Spaces:
Runtime error
Runtime error
File size: 10,394 Bytes
3424266 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 |
# Copyright 2024 EPFL and Apple Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Union, Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from diffusers.models import ControlNetModel
from diffusers.configuration_utils import ConfigMixin
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.controlnet import zero_module
from fourm.utils import to_2tuple
from .lm_models import create_model
class ControlNetAdapterEmbedding(nn.Module):
def __init__(
self,
conditioning_embedding_channels,
adapter,
conditioning_channels=3,
):
super().__init__()
self.adapter_model = create_model(
in_channels=conditioning_channels,
output_type="stats",
)
self._load_adapter(adapter)
self.conv_out = zero_module(
nn.Conv2d(8, conditioning_embedding_channels, kernel_size=3, padding=1)
)
def forward(self, conditioning):
embedding = self.adapter_model(quant=conditioning)
embedding = self.conv_out(embedding)
return embedding
def _load_adapter(self, path):
ckpt = torch.load(path)['model']
for key in list(ckpt.keys()):
if 'vq_model' in key or 'vae' in key:
del ckpt[key]
self.adapter_model.load_state_dict(ckpt)
print("Loaded the adapter model")
class ControlNetConditioningEmbedding(nn.Module):
def __init__(
self,
conditioning_embedding_channels,
conditioning_channels = 3,
block_out_channels = (16, 32, 96, 256),
):
super().__init__()
self.conv_in = nn.Conv2d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
self.blocks = nn.ModuleList([])
for i in range(len(block_out_channels) - 1):
channel_in = block_out_channels[i]
channel_out = block_out_channels[i + 1]
self.blocks.append(nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1))
self.blocks.append(nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1))
self.conv_out = zero_module(
nn.Conv2d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
)
def forward(self, conditioning):
embedding = self.conv_in(conditioning)
embedding = F.silu(embedding)
for block in self.blocks:
embedding = block(embedding)
embedding = F.silu(embedding)
embedding = self.conv_out(embedding)
return embedding
class ControlnetCond(ModelMixin, ConfigMixin):
def __init__(self,
in_channels,
cond_channels,
sd_pipeline,
image_size,
freeze_params=True,
block_out_channels = (320, 640, 1280, 1280),
conditioning_embedding_out_channels = (32, 32, 96, 256),
pretrained_cn=False,
enable_xformer=False,
adapter=None,
*args,
**kwargs
):
super().__init__()
self.in_channels = in_channels
self.cond_channels = cond_channels
self.sd_pipeline = sd_pipeline
self.unet = sd_pipeline.unet
self.text_encoder = sd_pipeline.text_encoder
self.tokenizer = sd_pipeline.tokenizer
if pretrained_cn:
self.controlnet = ControlNetModel.from_unet(self.unet, conditioning_embedding_out_channels=conditioning_embedding_out_channels)
self.controlnet.conditioning_channels = cond_channels
self.controlnet.config.conditioning_channels = cond_channels
else:
self.controlnet = ControlNetModel(
in_channels=in_channels,
conditioning_channels=cond_channels,
block_out_channels=block_out_channels,
conditioning_embedding_out_channels=conditioning_embedding_out_channels,
*args,
**kwargs,
)
self.use_adapter = adapter is not None
if adapter is not None:
self.controlnet.controlnet_cond_embedding = ControlNetAdapterEmbedding(
conditioning_embedding_channels=self.controlnet.config.block_out_channels[0],
adapter=adapter,
conditioning_channels=cond_channels,
)
else:
self.controlnet.controlnet_cond_embedding = ControlNetConditioningEmbedding(
conditioning_embedding_channels=self.controlnet.config.block_out_channels[0],
block_out_channels=self.controlnet.config.conditioning_embedding_out_channels,
