Triplet_Mining/predict_async.py at main · morgan76/Triplet_Mining · GitHub

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
# -*- coding: utf-8 -*-
import argparse
import os
from tqdm import tqdm
import torch
import numpy as np

from configuration import config
import input_output as io
from models import EmbedNet
from eval_segmentation import *
import joblib
import warnings
warnings.filterwarnings("ignore")
from utils import clean_tracklist
from embed import embed_tracks


P3s = []
R3s = []
F3s = []


def apply_async_with_callback(embeddings_list, tracklist, config, segmentation, clustering, level, masks, embedding_levels=[], annot=0):

    if segmentation:
        if config.model_name == 'features' and 'SALAMI_2annot' in embeddings_list[0][0]:
            print('eval_segmentation_async_baseline salami')
            #eval_segmentation_async_baseline_salami
            jobs = [ joblib.delayed(eval_segmentation_async_baseline_salami)(audio_file=i[0], config=config, level=level, annot=annot) for i in tqdm(embeddings_list) ]
            out = joblib.Parallel(n_jobs=32, verbose=1)(jobs)
            if not config.test:
                out = out[:,1:].astype(float)
            print('P1 =', np.mean(out[:,0]),'+/-', np.std(out[:,0]))
            print('R1 =', np.mean(out[:,1]),'+/-', np.std(out[:,1]))
            print('F1 =', np.mean(out[:,2]),'+/-', np.std(out[:,2]))
            print('P3 =', np.mean(out[:,3]),'+/-', np.std(out[:,3]))
            print('R3 =', np.mean(out[:,4]),'+/-', np.std(out[:,4]))
            print('F3 =', np.mean(out[:,5]),'+/-', np.std(out[:,5]))
            print('PFC =', np.mean(out[:,6]),'+/-', np.std(out[:,6]))
            print('NCE =', np.mean(out[:,7]),'+/-', np.std(out[:,7]))
            return np.mean(out[:,2]), np.mean(out[:,3]), np.mean(out[:,4]), np.mean(out[:,5])

        elif config.model_name == 'features' and 'SALAMI_2annot' not in embeddings_list[0][0]:
            print('eval_segmentation_async_baseline')
            jobs = [ joblib.delayed(eval_segmentation_async_baseline)(audio_file=i[0], config=config, level=level, annot=annot) for i in tqdm(embeddings_list) ]
            out = joblib.Parallel(n_jobs=32, verbose=1)(jobs)
            out = np.array(out)
            if not config.test:
                out = out[:,1:].astype(float)
            print('P1 =', np.mean(out[:,0]),'+/-', np.std(out[:,0]))
            print('R1 =', np.mean(out[:,1]),'+/-', np.std(out[:,1]))
            print('F1 =', np.mean(out[:,2]),'+/-', np.std(out[:,2]))
            print('P3 =', np.mean(out[:,3]),'+/-', np.std(out[:,3]))
            print('R3 =', np.mean(out[:,4]),'+/-', np.std(out[:,4]))
            print('F3 =', np.mean(out[:,5]),'+/-', np.std(out[:,5]))
            print('PFC =', np.mean(out[:,6]),'+/-', np.std(out[:,6]))
            print('NCE =', np.mean(out[:,7]),'+/-', np.std(out[:,7]))
            return np.mean(out[:,2]), np.mean(out[:,3]), np.mean(out[:,4]), np.mean(out[:,5])

        elif 'SALAMI_2annot' in embeddings_list[0][0] and (embedding_levels == [] or config.training_strategy not in ['csn', 'csn_supervised']):
            print('eval_segmentation_salami whole')
            jobs = [ joblib.delayed(eval_segmentation_salami_whole)(audio_file=i[0], embeddings=i[1], config=config, level=level) for i in tqdm(embeddings_list) ]
            out = joblib.Parallel(n_jobs=32, verbose=1)(jobs)
            out = np.array(out)
            if not config.test:
                out = out[:,1:].astype(float)
            print('Results on SALAMI, annotation level =', level, ', whole embedding matrix')
            print('P1 =', np.mean(out[:,0]),'+/-', np.std(out[:,0]))
            print('R1 =', np.mean(out[:,1]),'+/-', np.std(out[:,1]))
            print('F1 =', np.mean(out[:,2]),'+/-', np.std(out[:,2]))
            print('P3 =', np.mean(out[:,3]),'+/-', np.std(out[:,3]))
            print('R3 =', np.mean(out[:,4]),'+/-', np.std(out[:,4]))
            print('F3 =', np.mean(out[:,5]),'+/-', np.std(out[:,5]))
            print('PFC =', np.mean(out[:,6]),'+/-', np.std(out[:,6]))
            print('NCE =', np.mean(out[:,7]),'+/-', np.std(out[:,7]))
            return np.mean(out[:,2]), np.mean(out[:,3]), np.mean(out[:,4]), np.mean(out[:,5])


