sound-generation-dynamic-objects/Engine.py at main · objectfolder/sound-generation-dynamic-objects · 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
import os
import os.path as osp
import sys
import json
from pprint import pprint

from tqdm import tqdm, trange
import numpy as np
import matplotlib.pyplot as plt
import torch
import torchvision.utils as vutils
import torch.optim as optim
import math
import librosa

from dataset.build import build as build_dataset
from models.build import build as build_model
from utils.wavenet_vocoder import builder
from utils.metrics import *
from utils.meters import AverageMeter
sys.path.append('../vocoders')
from waveglow.glow import WaveGlow

class Engine():
    def __init__(self, args, cfg):
        self.args = args
        self.cfg = cfg
        # set seeds
        np.random.seed(self.args.seed)
        torch.manual_seed(self.args.seed)
        torch.cuda.manual_seed(self.args.seed)
        # build dataloaders
        self.train_loader, self.val_loader, self.test_loader = build_dataset(
            self.args, self.cfg)
        # build model & optimizer
        self.model = build_model(self.args, self.cfg)
        self.model.cuda()
        # experiment dir
        self.exp_dir = osp.join('./exp', self.args.exp)
        os.makedirs(self.exp_dir, exist_ok=True)
        self.visualization=True

    def train_epoch(self, epoch):
        self.model.train()
        for i, batch in tqdm(enumerate(self.train_loader)):
            self.model.zero_grad()
            self.model.optimize_parameters(batch)
            loss = self.model.criterionL1((self.model.fake_B, self.model.fake_B_postnet), self.model.real_B)
            reduced_loss = loss.item()
            if i % 10 == 0:
                message = "epoch:{} iter:{} loss:{:.6f} G:{:.6f} D:{:.6f} D_r-f:{:.6f} G_s:{:.6f}".format(
                        epoch, i,
                        reduced_loss, self.model.loss_G,
                        self.model.loss_D,
                        (self.model.pred_real - self.model.pred_fake).mean(),
                        self.model.loss_G_silence)
                tqdm.write(message)
        tqdm.write(f"Finish Training Epoch {epoch}")

    def eval_epoch(self, epoch=0, test=False):
        self.model.eval()
        data_loader = self.test_loader if test else self.val_loader
        epoch_l2_distance = AverageMeter()
        with torch.no_grad():
            for i, batch in tqdm(enumerate(data_loader)):
                self.model(batch)
                pred_spec = self.model.real_B.data.cpu().numpy()
                gt_spec = self.model.fake_B_postnet.data.cpu().numpy()
                l2_distance = STFT_L2_distance(pred_spec,gt_spec)
                epoch_l2_distance.update(l2_distance,pred_spec.shape[0])
                if self.visualization:
                    for j in range(len(self.model.fake_B)):
                        plt.figure(figsize=(8, 9))
                        plt.subplot(311)
                        plt.imshow(self.model.real_B[j].data.cpu().numpy(),
                                        aspect='auto', origin='lower')
                        plt.title(self.model.video_name[j]+"_ground_truth")
                        plt.subplot(312)
                        plt.imshow(self.model.fake_B[j].data.cpu().numpy(),
                                        aspect='auto', origin='lower')
                        plt.title(self.model.video_name[j]+"_predict")
                        plt.subplot(313)
                        plt.imshow(self.model.fake_B_postnet[j].data.cpu().numpy(),
                                        aspect='auto', origin='lower')
                        plt.title(self.model.video_name[j]+"_postnet")
                        plt.tight_layout()
                        viz_dir = os.path.join(self.exp_dir, "viz", f'epoch_{epoch:05d}')
                        os.makedirs(viz_dir, exist_ok=True)
                        plt.savefig(os.path.join(viz_dir, self.model.video_name[j]+".jpg"))
                        plt.close()
                        spec_dir = os.path.join(self.exp_dir, "spectrogram", f'epoch_{epoch:05d}')
                        os.makedirs(spec_dir, exist_ok=True)
                        np.save(os.path.join(spec_dir, self.model.video_name[j]+".npy"),
                                self.model.fake_B[j].data.cpu().numpy())
        return epoch_l2_distance.avg

    def train(self):
        bst_epoch_l2_distance = 1e8
        for epoch in range(self.args.epochs):
            print("Start Validation Epoch {}".format(epoch))
            epoch_l2_distance = self.eval_epoch(epoch)
            print("Finish Validation Epoch {}, L2 Distance = {:.8f}".format(epoch, epoch_l2_distance))
            if epoch_l2_distance < bst_epoch_l2_distance:
                print("New best L2 Distance {:.8f} reached, saving best model".format(epoch_l2_distance))
                bst_epoch_l2_distance = epoch_l2_distance
                torch.save(self.model.state_dict(), osp.join(self.exp_dir,'bst.pth'))
            torch.save(self.model.state_dict(), osp.join(self.exp_dir,'latest.pth'))
            print("Start Training Epoch {}".format(epoch))
            self.train_epoch(epoch)
        print("Finish Training Process")

    def load_waveglow(self, ckpt_path='../vocoders/exp/waveglow_small/ckpts/epoch_178.pth'):
        waveglow = WaveGlow(
            n_mel_channels=80,
            n_flows=12,
            n_group=6,
            n_early_every=4,
            n_early_size=2,
            WN_config={
                "n_layers": 4,
                "n_channels": 128,
                "kernel_size": 3
            }
        )
        state_dict = torch.load(ckpt_path)
        waveglow.load_state_dict(state_dict)
        return waveglow

    def evaluate_on_waveform(self):
        # evaluate on waveform
        waveglow = self.load_waveglow().cuda()
        test_envelope_distance = AverageMeter()
        test_cdpam_distance = AverageMeter()
        for root,dirs,files in os.walk(osp.join(self.exp_dir,'spectrogram','epoch_00000')):
            for file in tqdm(files):
                spec_path = osp.join(root,file)
                spec = np.load(spec_path)
                spec = torch.FloatTensor(spec).cuda()
                pred_wave = waveglow.infer(spec.unsqueeze(0), 1.0).cpu().numpy()[0]
                gt_wave,sr = librosa.load(osp.join(self.args.audio_location,'{}.wav'.format(spec_path.split('/')[-1].split('.')[0])))

                envelope_distance = Envelope_distance(pred_wave[:sr*5],gt_wave[:sr*5])
                cdpam_distance = CDPAM_distance(torch.tensor(pred_wave).unsqueeze(0).cuda(),
                               torch.tensor(gt_wave).unsqueeze(0).cuda())

                test_envelope_distance.update(envelope_distance)
                test_cdpam_distance.update(float(cdpam_distance))

        return test_envelope_distance.avg, test_cdpam_distance.avg

    def test(self):
        print("Start Testing")
        print("Loading best model from {}".format(osp.join(self.exp_dir,'bst.pth')))
        self.model.load_state_dict(torch.load(osp.join(self.exp_dir, 'bst.pth')))
        test_l2_distance = self.eval_epoch(test = True)
        test_envelope_distance, test_cdpam_distance = self.evaluate_on_waveform()
        result = {
            "Test Result":{
                'L2 Distance': test_l2_distance,
                'Envelope Distance': test_envelope_distance,
                'CDPAM Distance': test_cdpam_distance,
            }
        }
        json.dump(result, open(osp.join(self.exp_dir, 'result.json'),'w'))
        print("Finish Testing, L2 Distance = {:.8f}".format(test_l2_distance))

    def __call__(self):
        if not self.args.eval:
            self.train()
        self.test()