conditioning_channels=cond_channels,
)
if enable_xformer:
print('xFormer enabled')
self.unet.enable_xformers_memory_efficient_attention()
self.controlnet.enable_xformers_memory_efficient_attention()
self.empty_str_encoding = nn.Parameter(self._encode_prompt(""), requires_grad=False)
if freeze_params:
self.freeze_params()
self.sample_size = image_size // sd_pipeline.vae_scale_factor
self.H, self.W = to_2tuple(self.sample_size)
def forward(self,
sample: torch.FloatTensor, # Shape (B, C, H, W),
timestep: Union[torch.Tensor, float, int],
encoder_hidden_states: torch.Tensor = None, # Shape (B, D_C, H_C, W_C)
cond_mask: Optional[torch.BoolTensor] = None, # Boolen tensor of shape (B, H_C, W_C). True for masked out pixels,
prompt = None,
unconditional = False,
cond_scale = 1.0,
**kwargs):
# Optionally mask out conditioning
if cond_mask is not None:
encoder_hidden_states = torch.where(cond_mask[:,None,:,:], 0.0, encoder_hidden_states)
if not self.use_adapter:
controlnet_cond = F.interpolate(encoder_hidden_states, (self.H, self.W), mode="nearest")
else:
controlnet_cond = F.interpolate(encoder_hidden_states, (self.H // 2, self.W // 2), mode="nearest")
# encoder_hidden_states is the propmp embedding in the controlnet model, for now it's set to zeros.
if prompt is None or unconditional:
encoder_hidden_states = torch.cat([self.empty_str_encoding] * sample.shape[0])
else:
encoder_hidden_states = self._encode_prompt(prompt)
down_block_res_samples, mid_block_res_sample = self.controlnet(
sample,
timestep,
encoder_hidden_states=encoder_hidden_states,
controlnet_cond=controlnet_cond,
conditioning_scale=cond_scale,
return_dict=False,
)
# TODO not the most efficient way
if unconditional:
down_block_res_samples = [torch.zeros_like(s) for s in down_block_res_samples]
controlnet_cond = torch.zeros_like(controlnet_cond)
noise_pred = self.unet(
sample,
timestep,
encoder_hidden_states=encoder_hidden_states,
down_block_additional_residuals=down_block_res_samples,
mid_block_additional_residual=mid_block_res_sample,
return_dict=False,
)[0]
return noise_pred
def freeze_params(self):
for param in self.unet.parameters():
param.requires_grad = False
for param in self.text_encoder.parameters():
param.requires_grad = False
def unfreeze_params(self):
for param in self.unet.parameters():
param.requires_grad = True
for param in self.text_encoder.parameters():
param.requires_grad = True
@torch.no_grad()
def _encode_prompt(self, prompt):
text_inputs = self.tokenizer(
prompt,
padding="max_length",
max_length=self.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
attention_mask = text_inputs.attention_mask.to(self.device)
else:
attention_mask = None
prompt_embeds = self.text_encoder(
text_input_ids.to(self.device),
attention_mask=attention_mask,
)[0]
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=self.device)
return prompt_embeds
def controlnet(*args, **kwargs):
return ControlnetCond(
flip_sin_to_cos=True,
freq_shift=0,
down_block_types=
['CrossAttnDownBlock2D',
'CrossAttnDownBlock2D',
'CrossAttnDownBlock2D',
'DownBlock2D'],
only_cross_attention=False,
block_out_channels=[320, 640, 1280, 1280],
layers_per_block=2,
downsample_padding=1,
mid_block_scale_factor=1,
act_fn='silu',
norm_num_groups=32,
norm_eps=1e-05,
cross_attention_dim=768,
attention_head_dim=8,
num_attention_heads=None,
use_linear_projection=False,
class_embed_type=None,
num_class_embeds=None,
upcast_attention=False,
resnet_time_scale_shift='default',
projection_class_embeddings_input_dim=None,
controlnet_conditioning_channel_order='rgb',
conditioning_embedding_out_channels=[kwargs['cond_channels'], 32, 96, 256],
global_pool_conditions=False,
freeze_params=True,
*args,
**kwargs,
) |