        else:
            print('eval_segmentation_async')
            jobs = [ joblib.delayed(eval_segmentation_async)(audio_file=i[0], embeddings=i[1], config=config, level=level) for i in tqdm(embeddings_list) ]
            out = joblib.Parallel(n_jobs=32, verbose=1)(jobs)
            if config.test:
                out = np.array(out)
            else:
                out = np.array(out)
                out = out[:,1:].astype(float)
            if len(out.shape)<2:
                print(out)
            print('Results on annotation level =', level, ', whole embedding matrix')
            print('P1 =', np.nanmean(out[:,0]),'+/-', np.nanstd(out[:,0]))
            print('R1 =', np.nanmean(out[:,1]),'+/-', np.nanstd(out[:,1]))
            print('F1 =', np.nanmean(out[:,2]),'+/-', np.nanstd(out[:,2]))
            print('P3 =', np.nanmean(out[:,3]),'+/-', np.nanstd(out[:,3]))
            print('R3 =', np.nanmean(out[:,4]),'+/-', np.nanstd(out[:,4]))
            print('F3 =', np.nanmean(out[:,5]),'+/-', np.nanstd(out[:,5]))
            print('PFC =', np.nanmean(out[:,6]),'+/-', np.nanstd(out[:,6]))
            print('NCE =', np.nanmean(out[:,7]),'+/-', np.nanstd(out[:,7]))
            return out


def main(config):
    feature_dir = os.path.join(config.ds_path, 'features')
    embedding_dir = os.path.join(feature_dir, 'Embeddings')
    if not os.path.exists(embedding_dir):
        os.makedirs(embedding_dir)

    # loading cuda and model
    use_cuda = torch.cuda.is_available() and not config.no_cuda
    device = torch.device('cuda' if use_cuda else 'cpu')

    if config.model_name != 'features':

        embedding_net = EmbedNet(config)
        model_path = os.path.join(config.models_dir, config.model_name + "." + 'pt')
        embedding_net.load_state_dict(torch.load(model_path, map_location=device)['state_dict'])
        embedding_net.to(device)
        embedding_net.eval()


    print('model loaded')
    # loading and cleaning tracklist
    tracklist = clean_tracklist(config, annotations=True)
    print('tracklist cleaned', len(tracklist))

    if config.model_name != 'features':
        embeddings_list = embed_tracks(tracklist, embedding_net, config, device)
    else:
        embeddings_list = [(i,0) for i in tracklist]
    mask = 0

    #embeddings_list = embed_tracklist(embedding_net, tracklist, device, config)
    print('Device =', device)
    device = 'cpu'
    print('Device =', device)
    print('Length embeddings list =', len(embeddings_list))
    return embeddings_list, tracklist, mask


if __name__ == '__main__':

    parser = argparse.ArgumentParser(description='predict')
    parser.add_argument('ds_path', type=str, help='Path to the  dataset.')
    parser.add_argument('feat_id', type=str, help='Type of feature before embedding.')
    parser.add_argument('model_name', type=str, help='Model Name')
    parser.add_argument('n_embedding', type=int, help='n_embedding')
    parser.add_argument('level', type=int, help='level')
    parser.add_argument('embed_levels', type=int, help='embed_levels')
    parser.add_argument('mode', type=str, help='mode')
    args, _ = parser.parse_known_args()
    warnings.filterwarnings("ignore")
    config.model_name = args.model_name
    config.ds_path = args.ds_path
    if args.mode == 'seg':
        segmentation = True
        clustering = False
    else:
        segmentation = False
        clustering = True
    config.training_strategy = 'triplet_features'
    config.architecture = 'EmbedNet'
    config.use_batch_norm = True
    config.use_dropout = False
    config.no_cuda = False
    config.feat_id = args.feat_id
    config.nb_workers = 0
    config.feat_type = 'beat_sync'
    config.min_samples = 1
    config.batch_size = 128
    config.embedding.n_embedding = args.n_embedding
    embeddings_list, tracklist, mask = main(config)
    print('Length embeddings list =', len(embeddings_list))
    level = int(args.level)
    e_levels = args.embed_levels
    annot = 0
    config.test = False
    print(config)
    if e_levels == 0:
        embedding_levels = [0]
    elif e_levels == 1:
        embedding_levels = [1]
    else:
        embedding_levels = [0,1]
    print('Embedding levels =',embedding_levels)
    print('Annotation level =', level)
    print('Segmentation =', segmentation)
    print('Clustering =', clustering)
    apply_async_with_callback(embeddings_list, tracklist, config, segmentation, clustering, level, mask, embedding_levels, annot)