Spaces:
Running
Running
File size: 7,803 Bytes
891b88f |
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 |
import torch as th
import torch.nn as nn
import torch.nn.functional as F
from .nn import timestep_embedding
from .unet import UNetModel
from .xf import LayerNorm, Transformer, convert_module_to_f16
class Text2ImUNet(UNetModel):
"""
A UNetModel that conditions on text with an encoding transformer.
Expects an extra kwarg `tokens` of text.
:param text_ctx: number of text tokens to expect.
:param xf_width: width of the transformer.
:param xf_layers: depth of the transformer.
:param xf_heads: heads in the transformer.
:param xf_final_ln: use a LayerNorm after the output layer.
:param tokenizer: the text tokenizer for sampling/vocab size.
"""
def __init__(
self,
text_ctx,
xf_width,
xf_layers,
xf_heads,
xf_final_ln,
tokenizer,
*args,
cache_text_emb=False,
xf_ar=0.0,
xf_padding=False,
share_unemb=False,
**kwargs,
):
self.text_ctx = text_ctx
self.xf_width = xf_width
self.xf_ar = xf_ar
self.xf_padding = xf_padding
self.tokenizer = tokenizer
if not xf_width:
super().__init__(*args, **kwargs, encoder_channels=None)
else:
super().__init__(*args, **kwargs, encoder_channels=xf_width)
if self.xf_width:
self.transformer = Transformer(
text_ctx,
xf_width,
xf_layers,
xf_heads,
)
if xf_final_ln:
self.final_ln = LayerNorm(xf_width)
else:
self.final_ln = None
self.token_embedding = nn.Embedding(self.tokenizer.n_vocab, xf_width)
self.positional_embedding = nn.Parameter(th.empty(text_ctx, xf_width, dtype=th.float32))
self.transformer_proj = nn.Linear(xf_width, self.model_channels * 4)
if self.xf_padding:
self.padding_embedding = nn.Parameter(
th.empty(text_ctx, xf_width, dtype=th.float32)
)
if self.xf_ar:
self.unemb = nn.Linear(xf_width, self.tokenizer.n_vocab)
if share_unemb:
self.unemb.weight = self.token_embedding.weight
self.cache_text_emb = cache_text_emb
self.cache = None
def convert_to_fp16(self):
super().convert_to_fp16()
if self.xf_width:
self.transformer.apply(convert_module_to_f16)
self.transformer_proj.to(th.float16)
self.token_embedding.to(th.float16)
self.positional_embedding.to(th.float16)
if self.xf_padding:
self.padding_embedding.to(th.float16)
if self.xf_ar:
self.unemb.to(th.float16)
def get_text_emb(self, tokens, mask):
assert tokens is not None
if self.cache_text_emb and self.cache is not None:
assert (
tokens == self.cache["tokens"]
).all(), f"Tokens {tokens.cpu().numpy().tolist()} do not match cache {self.cache['tokens'].cpu().numpy().tolist()}"
return self.cache
xf_in = self.token_embedding(tokens.long())
xf_in = xf_in + self.positional_embedding[None]
if self.xf_padding:
assert mask is not None
xf_in = th.where(mask[..., None], xf_in, self.padding_embedding[None])
xf_out = self.transformer(xf_in.to(self.dtype))
if self.final_ln is not None:
xf_out = self.final_ln(xf_out)
xf_proj = self.transformer_proj(xf_out[:, -1])
xf_out = xf_out.permute(0, 2, 1) # NLC -> NCL
outputs = dict(xf_proj=xf_proj, xf_out=xf_out)
if self.cache_text_emb:
self.cache = dict(
tokens=tokens,
xf_proj=xf_proj.detach(),
xf_out=xf_out.detach() if xf_out is not None else None,
)
return outputs
def del_cache(self):
self.cache = None
def forward(self, x, timesteps, tokens=None, mask=None):
hs = []
emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
if self.xf_width:
text_outputs = self.get_text_emb(tokens, mask)
xf_proj, xf_out = text_outputs["xf_proj"], text_outputs["xf_out"]
emb = emb + xf_proj.to(emb)
else:
xf_out = None
h = x.type(self.dtype)
for module in self.input_blocks:
h = module(h, emb, xf_out)
hs.append(h)
h = self.middle_block(h, emb, xf_out)
for module in self.output_blocks:
h = th.cat([h, hs.pop()], dim=1)
h = module(h, emb, xf_out)
h = h.type(x.dtype)
h = self.out(h)
return h
class SuperResText2ImUNet(Text2ImUNet):
"""
A text2im model that performs super-resolution.
Expects an extra kwarg `low_res` to condition on a low-resolution image.
"""
def __init__(self, *args, **kwargs):
if "in_channels" in kwargs:
kwargs = dict(kwargs)
kwargs["in_channels"] = kwargs["in_channels"] * 2
else:
# Curse you, Python. Or really, just curse positional arguments :|.
args = list(args)
args[1] = args[1] * 2
super().__init__(*args, **kwargs)
def forward(self, x, timesteps, low_res=None, **kwargs):
_, _, new_height, new_width = x.shape
upsampled = F.interpolate(
low_res, (new_height, new_width), mode="bilinear", align_corners=False
)
x = th.cat([x, upsampled], dim=1)
return super().forward(x, timesteps, **kwargs)
class InpaintText2ImUNet(Text2ImUNet):
"""
A text2im model which can perform inpainting.
"""
def __init__(self, *args, **kwargs):
if "in_channels" in kwargs:
kwargs = dict(kwargs)
kwargs["in_channels"] = kwargs["in_channels"] * 2 + 1
else:
# Curse you, Python. Or really, just curse positional arguments :|.
args = list(args)
args[1] = args[1] * 2 + 1
super().__init__(*args, **kwargs)
def forward(self, x, timesteps, inpaint_image=None, inpaint_mask=None, **kwargs):
if inpaint_image is None:
inpaint_image = th.zeros_like(x)
if inpaint_mask is None:
inpaint_mask = th.zeros_like(x[:, :1])
return super().forward(
th.cat([x, inpaint_image * inpaint_mask, inpaint_mask], dim=1),
timesteps,
**kwargs,
)
class SuperResInpaintText2ImUnet(Text2ImUNet):
"""
A text2im model which can perform both upsampling and inpainting.
"""
def __init__(self, *args, **kwargs):
if "in_channels" in kwargs:
kwargs = dict(kwargs)
kwargs["in_channels"] = kwargs["in_channels"] * 3 + 1
else:
# Curse you, Python. Or really, just curse positional arguments :|.
args = list(args)
args[1] = args[1] * 3 + 1
super().__init__(*args, **kwargs)
def forward(
self,
x,
timesteps,
inpaint_image=None,
inpaint_mask=None,
low_res=None,
**kwargs,
):
if inpaint_image is None:
inpaint_image = th.zeros_like(x)
if inpaint_mask is None:
inpaint_mask = th.zeros_like(x[:, :1])
_, _, new_height, new_width = x.shape
upsampled = F.interpolate(
low_res, (new_height, new_width), mode="bilinear", align_corners=False
)
return super().forward(
th.cat([x, inpaint_image * inpaint_mask, inpaint_mask, upsampled], dim=1),
timesteps,
**kwargs,
)
|