-
Notifications
You must be signed in to change notification settings - Fork 1
/
hmmopt.py
316 lines (258 loc) · 12.4 KB
/
hmmopt.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
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
# -*- coding: utf-8 -*-
"""
Created on Tue Jul 7 16:00:00 2020
@author: CITI
"""
#%%
import os
import numpy as np
from models.BaselineBLSTM import RNNDownBeatProc as bsl_blstm
from downbeat_dataset import EvalDataset
from multiprocessing import Pool
import tqdm
import torch
from pathlib import Path
import pandas as pd
import mir_eval.util as mir_util
from scipy.special import softmax
from madmom.features.downbeats import DBNDownBeatTrackingProcessor as DownBproc
fps = 100
f_measure_threshold=0.07 # 70ms tolerance as set in paper
beats_per_bar = [3, 4]
def beat_eval(beat_label, beat_est, withdownbeat = True):
if type(beat_est) ==list and len(beat_est)==0: ###1116: added for old models
return []
else:
bmatching = mir_util.match_events(beat_label[:, 0], beat_est[:, 0], f_measure_threshold)
b_reflen = beat_label.shape[0]
b_estlen = beat_est.shape[0]
# downbeat evaluation
if withdownbeat:
downbeat_label = beat_label[:,0][np.where(beat_label[:,1]==1)]
downbeat_est = beat_est[:, 0][np.where(beat_est[:,1]==1)]
dbmatching = mir_util.match_events(downbeat_label, downbeat_est, f_measure_threshold)
db_reflen = len(downbeat_label)
db_estlen = len(downbeat_est)
return len(bmatching), b_reflen, b_estlen, len(dbmatching), db_reflen, db_estlen
else:
return len(bmatching), b_reflen, b_estlen
def res2fmeasure(all_results, head = 'fuser'):
### 1116: added to deal with empty head for old models
all_results_array = np.array(all_results)
if all_results_array.shape[1]>0:
beat_precision = float( all_results_array[:, 0].sum()/all_results_array[:, 2].sum())
beat_recall = float( all_results_array[:, 0].sum()/all_results_array[:, 1].sum())
beat_fmeasure = mir_util.f_measure(beat_precision, beat_recall)
# downbeat
dbeat_precision = float( all_results_array[:, 3].sum()/ all_results_array[:, 5 ].sum())
dbeat_recall = float( all_results_array[:, 3].sum()/ all_results_array[:, 4].sum())
dbeat_fmeasure = mir_util.f_measure(dbeat_precision, dbeat_recall)
result_dict = {
head+'_f sum': beat_fmeasure + dbeat_fmeasure,
head+'_beat F-score': beat_fmeasure,
head+'_beat_precision': beat_precision,
head+'_beat_recall': beat_recall,
head+'_downbeat F-score': dbeat_fmeasure,
head+'_downbeat_precision': dbeat_precision,
head+'_downbeat_recall': dbeat_recall,
}
else:
result_dict = {
head+'_f sum': None,
head+'_beat F-score': None,
head+'_beat_precision': None,
head+'_beat_recall': None,
head+'_downbeat F-score': None,
head+'_downbeat_precision': None,
head+'_downbeat_recall': None,
}
return result_dict
def process_worker(arg_list):
param_dict, beat_list, activation_list = arg_list
hmm_proc = DownBproc(beats_per_bar = beats_per_bar, num_tempi = param_dict['ntempi'],
transition_lambda = param_dict['tlambda'],
observation_lambda = param_dict['olambda'],
threshold = param_dict['thres'], fps = fps)
fuser_results = []
mix_results = []
nodrum_results = []
drum_results = []
for ind, activation in enumerate(activation_list):
# break
if type(activation) ==list:
if len(activation) == 4:
### For drum-aware ensemble models:
(fuser_activation, mix_activation, nodrum_activation, drum_activation) = activation
beat_fuser_est = hmm_proc( fuser_activation)
