Dragunflie-420
commited on
Commit
•
288a8da
1
Parent(s):
73b4102
Create models.py
Browse files
models.py
ADDED
@@ -0,0 +1,370 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
# --------------------------------------------------------
|
7 |
+
# References:
|
8 |
+
# GLIDE: https://github.com/openai/glide-text2im
|
9 |
+
# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py
|
10 |
+
# --------------------------------------------------------
|
11 |
+
|
12 |
+
import torch
|
13 |
+
import torch.nn as nn
|
14 |
+
import numpy as np
|
15 |
+
import math
|
16 |
+
from timm.models.vision_transformer import PatchEmbed, Attention, Mlp
|
17 |
+
|
18 |
+
|
19 |
+
def modulate(x, shift, scale):
|
20 |
+
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
|
21 |
+
|
22 |
+
|
23 |
+
#################################################################################
|
24 |
+
# Embedding Layers for Timesteps and Class Labels #
|
25 |
+
#################################################################################
|
26 |
+
|
27 |
+
class TimestepEmbedder(nn.Module):
|
28 |
+
"""
|
29 |
+
Embeds scalar timesteps into vector representations.
|
30 |
+
"""
|
31 |
+
def __init__(self, hidden_size, frequency_embedding_size=256):
|
32 |
+
super().__init__()
|
33 |
+
self.mlp = nn.Sequential(
|
34 |
+
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
|
35 |
+
nn.SiLU(),
|
36 |
+
nn.Linear(hidden_size, hidden_size, bias=True),
|
37 |
+
)
|
38 |
+
self.frequency_embedding_size = frequency_embedding_size
|
39 |
+
|
40 |
+
@staticmethod
|
41 |
+
def timestep_embedding(t, dim, max_period=10000):
|
42 |
+
"""
|
43 |
+
Create sinusoidal timestep embeddings.
|
44 |
+
:param t: a 1-D Tensor of N indices, one per batch element.
|
45 |
+
These may be fractional.
|
46 |
+
:param dim: the dimension of the output.
|
47 |
+
:param max_period: controls the minimum frequency of the embeddings.
|
48 |
+
:return: an (N, D) Tensor of positional embeddings.
|
49 |
+
"""
|
50 |
+
# https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
|
51 |
+
half = dim // 2
|
52 |
+
freqs = torch.exp(
|
53 |
+
-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
|
54 |
+
).to(device=t.device)
|
55 |
+
args = t[:, None].float() * freqs[None]
|
56 |
+
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
57 |
+
if dim % 2:
|
58 |
+
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
59 |
+
return embedding
|
60 |
+
|
61 |
+
def forward(self, t):
|
62 |
+
t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
|
63 |
+
t_emb = self.mlp(t_freq)
|
64 |
+
return t_emb
|
65 |
+
|
66 |
+
|
67 |
+
class LabelEmbedder(nn.Module):
|
68 |
+
"""
|
69 |
+
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
|
70 |
+
"""
|
71 |
+
def __init__(self, num_classes, hidden_size, dropout_prob):
|
72 |
+
super().__init__()
|
73 |
+
use_cfg_embedding = dropout_prob > 0
|
74 |
+
self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size)
|
75 |
+
self.num_classes = num_classes
|
76 |
+
self.dropout_prob = dropout_prob
|
77 |
+
|
78 |
+
def token_drop(self, labels, force_drop_ids=None):
|
79 |
+
"""
|
80 |
+
Drops labels to enable classifier-free guidance.
|
81 |
+
"""
|
82 |
+
if force_drop_ids is None:
|
83 |
+
drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
|
84 |
+
else:
|
85 |
+
drop_ids = force_drop_ids == 1
|
86 |
+
labels = torch.where(drop_ids, self.num_classes, labels)
|
87 |
+
return labels
|
88 |
+
|
89 |
+
def forward(self, labels, train, force_drop_ids=None):
|
90 |
+
use_dropout = self.dropout_prob > 0
|
91 |
+
if (train and use_dropout) or (force_drop_ids is not None):
|
92 |
+
labels = self.token_drop(labels, force_drop_ids)
|
93 |
+
embeddings = self.embedding_table(labels)
|
94 |
+
return embeddings
|
95 |
+
|
96 |
+
|
97 |
+
#################################################################################
|
98 |
+
# Core DiT Model #
|
99 |
+
#################################################################################
|
100 |
+
|
101 |
+
class DiTBlock(nn.Module):
|
102 |
+
"""
|
103 |
+
A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning.
