File size: 7,885 Bytes
ac6acf2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from contextlib import contextmanager
from typing import Any, Dict, List, Optional, Tuple, Union

from comfy.ldm.modules.distributions.distributions import DiagonalGaussianDistribution

from comfy.ldm.util import instantiate_from_config
from comfy.ldm.modules.ema import LitEma
import comfy.ops

class DiagonalGaussianRegularizer(torch.nn.Module):
    def __init__(self, sample: bool = True):
        super().__init__()
        self.sample = sample

    def get_trainable_parameters(self) -> Any:
        yield from ()

    def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]:
        log = dict()
        posterior = DiagonalGaussianDistribution(z)
        if self.sample:
            z = posterior.sample()
        else:
            z = posterior.mode()
        kl_loss = posterior.kl()
        kl_loss = torch.sum(kl_loss) / kl_loss.shape[0]
        log["kl_loss"] = kl_loss
        return z, log


class AbstractAutoencoder(torch.nn.Module):
    """

    This is the base class for all autoencoders, including image autoencoders, image autoencoders with discriminators,

    unCLIP models, etc. Hence, it is fairly general, and specific features

    (e.g. discriminator training, encoding, decoding) must be implemented in subclasses.

    """

    def __init__(

        self,

        ema_decay: Union[None, float] = None,

        monitor: Union[None, str] = None,

        input_key: str = "jpg",

        **kwargs,

    ):
        super().__init__()

        self.input_key = input_key
        self.use_ema = ema_decay is not None
        if monitor is not None:
            self.monitor = monitor

        if self.use_ema:
            self.model_ema = LitEma(self, decay=ema_decay)
            logpy.info(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")

    def get_input(self, batch) -> Any:
        raise NotImplementedError()

    def on_train_batch_end(self, *args, **kwargs):
        # for EMA computation
        if self.use_ema:
            self.model_ema(self)

    @contextmanager
    def ema_scope(self, context=None):
        if self.use_ema:
            self.model_ema.store(self.parameters())
            self.model_ema.copy_to(self)
            if context is not None:
                logpy.info(f"{context}: Switched to EMA weights")
        try:
            yield None
        finally:
            if self.use_ema:
                self.model_ema.restore(self.parameters())
                if context is not None:
                    logpy.info(f"{context}: Restored training weights")

    def encode(self, *args, **kwargs) -> torch.Tensor:
        raise NotImplementedError("encode()-method of abstract base class called")

    def decode(self, *args, **kwargs) -> torch.Tensor:
        raise NotImplementedError("decode()-method of abstract base class called")

    def instantiate_optimizer_from_config(self, params, lr, cfg):
        logpy.info(f"loading >>> {cfg['target']} <<< optimizer from config")
        return get_obj_from_str(cfg["target"])(
            params, lr=lr, **cfg.get("params", dict())
        )

    def configure_optimizers(self) -> Any:
        raise NotImplementedError()


class AutoencodingEngine(AbstractAutoencoder):
    """

    Base class for all image autoencoders that we train, like VQGAN or AutoencoderKL

    (we also restore them explicitly as special cases for legacy reasons).

    Regularizations such as KL or VQ are moved to the regularizer class.

    """

    def __init__(

        self,

        *args,

        encoder_config: Dict,

        decoder_config: Dict,

        regularizer_config: Dict,

        **kwargs,

    ):
        super().__init__(*args, **kwargs)

        self.encoder: torch.nn.Module = instantiate_from_config(encoder_config)
        self.decoder: torch.nn.Module = instantiate_from_config(decoder_config)
        self.regularization: AbstractRegularizer = instantiate_from_config(
            regularizer_config
        )

    def get_last_layer(self):
        return self.decoder.get_last_layer()

    def encode(

        self,

        x: torch.Tensor,

        return_reg_log: bool = False,

        unregularized: bool = False,

    ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]:
        z = self.encoder(x)
        if unregularized:
            return z, dict()
        z, reg_log = self.regularization(z)
        if return_reg_log:
            return z, reg_log
        return z

    def decode(self, z: torch.Tensor, **kwargs) -> torch.Tensor:
        x = self.decoder(z, **kwargs)
        return x

    def forward(

        self, x: torch.Tensor, **additional_decode_kwargs

    ) -> Tuple[torch.Tensor, torch.Tensor, dict]:
        z, reg_log = self.encode(x, return_reg_log=True)
        dec = self.decode(z, **additional_decode_kwargs)
        return z, dec, reg_log


class AutoencodingEngineLegacy(AutoencodingEngine):
    def __init__(self, embed_dim: int, **kwargs):
        self.max_batch_size = kwargs.pop("max_batch_size", None)
        ddconfig = kwargs.pop("ddconfig")
        super().__init__(
            encoder_config={
                "target": "comfy.ldm.modules.diffusionmodules.model.Encoder",
                "params": ddconfig,
            },
            decoder_config={
                "target": "comfy.ldm.modules.diffusionmodules.model.Decoder",
                "params": ddconfig,
            },
            **kwargs,
        )
        self.quant_conv = comfy.ops.disable_weight_init.Conv2d(
            (1 + ddconfig["double_z"]) * ddconfig["z_channels"],
            (1 + ddconfig["double_z"]) * embed_dim,
            1,
        )
        self.post_quant_conv = comfy.ops.disable_weight_init.Conv2d(embed_dim, ddconfig["z_channels"], 1)
        self.embed_dim = embed_dim

    def get_autoencoder_params(self) -> list:
        params = super().get_autoencoder_params()
        return params

    def encode(

        self, x: torch.Tensor, return_reg_log: bool = False

    ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]:
        if self.max_batch_size is None:
            z = self.encoder(x)
            z = self.quant_conv(z)
        else:
            N = x.shape[0]
            bs = self.max_batch_size
            n_batches = int(math.ceil(N / bs))
            z = list()
            for i_batch in range(n_batches):
                z_batch = self.encoder(x[i_batch * bs : (i_batch + 1) * bs])
                z_batch = self.quant_conv(z_batch)
                z.append(z_batch)
            z = torch.cat(z, 0)

        z, reg_log = self.regularization(z)
        if return_reg_log:
            return z, reg_log
        return z

    def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor:
        if self.max_batch_size is None:
            dec = self.post_quant_conv(z)
            dec = self.decoder(dec, **decoder_kwargs)
        else:
            N = z.shape[0]
            bs = self.max_batch_size
            n_batches = int(math.ceil(N / bs))
            dec = list()
            for i_batch in range(n_batches):
                dec_batch = self.post_quant_conv(z[i_batch * bs : (i_batch + 1) * bs])
                dec_batch = self.decoder(dec_batch, **decoder_kwargs)
                dec.append(dec_batch)
            dec = torch.cat(dec, 0)

        return dec


class AutoencoderKL(AutoencodingEngineLegacy):
    def __init__(self, **kwargs):
        if "lossconfig" in kwargs:
            kwargs["loss_config"] = kwargs.pop("lossconfig")
        super().__init__(
            regularizer_config={
                "target": (
                    "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"
                )
            },
            **kwargs,
        )