beat_mix_est = hmm_proc( mix_activation)
beat_nodrum_est = hmm_proc( nodrum_activation)
beat_drum_est = hmm_proc( drum_activation)
else:
print("unexpected len of activation")
else:
fuser_activation = activation
beat_fuser_est = hmm_proc( fuser_activation)
beat_mix_est = [] # not implemented for old models
beat_nodrum_est = [] # not implemented for old models
beat_drum_est = [] # not implemented for old models
beat = beat_list[ind]
fuser_results.append(beat_eval(beat, beat_fuser_est))
mix_results.append(beat_eval(beat, beat_mix_est))
nodrum_results.append(beat_eval(beat, beat_nodrum_est))
drum_results.append(beat_eval(beat, beat_drum_est))
fuser_dict = res2fmeasure(fuser_results, head = 'fuser')
mix_dict = res2fmeasure(mix_results, head = 'mix')
nodrum_dict = res2fmeasure(nodrum_results, head = 'nodrum')
drum_dict = res2fmeasure(drum_results, head= 'drum')
result_dict = {
'n_tempi': param_dict['ntempi'],
'transition_lambda': param_dict['tlambda'],
'observation_lambda': param_dict['olambda'],
'threshold': param_dict['thres'],
}
result_dict.update(fuser_dict)
result_dict.update(mix_dict)
result_dict.update(nodrum_dict)
result_dict.update(drum_dict)
return result_dict
def prediction_conversion(prediction):
if len(prediction.shape) == 2:
prediction = prediction.unsqueeze(0)
pred_arr = prediction.detach().cpu().numpy()
pred_acti = softmax(pred_arr, axis = 2)
pred_acti = pred_acti.squeeze()
model_activation = np.zeros((pred_acti.shape[0], 2))
model_activation[:, 0] = pred_acti[:, 2] # beat class
model_activation[:, 1] = pred_acti[:, 1] # downbeat class
return model_activation
def get_dlm_activation(rnn, device, np_2dfeature):
""" get deep learning model activations"""
input_feature = torch.tensor(np_2dfeature[np.newaxis, :, :]).float().to(device)
rnn.eval()
with torch.no_grad():
activation = rnn(input_feature)
### drum-aware models
if type(activation)==tuple and len(activation) ==6:
beat_fused, beat_mix, beat_nodrum, beat_drum, x_nodrum_hat, x_drum_hat = activation
fuser_activation = prediction_conversion(beat_fused)
mix_activation = prediction_conversion(beat_mix)
nodrum_activation = prediction_conversion(beat_nodrum)
drum_activation = prediction_conversion(beat_drum)
model_activation = [fuser_activation, mix_activation, nodrum_activation, drum_activation]
return model_activation
else:
beat_fused = activation
fuser_activation = prediction_conversion(beat_fused)
return fuser_activation
def getValidtxt_paths(datasets_list, main_dataset_dir, valid_fn = 'valid_audiofiles.txt'):
validtxt_paths = []
for eachdataset in datasets_list:
# break
valid_txt_path = os.path.join(main_dataset_dir, eachdataset, valid_fn)
if os.path.exists(valid_txt_path):
validtxt_paths.append(valid_txt_path)
else:
print("can't find validtxt:", valid_txt_path)
return validtxt_paths
def getEvaldataset_objlist(validtxt_paths):
obj_list = []
for eachvalidtxt in validtxt_paths:
# break
print("======loading {} evaldataset ======".format(Path(eachvalidtxt).parents[0].stem))
evaldataset = EvalDataset(eachvalidtxt)
print("len:", len(evaldataset.datasets))
obj_list += evaldataset.datasets
return obj_list
#%%
def main():
date = '0529'
cuda_num = 0
cuda_str = 'cuda:'+str(cuda_num)
device = torch.device(cuda_str if torch.cuda.is_available() else 'cpu')
process_num = 3
eval_dir = os.path.join('./hmm_optimization/opt_history')
