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
|
use candle::{bail, DType, Module, Result, Tensor};
use candle_nn as nn;
pub struct PatchEmbedder {
proj: nn::Conv2d,
}
impl PatchEmbedder {
pub fn new(
patch_size: usize,
in_channels: usize,
embed_dim: usize,
vb: nn::VarBuilder,
) -> Result<Self> {
let proj = nn::conv2d(
in_channels,
embed_dim,
patch_size,
nn::Conv2dConfig {
stride: patch_size,
..Default::default()
},
vb.pp("proj"),
)?;
Ok(Self { proj })
}
}
impl Module for PatchEmbedder {
fn forward(&self, x: &Tensor) -> Result<Tensor> {
let x = self.proj.forward(x)?;
// flatten spatial dim and transpose to channels last
let (b, c, h, w) = x.dims4()?;
x.reshape((b, c, h * w))?.transpose(1, 2)
}
}
pub struct Unpatchifier {
patch_size: usize,
out_channels: usize,
}
impl Unpatchifier {
pub fn new(patch_size: usize, out_channels: usize) -> Result<Self> {
Ok(Self {
patch_size,
out_channels,
})
}
pub fn unpatchify(&self, x: &Tensor, h: usize, w: usize) -> Result<Tensor> {
let h = (h + 1) / self.patch_size;
let w = (w + 1) / self.patch_size;
let x = x.reshape((
x.dim(0)?,
h,
w,
self.patch_size,
self.patch_size,
self.out_channels,
))?;
let x = x.permute((0, 5, 1, 3, 2, 4))?; // "nhwpqc->nchpwq"
x.reshape((
x.dim(0)?,
self.out_channels,
self.patch_size * h,
self.patch_size * w,
))
}
}
pub struct PositionEmbedder {
pos_embed: Tensor,
patch_size: usize,
pos_embed_max_size: usize,
}
impl PositionEmbedder {
pub fn new(
hidden_size: usize,
patch_size: usize,
pos_embed_max_size: usize,
vb: nn::VarBuilder,
) -> Result<Self> {
let pos_embed = vb.get(
(1, pos_embed_max_size * pos_embed_max_size, hidden_size),
"pos_embed",
)?;
Ok(Self {
pos_embed,
patch_size,
pos_embed_max_size,
})
}
pub fn get_cropped_pos_embed(&self, h: usize, w: usize) -> Result<Tensor> {
let h = (h + 1) / self.patch_size;
let w = (w + 1) / self.patch_size;
if h > self.pos_embed_max_size || w > self.pos_embed_max_size {
bail!("Input size is too large for the position embedding")
}
let top = (self.pos_embed_max_size - h) / 2;
let left = (self.pos_embed_max_size - w) / 2;
let pos_embed =
self.pos_embed
.reshape((1, self.pos_embed_max_size, self.pos_embed_max_size, ()))?;
let pos_embed = pos_embed.narrow(1, top, h)?.narrow(2, left, w)?;
pos_embed.reshape((1, h * w, ()))
}
}
pub struct TimestepEmbedder {
mlp: nn::Sequential,
frequency_embedding_size: usize,
}
impl TimestepEmbedder {
pub fn new(
hidden_size: usize,
frequency_embedding_size: usize,
vb: nn::VarBuilder,
) -> Result<Self> {
let mlp = nn::seq()
.add(nn::linear(
frequency_embedding_size,
hidden_size,
vb.pp("mlp.0"),
)?)
.add(nn::Activation::Silu)
.add(nn::linear(hidden_size, hidden_size, vb.pp("mlp.2"))?);
Ok(Self {
mlp,
frequency_embedding_size,
})
}
fn timestep_embedding(t: &Tensor, dim: usize, max_period: f64) -> Result<Tensor> {
if dim % 2 != 0 {
bail!("Embedding dimension must be even")
}
if t.dtype() != DType::F32 && t.dtype() != DType::F64 {
bail!("Input tensor must be floating point")
}
let half = dim / 2;
let freqs = Tensor::arange(0f32, half as f32, t.device())?
.to_dtype(candle::DType::F32)?
.mul(&Tensor::full(
(-f64::ln(max_period) / half as f64) as f32,
half,
t.device(),
)?)?
.exp()?;
let args = t
.unsqueeze(1)?
.to_dtype(candle::DType::F32)?
.matmul(&freqs.unsqueeze(0)?)?;
let embedding = Tensor::cat(&[args.cos()?, args.sin()?], 1)?;
embedding.to_dtype(candle::DType::F16)
}
}
impl Module for TimestepEmbedder {
fn forward(&self, t: &Tensor) -> Result<Tensor> {
let t_freq = Self::timestep_embedding(t, self.frequency_embedding_size, 10000.0)?;
self.mlp.forward(&t_freq)
}
}
pub struct VectorEmbedder {
mlp: nn::Sequential,
}
impl VectorEmbedder {
pub fn new(input_dim: usize, hidden_size: usize, vb: nn::VarBuilder) -> Result<Self> {
let mlp = nn::seq()
.add(nn::linear(input_dim, hidden_size, vb.pp("mlp.0"))?)
.add(nn::Activation::Silu)
.add(nn::linear(hidden_size, hidden_size, vb.pp("mlp.2"))?);
Ok(Self { mlp })
}
}
impl Module for VectorEmbedder {
fn forward(&self, x: &Tensor) -> Result<Tensor> {
self.mlp.forward(x)
}
}
|
