
vas = val_waves_dict['s']
van = val_waves_dict['n']
vax = val_waves_dict['x']
# Load projections and apply
if args.is_proj:
print ("Loading projections, ", args.load_model)
if args.load_model:
projections = np.load(
"{}_projs.npy".format(args.load_model))
projections = projections.squeeze().T
projections = projections[:,:args.n_proj]
else:
# LSH Random projection baseline
np.random.seed(args.seed)
# projections = np.random.rand(Xtr.shape[1], args.n_proj)
projections = np.random.normal(loc=0.0,
scale=1./args.n_proj,
size=(Xtr.shape[1], args.n_proj))
print ("Projections Shape: ", projections.shape)
if args.gpu_id != -1:
Xtr = torch.cuda.FloatTensor(Xtr)
Xtr_bias = torch.cat((Xtr, torch.ones((len(Xtr),1)).cuda()), 1)
projections = torch.cuda.FloatTensor(projections)
applied_tr = torch.sign(Xtr_bias.mm(projections))
else:
applied_tr = np.sign(np.dot(Xtr, projections))
else:
# Oracle kNN baseline
if args.gpu_id != -1:
applied_tr = torch.cuda.FloatTensor(Xtr)
else:
applied_tr = Xtr
# Get scores
N_va = len(Xva)
SDRlist=np.zeros(N_va)
ml=np.zeros(N_va)
mirSDRlist=np.zeros(N_va)
mirSIRlist=np.zeros(N_va)
mirSARlist=np.zeros(N_va)
for i in range(N_va):
# Apply projections on validation data
if args.is_proj:
if args.gpu_id != -1:
Xva_i = torch.cuda.FloatTensor(Xva[i])
Xva_i_bias = torch.cat((Xva_i, torch.ones((len(Xva_i),1)).cuda()), 1)
applied_vate = torch.sign(Xva_i_bias.mm(projections))
else:
applied_vate = np.sign(np.dot(Xva[i], projections))
else:
if args.gpu_id != -1:
applied_vate = torch.cuda.FloatTensor(Xva[i])
else:
applied_vate = Xva[i]
# Compute cosine similarity scores
if args.gpu_id != -1:
scores = applied_tr.mm(applied_vate.t())
scores = scores.detach().cpu().numpy()
else:
scores = np.dot(applied_tr, applied_vate.T)
if args.is_proj:
scores = ((scores+args.n_proj)/2)
# Apply average of K IBMs
K_locs = np.argpartition(-scores, args.K, 0)[:args.K]
Yhat = np.mean(Ytr[K_locs],0)
applied = vaX[i] * Yhat
recon = librosa.istft(applied.T, hop_length=256)
ml[i], SDRlist[i] = SDR(vas[i], recon)
msdr, msir, msar = get_mir_scores(
vas[i], van[i], vax[i], recon)
mirSDRlist[i] = msdr
mirSIRlist[i] = msir
mirSARlist[i] = msar
# Print intermediate results
if (i+1) % args.print_every == 0:
curr_tot_SDR = np.sum(ml*SDRlist/np.sum(ml))
curr_tot_mSDR = np.sum(ml*mirSDRlist/np.sum(ml))
curr_tot_mSIR = np.sum(ml*mirSIRlist/np.sum(ml))
curr_tot_mSAR = np.sum(ml*mirSARlist/np.sum(ml))
prog = (i+1)/len(vas)*100
print ("{}: {:.2f} SDR {:.2f} mSDR {:.2f} mSIR {:.2f} mSAR {:.2f}"
.format(i+1, prog, curr_tot_SDR, curr_tot_mSDR,
curr_tot_mSIR, curr_tot_mSAR))
# Save results
result_dict = {
'sdr': SDRlist,
'msdr': mirSDRlist,
'msir': mirSIRlist,
'msar': mirSARlist,
'ml': ml
}
with open('{}.pkl'.format(save_nm), 'wb') as handle:
pickle.dump(result_dict, handle)
print ("Results saved!")
if __name__ == "__main__":
main()