### assign the search range here:
### note: it take a long time to finish if search space is too large
temp_range = np.arange(60, 70, 10, dtype = int)
transition_lambda_range = np.arange(100, 160, 10, dtype = float)
observation_lambda_range = np.arange(16, 20, 8, dtype = int)
threshold_range = np.arange(0.40, 0.80, 0.05, dtype = float)
### get paths of trained models ready for hmmopt
ft_notes_dir = './hmm_optimization/hmmopt_todolist'
ft_notes_csv = os.path.join(ft_notes_dir, 'hmmopt_todolist_0529.csv')
df = pd.read_csv(ft_notes_csv, index_col = 0)
target_model_dict = df.loc[df['hmm']=='ready']
target_model_dict = target_model_dict.to_dict('records')
### get dataset for validation set
main_dataset_dir = './datasets/original/'
datasets_list = os.listdir(main_dataset_dir)
validtxt_paths = getValidtxt_paths(datasets_list, main_dataset_dir, valid_fn = 'valid_audiofiles.txt')
obj_list = getEvaldataset_objlist(validtxt_paths)
song_num ='all'
if song_num != 'all':
obj_list = obj_list[0:song_num]
else:
song_num = len(obj_list)
for modeldict in target_model_dict:
# break
exp_name = os.path.basename(modeldict['model_dir'])
print("======Processing {} ======".format( exp_name ))
### save finetune information
evaluation_folder = eval_dir
if not os.path.exists(evaluation_folder):
Path(evaluation_folder).mkdir(parents = True, exist_ok = True)
target_jsonpath = modeldict['model_dir']
# if 'model_setting' in modeldict.keys():
# model_setting = modeldict['model_setting']
### model initialization
if modeldict['model_type'] =='bsl_blstm':
model = bsl_blstm()
else:
print("===!!!===> unknown model_type")
### finished training, not plot loss curve and apply hmm finetune
model_fn = 'RNNBeatProc.pth'
model_path = os.path.join(target_jsonpath , model_fn)
## load the best model just trained
state = torch.load(model_path, map_location = device)
model.load_state_dict(state)
model.cuda(device.index)
# csv save path
csv_sdir = evaluation_folder
evaluation_title = "HmmFT_"+str(song_num)+'songs_'+exp_name+ '_'
best_para_spath = os.path.join(csv_sdir, evaluation_title+'bestParams_'+ date+'.csv')
csv_spath = os.path.join(csv_sdir, evaluation_title+ date+'.csv')
### collect all activations and beat labels into list for multiprocess
activation_list = []
beatlabel_list = []
for eachsong_obj in tqdm.tqdm(obj_list, desc = 'processing song:'):
# break
feat, beat, audiofile = eachsong_obj.get_data()
activation = get_dlm_activation(model, device, feat)
activation_list.append(activation)
beatlabel_list.append(beat)
args_list = []
for ntempi in temp_range:
for tlambda in transition_lambda_range:
for olambda in observation_lambda_range:
for thres in threshold_range:
param_dict = {
'ntempi': ntempi,
'tlambda': tlambda,
'olambda': olambda,
'thres': thres, }
args_list.append([param_dict, beatlabel_list, activation_list] )
pool = Pool(process_num)
results_dictlist = pool.map(process_worker, tqdm.tqdm(args_list))
csv_results = pd.DataFrame(results_dictlist)
csv_results['model_dir'] = modeldict['model_dir']
csv_results['model_simpname'] = modeldict['model_simpname']
if 'model_setting' in modeldict.keys():
csv_results['model_setting'] = modeldict['model_setting']
else:
csv_results['model_setting'] = ''
csv_results['model_type'] = modeldict['model_type']
best_parameters = csv_results.nlargest(5, "fuser_f sum")
best_parameters.to_csv(best_para_spath)
csv_results.to_csv(csv_spath)
if __name__ == "__main__":
main()