File size: 8,090 Bytes
78e32cc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
###
# Author: Kai Li
# Date: 2022-05-26 18:09:54
# Email: [email protected]
# LastEditTime: 2024-01-24 00:00:28
###
import gc
from omegaconf import OmegaConf
import torch
import pytorch_lightning as pl
from torch.optim.lr_scheduler import ReduceLROnPlateau
from collections.abc import MutableMapping
from omegaconf import ListConfig

def flatten_dict(d, parent_key="", sep="_"):
    """Flattens a dictionary into a single-level dictionary while preserving
    parent keys. Taken from
    `SO <https://stackoverflow.com/questions/6027558/flatten-nested-dictionaries-compressing-keys>`_

    Args:
        d (MutableMapping): Dictionary to be flattened.
        parent_key (str): String to use as a prefix to all subsequent keys.
        sep (str): String to use as a separator between two key levels.

    Returns:
        dict: Single-level dictionary, flattened.
    """
    items = []
    for k, v in d.items():
        new_key = parent_key + sep + k if parent_key else k
        if isinstance(v, MutableMapping):
            items.extend(flatten_dict(v, new_key, sep=sep).items())
        else:
            items.append((new_key, v))
    return dict(items)


class AudioLightningModule(pl.LightningModule):
    def __init__(
        self,
        model=None,
        discriminator=None,
        optimizer=None,
        loss_func=None,
        metrics=None,
        scheduler=None,
    ):
        super().__init__()
        self.audio_model = model
        self.discriminator = discriminator
        self.optimizer = list(optimizer)
        self.loss_func = loss_func
        self.metrics = metrics
        self.scheduler = list(scheduler)
        
        # Save lightning"s AttributeDict under self.hparams
        self.default_monitor = "val_loss"
        # self.print(self.audio_model)
        self.validation_step_outputs = []
        self.test_step_outputs = []
        self.automatic_optimization = False

    def forward(self, wav):
        """Applies forward pass of the model.

        Returns:
            :class:`torch.Tensor`
        """
        return self.audio_model(wav)

    def training_step(self, batch, batch_nb):
        ori_data, codec_data = batch
        optimizer_g, optimizer_d = self.optimizers()
        # multiple schedulers
        scheduler_g, scheduler_d = self.lr_schedulers()
        
        # train discriminator
        optimizer_g.zero_grad()
        output = self(codec_data)
        
        optimizer_d.zero_grad()
        est_outputs, _ = self.discriminator(output.detach(), sample_rate=44100)
        target_outputs, _ = self.discriminator(ori_data, sample_rate=44100)
        
        loss_d = self.loss_func["d"](target_outputs, est_outputs)
        self.manual_backward(loss_d)
        self.clip_gradients(optimizer_d, gradient_clip_val=5, gradient_clip_algorithm="norm")
        optimizer_d.step()
        # train generator
        est_outputs, est_feature_maps = self.discriminator(output, sample_rate=44100)
        _, targets_feature_maps = self.discriminator(ori_data, sample_rate=44100)
        
        loss_g = self.loss_func["g"](est_outputs, est_feature_maps, targets_feature_maps, output, ori_data)
        self.manual_backward(loss_g)
        self.clip_gradients(optimizer_g, gradient_clip_val=5, gradient_clip_algorithm="norm")
        optimizer_g.step()
        # print(loss)
        
        if self.trainer.is_last_batch:
            scheduler_g.step()
            scheduler_d.step()

        self.log(
            "train_loss_d",
            loss_d,
            on_epoch=True,
            prog_bar=True,
            sync_dist=True,
            logger=True,
        )
        
        self.log(
            "train_loss_g",
            loss_g,
            on_epoch=True,
            prog_bar=True,
            sync_dist=True,
            logger=True,
        )


    def validation_step(self, batch, batch_nb):
        # cal val loss
        ori_data, codec_data = batch
        # print(mixtures.shape)
        est_sources = self(codec_data)
        loss = self.metrics(est_sources, ori_data)
        
        self.log(
            "val_loss",
            loss,
            on_epoch=True,
            prog_bar=True,
            sync_dist=True,
            logger=True,
        )
        
        self.validation_step_outputs.append(loss)
        
        return {"val_loss": loss}

    def on_validation_epoch_end(self):
        # val
        avg_loss = torch.stack(self.validation_step_outputs).mean()
        val_loss = torch.mean(self.all_gather(avg_loss))
        self.log(
            "lr",
            self.optimizer[0].param_groups[0]["lr"],
            on_epoch=True,
            prog_bar=True,
            sync_dist=True,
        )
        self.logger.experiment.log(
            {"learning_rate": self.optimizer[0].param_groups[0]["lr"], "epoch": self.current_epoch}
        )
        self.logger.experiment.log(
            {"val_pit_sisnr": -val_loss, "epoch": self.current_epoch}
        )

        self.validation_step_outputs.clear()  # free memory
        torch.cuda.empty_cache()
        
    def test_step(self, batch, batch_nb):
        mixtures, targets = batch
        est_sources = self(mixtures)
        loss = self.metrics(est_sources, targets)
        self.log(
            "test_loss",
            loss,
            on_epoch=True,
            prog_bar=True,
            sync_dist=True,
            logger=True,
        )
        self.test_step_outputs.append(loss)
        return {"test_loss": loss}
    
    def on_test_epoch_end(self):
        # val
        avg_loss = torch.stack(self.test_step_outputs).mean()
        test_loss = torch.mean(self.all_gather(avg_loss))
        self.log(
            "lr",
            self.optimizer.param_groups[0]["lr"],
            on_epoch=True,
            prog_bar=True,
            sync_dist=True,
        )
        self.logger.experiment.log(
            {"learning_rate": self.optimizer.param_groups[0]["lr"], "epoch": self.current_epoch}
        )
        self.logger.experiment.log(
            {"test_pit_sisnr": -test_loss, "epoch": self.current_epoch}
        )

        self.test_step_outputs.clear()

    def configure_optimizers(self):
        """Initialize optimizers, batch-wise and epoch-wise schedulers."""
        if self.scheduler is None:
            return self.optimizer
        if not isinstance(self.scheduler, (list, tuple)):
            self.scheduler = [self.scheduler]  # support multiple schedulers
            
        if not isinstance(self.optimizer, (list, tuple)):
            self.optimizer = [self.optimizer]  # support multiple schedulers
        
        epoch_schedulers = []
        for sched in self.scheduler:
            if not isinstance(sched, dict):
                if isinstance(sched, ReduceLROnPlateau):
                    sched = {"scheduler": sched, "monitor": self.default_monitor}
                epoch_schedulers.append(sched)
            else:
                sched.setdefault("monitor", self.default_monitor)
                sched.setdefault("frequency", 1)
                # Backward compat
                if sched["interval"] == "batch":
                    sched["interval"] = "step"
                assert sched["interval"] in [
                    "epoch",
                    "step",
                ], "Scheduler interval should be either step or epoch"
                epoch_schedulers.append(sched)
        return self.optimizer, epoch_schedulers
            
    @staticmethod
    def config_to_hparams(dic):
        """Sanitizes the config dict to be handled correctly by torch
        SummaryWriter. It flatten the config dict, converts ``None`` to
        ``"None"`` and any list and tuple into torch.Tensors.

        Args:
            dic (dict): Dictionary to be transformed.

        Returns:
            dict: Transformed dictionary.
        """
        dic = flatten_dict(dic)
        for k, v in dic.items():
            if v is None:
                dic[k] = str(v)
            elif isinstance(v, (list, tuple)):
                dic[k] = torch.tensor(v)
        return dic