File size: 6,840 Bytes
3ed0796 |
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 |
import torch
from torch.utils.checkpoint import checkpoint
from diffusers.models.autoencoders.autoencoder_kl import Encoder, Decoder, AutoencoderKL
from typing import Optional
class MyEncoder(Encoder):
def __init__(
self,
in_channels = 3,
out_channels = 3,
down_block_types = ...,
block_out_channels = ...,
layers_per_block = 2,
norm_num_groups = 32,
act_fn = "silu",
double_z = True,
mid_block_add_attention=True
):
super().__init__(
in_channels, out_channels, down_block_types, block_out_channels,
layers_per_block, norm_num_groups, act_fn, double_z, mid_block_add_attention
)
def forward(self, sample: torch.Tensor) -> torch.Tensor:
r"""The forward method of the `Encoder` class."""
sample = self.conv_in(sample)
if torch.is_grad_enabled() and self.gradient_checkpointing:
# down
for down_block in self.down_blocks:
sample = checkpoint(self.ckpt_wrapper(down_block), sample)
# middle
sample = checkpoint(self.ckpt_wrapper(self.mid_block), sample)
else:
# down
for down_block in self.down_blocks:
sample = down_block(sample)
# middle
sample = self.mid_block(sample)
# post-process
sample = self.conv_norm_out(sample)
sample = self.conv_act(sample)
sample = self.conv_out(sample)
return sample
def ckpt_wrapper(self, module):
def ckpt_forward(*inputs):
outputs = module(*inputs)
return outputs
return ckpt_forward
class MyDecoder(Decoder):
def __init__(
self,
in_channels = 3,
out_channels = 3,
up_block_types = ...,
block_out_channels = ...,
layers_per_block = 2,
norm_num_groups = 32,
act_fn = "silu",
norm_type = "group",
mid_block_add_attention=True
):
super().__init__(
in_channels, out_channels, up_block_types, block_out_channels,
layers_per_block, norm_num_groups, act_fn, norm_type, mid_block_add_attention
)
def forward(
self,
sample: torch.Tensor,
latent_embeds: Optional[torch.Tensor] = None,
) -> torch.Tensor:
r"""The forward method of the `Decoder` class."""
sample = self.conv_in(sample)
upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
if torch.is_grad_enabled() and self.gradient_checkpointing:
# middle
sample = checkpoint(self.ckpt_wrapper(self.mid_block), sample, latent_embeds)
sample = sample.to(upscale_dtype)
# up
for up_block in self.up_blocks:
sample = checkpoint(self.ckpt_wrapper(up_block), sample, latent_embeds)
else:
# middle
sample = self.mid_block(sample, latent_embeds)
sample = sample.to(upscale_dtype)
# up
for up_block in self.up_blocks:
sample = up_block(sample, latent_embeds)
# post-process
if latent_embeds is None:
sample = self.conv_norm_out(sample)
else:
sample = self.conv_norm_out(sample, latent_embeds)
sample = self.conv_act(sample)
sample = self.conv_out(sample)
return sample
def ckpt_wrapper(self, module):
def ckpt_forward(*inputs):
outputs = module(*inputs)
return outputs
return ckpt_forward
class MyAutoencoderKL(AutoencoderKL):
def __init__(
self,
in_channels = 3,
out_channels = 3,
down_block_types = ...,
up_block_types = ...,
block_out_channels = ...,
layers_per_block = 1,
act_fn = "silu",
latent_channels = 4,
norm_num_groups = 32,
sample_size = 32,
scaling_factor = 0.18215,
shift_factor = None,
latents_mean = None,
latents_std = None,
force_upcast = True,
use_quant_conv = True,
use_post_quant_conv = True,
mid_block_add_attention = True,
bn_momentum = 0.1,
):
super().__init__(
in_channels, out_channels, down_block_types, up_block_types, block_out_channels,
layers_per_block, act_fn, latent_channels, norm_num_groups, sample_size,
scaling_factor, shift_factor, latents_mean, latents_std, force_upcast,
use_quant_conv, use_post_quant_conv, mid_block_add_attention
)
self.encoder = MyEncoder(
in_channels=in_channels,
out_channels=latent_channels,
down_block_types=down_block_types,
block_out_channels=block_out_channels,
layers_per_block=layers_per_block,
act_fn=act_fn,
norm_num_groups=norm_num_groups,
double_z=True,
mid_block_add_attention=mid_block_add_attention,
)
# pass init params to Decoder
self.decoder = MyDecoder(
in_channels=latent_channels,
out_channels=out_channels,
up_block_types=up_block_types,
block_out_channels=block_out_channels,
layers_per_block=layers_per_block,
norm_num_groups=norm_num_groups,
act_fn=act_fn,
mid_block_add_attention=mid_block_add_attention,
)
self.bn = torch.nn.BatchNorm2d(
latent_channels, eps=1e-4, momentum=bn_momentum, affine=False, track_running_stats=True
)
self.bn.reset_running_stats()
self.init_bn()
def init_bn(self):
# self.bn.running_mean = torch.zeros_like(self.bn.running_mean).to(torch.float64)
# self.bn.running_var = torch.ones_like(self.bn.running_var).to(torch.float64) / self.config.scaling_factor ** 2
self.bn.running_mean = torch.zeros_like(self.bn.running_mean)
self.bn.running_var = torch.ones_like(self.bn.running_var) / self.config.scaling_factor ** 2
@property
def mean(self):
mean = self.bn.running_mean.unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
return mean
@property
def std(self):
std = self.bn.running_var.sqrt().unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
return std
def forward(self, x: torch.Tensor, use_checkpoint=False):
self.encoder.gradient_checkpointing = use_checkpoint
self.decoder.gradient_checkpointing = use_checkpoint
posterior = self.encode(x).latent_dist
z = posterior.sample()
latent = self.bn(z)
recon = self.decode(z).sample
return posterior, latent, recon
|