|
104 |
+
"""
|
105 |
+
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, **block_kwargs):
|
106 |
+
super().__init__()
|
107 |
+
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
108 |
+
self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, **block_kwargs)
|
109 |
+
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
110 |
+
mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
111 |
+
approx_gelu = lambda: nn.GELU(approximate="tanh")
|
112 |
+
self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)
|
113 |
+
self.adaLN_modulation = nn.Sequential(
|
114 |
+
nn.SiLU(),
|
115 |
+
nn.Linear(hidden_size, 6 * hidden_size, bias=True)
|
116 |
+
)
|
117 |
+
|
118 |
+
def forward(self, x, c):
|
119 |
+
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)
|
120 |
+
x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))
|
121 |
+
x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
|
122 |
+
return x
|
123 |
+
|
124 |
+
|
125 |
+
class FinalLayer(nn.Module):
|
126 |
+
"""
|
127 |
+
The final layer of DiT.
|
128 |
+
"""
|
129 |
+
def __init__(self, hidden_size, patch_size, out_channels):
|
130 |
+
super().__init__()
|
131 |
+
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
132 |
+
self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
|
133 |
+
self.adaLN_modulation = nn.Sequential(
|
134 |
+
nn.SiLU(),
|
135 |
+
nn.Linear(hidden_size, 2 * hidden_size, bias=True)
|
136 |
+
)
|
137 |
+
|
138 |
+
def forward(self, x, c):
|
139 |
+
shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
|
140 |
+
x = modulate(self.norm_final(x), shift, scale)
|
141 |
+
x = self.linear(x)
|
142 |
+
return x
|
143 |
+
|
144 |
+
|
145 |
+
class DiT(nn.Module):
|
146 |
+
"""
|
147 |
+
Diffusion model with a Transformer backbone.
|
148 |
+
"""
|
149 |
+
def __init__(
|
150 |
+
self,
|
151 |
+
input_size=32,
|
152 |
+
patch_size=2,
|
153 |
+
in_channels=4,
|
154 |
+
hidden_size=1152,
|
155 |
+
depth=28,
|
156 |
+
num_heads=16,
|
157 |
+
mlp_ratio=4.0,
|
158 |
+
class_dropout_prob=0.1,
|
159 |
+
num_classes=1000,
|
160 |
+
learn_sigma=True,
|
161 |
+
):
|
162 |
+
super().__init__()
|
163 |
+
self.learn_sigma = learn_sigma
|
164 |
+
self.in_channels = in_channels
|
165 |
+
self.out_channels = in_channels * 2 if learn_sigma else in_channels
|
166 |
+
self.patch_size = patch_size
|
167 |
+
self.num_heads = num_heads
|
168 |
+
|
169 |
+
self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
|
170 |
+
self.t_embedder = TimestepEmbedder(hidden_size)
|
171 |
+
self.y_embedder = LabelEmbedder(num_classes, hidden_size, class_dropout_prob)
|
172 |
+
num_patches = self.x_embedder.num_patches
|
173 |
+
# Will use fixed sin-cos embedding:
|
174 |
+
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, hidden_size), requires_grad=False)
|
175 |
+
|
176 |
+
self.blocks = nn.ModuleList([
|
177 |
+
DiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(depth)
|
178 |
+
])
|
179 |
+
self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels)
|
180 |
+
self.initialize_weights()
|
181 |
+
|
182 |
+
def initialize_weights(self):
|
183 |
+
# Initialize transformer layers:
|
184 |
+
def _basic_init(module):
|
185 |
+
if isinstance(module, nn.Linear):
|
186 |
+
torch.nn.init.xavier_uniform_(module.weight)
|
187 |
+
if module.bias is not None:
|
188 |
+
nn.init.constant_(module.bias, 0)
|
189 |
+
self.apply(_basic_init)
|
190 |
+
|
191 |
+
# Initialize (and freeze) pos_embed by sin-cos embedding:
|
192 |
+
pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches ** 0.5))
|
193 |
+
self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
|
194 |
+
|
195 |
+
# Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
|
196 |
+
w = self.x_embedder.proj.weight.data
|
197 |
+
nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
|
198 |
+
nn.init.constant_(self.x_embedder.proj.bias, 0)
|
199 |
+
|
200 |
+
# Initialize label embedding table:
|
201 |
+
nn.init.normal_(self.y_embedder.embedding_table.weight, std=0.02)
|
202 |
+
|
203 |
+
# Initialize timestep embedding MLP:
|
204 |
+
nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
|
205 |
+
nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
|
206 |
+
|
207 |
+
# Zero-out adaLN modulation layers in DiT blocks:
|
208 |
+
for block in self.blocks:
|
209 |
+
nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
|
210 |
+
nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
|
211 |
+
|
212 |
+
# Zero-out output layers:
|
213 |
+
nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
|
214 |
+
nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
|
215 |
+
nn.init.constant_(self.final_layer.linear.weight, 0)
|
216 |
+
nn.init.constant_(self.final_layer.linear.bias, 0)
|
217 |
+
|
218 |
+
def unpatchify(self, x):
|
219 |
+
"""
|
220 |
+
x: (N, T, patch_size**2 * C)
|
221 |
+
imgs: (N, H, W, C)
|
222 |
+
"""
|
223 |
+
c = self.out_channels
|
224 |
+
p = self.x_embedder.patch_size[0]
|
225 |
+
h = w = int(x.shape[1] ** 0.5)
|
226 |
+
assert h * w == x.shape[1]
|
227 |
+
|
228 |
+
x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
|
229 |
+
x = torch.einsum('nhwpqc->nchpwq', x)
|
230 |
+
imgs = x.reshape(shape=(x.shape[0], c, h * p, h * p))
|
231 |
+
return imgs
|
232 |
+
|
233 |
+
def forward(self, x, t, y):
|
234 |
+
"""
|
235 |
+
Forward pass of DiT.
|
236 |
+
x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
|
237 |
+
t: (N,) tensor of diffusion timesteps
|
238 |
+
y: (N,) tensor of class labels
|
239 |
+
"""
|
240 |
+
x = self.x_embedder(x) + self.pos_embed # (N, T, D), where T = H * W / patch_size ** 2
|
241 |
+
t = self.t_embedder(t) # (N, D)
|
242 |
+
y = self.y_embedder(y, self.training) # (N, D)
|
243 |
+
c = t + y # (N, D)
|
244 |
+
for block in self.blocks:
|
245 |
+
x = block(x, c) # (N, T, D)
|
246 |
+
x = self.final_layer(x, c) # (N, T, patch_size ** 2 * out_channels)
|
247 |
+
x = self.unpatchify(x) # (N, out_channels, H, W)
|
248 |
+
return x
|
249 |
+
|
250 |
+
def forward_with_cfg(self, x, t, y, cfg_scale):
|
251 |
+
"""
|
252 |
+
Forward pass of DiT, but also batches the unconditional forward pass for classifier-free guidance.
|
253 |
+
"""
|
254 |
+
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
|
255 |
+
half = x[: len(x) // 2]
|
256 |
+
combined = torch.cat([half, half], dim=0)
|
257 |
+
model_out = self.forward(combined, t, y)
|
258 |
+
# For exact reproducibility reasons, we apply classifier-free guidance on only
|
259 |
+
# three channels by default. The standard approach to cfg applies it to all channels.
|
260 |
+
# This can be done by uncommenting the following line and commenting-out the line following that.
|
261 |
+
# eps, rest = model_out[:, :self.in_channels], model_out[:, self.in_channels:]
|
262 |
+
eps, rest = model_out[:, :3], model_out[:, 3:]
|
263 |
+
cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
|
264 |
+
half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
|
265 |
+
eps = torch.cat([half_eps, half_eps], dim=0)
|
266 |
+
return torch.cat([eps, rest], dim=1)
|
267 |
+
|
268 |
+
|
269 |
+
#################################################################################
|
270 |
+
# Sine/Cosine Positional Embedding Functions #
|
271 |
+
#################################################################################
|
272 |
+
# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py
|
273 |
+
|
274 |
+
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
|
275 |
+
"""
|
276 |
+
grid_size: int of the grid height and width
|
277 |
+
return:
|
278 |
+
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
|
279 |
+
"""
|
280 |
+
grid_h = np.arange(grid_size, dtype=np.float32)
|
281 |
+
grid_w = np.arange(grid_size, dtype=np.float32)
|
282 |
+
grid = np.meshgrid(grid_w, grid_h) # here w goes first
|
283 |
+
grid = np.stack(grid, axis=0)
|
284 |
+
|
285 |
+
grid = grid.reshape([2, 1, grid_size, grid_size])
|
286 |
+
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
|
287 |
+
if cls_token and extra_tokens > 0:
|
288 |
+
pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
|
289 |
+
return pos_embed
|
290 |
+
|
291 |
+
|
292 |
+
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
|
293 |
+
assert embed_dim % 2 == 0
|
294 |
+
|
295 |
+
# use half of dimensions to encode grid_h
|
296 |
+
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
|
297 |
+
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
|
298 |
+
|
299 |
+
emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
|
300 |
+
return emb
|
301 |
+
|
302 |
+
|
303 |
+
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
|
304 |
+
"""
|
305 |
+
embed_dim: output dimension for each position
|
306 |
+
pos: a list of positions to be encoded: size (M,)
|
307 |
+
out: (M, D)
|
308 |
+
"""
|
309 |
+
assert embed_dim % 2 == 0
|
310 |
+
omega = np.arange(embed_dim // 2, dtype=np.float64)
|
311 |
+
omega /= embed_dim / 2.
|
312 |
+
omega = 1. / 10000**omega # (D/2,)
|
313 |
+
|
314 |
+
pos = pos.reshape(-1) # (M,)
|
315 |
+
out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product
|
316 |
+
|
317 |
+
emb_sin = np.sin(out) # (M, D/2)
|
318 |
+
emb_cos = np.cos(out) # (M, D/2)
|
319 |
+
|
320 |
+
emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
|
321 |
+
return emb
|
322 |
+
|
323 |
+
|
324 |
+
#################################################################################
|
325 |
+
# DiT Configs #
|
326 |
+
#################################################################################
|
327 |
+
|
328 |
+
def DiT_XL_2(**kwargs):
|
329 |
+
return DiT(depth=28, hidden_size=1152, patch_size=2, num_heads=16, **kwargs)
|
330 |
+
|
331 |
+
def DiT_XL_4(**kwargs):
|
332 |
+
return DiT(depth=28, hidden_size=1152, patch_size=4, num_heads=16, **kwargs)
|
333 |
+
|
334 |
+
def DiT_XL_8(**kwargs):
|
335 |
+
return DiT(depth=28, hidden_size=1152, patch_size=8, num_heads=16, **kwargs)
|
336 |
+
|
337 |
+
def DiT_L_2(**kwargs):
|
338 |
+
return DiT(depth=24, hidden_size=1024, patch_size=2, num_heads=16, **kwargs)
|
339 |
+
|
340 |
+
def DiT_L_4(**kwargs):
|
341 |
+
return DiT(depth=24, hidden_size=1024, patch_size=4, num_heads=16, **kwargs)
|
342 |
+
|
343 |
+
def DiT_L_8(**kwargs):
|
344 |
+
return DiT(depth=24, hidden_size=1024, patch_size=8, num_heads=16, **kwargs)
|
345 |
+
|
346 |
+
def DiT_B_2(**kwargs):
|
347 |
+
return DiT(depth=12, hidden_size=768, patch_size=2, num_heads=12, **kwargs)
|
348 |
+
|
349 |
+
def DiT_B_4(**kwargs):
|
350 |
+
return DiT(depth=12, hidden_size=768, patch_size=4, num_heads=12, **kwargs)
|
351 |
+
|
352 |
+
def DiT_B_8(**kwargs):
|
353 |
+
return DiT(depth=12, hidden_size=768, patch_size=8, num_heads=12, **kwargs)
|
354 |
+
|
355 |
+
def DiT_S_2(**kwargs):
|
356 |
+
return DiT(depth=12, hidden_size=384, patch_size=2, num_heads=6, **kwargs)
|
357 |
+
|
358 |
+
def DiT_S_4(**kwargs):
|
359 |
+
return DiT(depth=12, hidden_size=384, patch_size=4, num_heads=6, **kwargs)
|
360 |
+
|
361 |
+
def DiT_S_8(**kwargs):
|
362 |
+
return DiT(depth=12, hidden_size=384, patch_size=8, num_heads=6, **kwargs)
|
363 |
+
|
364 |
+
|
365 |
+
DiT_models = {
|
366 |
+
'DiT-XL/2': DiT_XL_2, 'DiT-XL/4': DiT_XL_4, 'DiT-XL/8': DiT_XL_8,
|
367 |
+
'DiT-L/2': DiT_L_2, 'DiT-L/4': DiT_L_4, 'DiT-L/8': DiT_L_8,
|
368 |
+
'DiT-B/2': DiT_B_2, 'DiT-B/4': DiT_B_4, 'DiT-B/8': DiT_B_8,
|
369 |
+
'DiT-S/2': DiT_S_2, 'DiT-S/4': DiT_S_4, 'DiT-S/8': DiT_S_8,
|
370 |
+
}
|