Commit
·
79b5829
1
Parent(s):
657d518
Initial CosAE release
Browse files- cosae/__init__.py +0 -0
- cosae/config.py +57 -0
- cosae/cosae.py +53 -0
- cosae/modules.py +267 -0
cosae/__init__.py
ADDED
File without changes
|
cosae/config.py
ADDED
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import PretrainedConfig, PreTrainedModel
|
2 |
+
|
3 |
+
|
4 |
+
class CosAEConfig(PretrainedConfig):
|
5 |
+
model_type = "cosae"
|
6 |
+
|
7 |
+
def __init__(
|
8 |
+
self,
|
9 |
+
image_size: tuple[int, int] = (256, 256),
|
10 |
+
# Encoder parameters
|
11 |
+
in_channels: int = 3,
|
12 |
+
hidden_dims: list[int] = (64, 128, 256, 512),
|
13 |
+
num_res_blocks: int = 2,
|
14 |
+
downsample_strides: list[int] = (2, 2, 2, 2),
|
15 |
+
use_encoder_attention: bool = True,
|
16 |
+
encoder_attention_heads: int = 8,
|
17 |
+
encoder_attention_layers: int = 1,
|
18 |
+
bottleneck_channels: int = 256,
|
19 |
+
basis_size: int = 32,
|
20 |
+
norm_type: str = "gn", # "gn" (GroupNorm) or "ln" (LayerNorm)
|
21 |
+
activation: str = "gelu", # "gelu" or "silu"
|
22 |
+
|
23 |
+
# Decoder parameters
|
24 |
+
decoder_hidden_dim: int = 256,
|
25 |
+
decoder_upsample_strides: list[int] = (2,), # e.g. (2,) for one 2× upsample
|
26 |
+
use_decoder_attention: bool = False,
|
27 |
+
decoder_attention_heads: int = 8,
|
28 |
+
decoder_attention_layers: int = 0,
|
29 |
+
|
30 |
+
**kwargs,
|
31 |
+
):
|
32 |
+
"""
|
33 |
+
Configuration for CosAEModel, including encoder, HCM, and decoder settings.
|
34 |
+
"""
|
35 |
+
super().__init__(**kwargs)
|
36 |
+
|
37 |
+
# Encoder settings
|
38 |
+
self.in_channels = in_channels
|
39 |
+
self.hidden_dims = list(hidden_dims)
|
40 |
+
self.num_res_blocks = num_res_blocks
|
41 |
+
self.downsample_strides = list(downsample_strides)
|
42 |
+
self.use_encoder_attention = use_encoder_attention
|
43 |
+
self.encoder_attention_heads = encoder_attention_heads
|
44 |
+
self.encoder_attention_layers = encoder_attention_layers
|
45 |
+
self.bottleneck_channels = bottleneck_channels
|
46 |
+
self.basis_size = basis_size
|
47 |
+
self.norm_type = norm_type
|
48 |
+
self.activation = activation
|
49 |
+
self.image_size = image_size
|
50 |
+
|
51 |
+
# Decoder settings
|
52 |
+
self.decoder_hidden_dim = decoder_hidden_dim
|
53 |
+
self.decoder_upsample_strides = list(decoder_upsample_strides)
|
54 |
+
self.use_decoder_attention = use_decoder_attention
|
55 |
+
self.decoder_attention_heads = decoder_attention_heads
|
56 |
+
self.decoder_attention_layers = decoder_attention_layers
|
57 |
+
|
cosae/cosae.py
ADDED
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import math
|
5 |
+
|
6 |
+
from transformers import PreTrainedModel
|
7 |
+
from .modules import *
|
8 |
+
from .config import CosAEConfig
|
9 |
+
|
10 |
+
class CosAEModel(PreTrainedModel):
|
11 |
+
config_class = CosAEConfig
|
12 |
+
base_model_prefix = "cosae"
|
13 |
+
|
14 |
+
def __init__(self, config: CosAEConfig):
|
15 |
+
super().__init__(config)
|
16 |
+
# 1) Encoder
|
17 |
+
self.encoder = CosAEEncoder(config)
|
18 |
+
|
19 |
+
# 2) Harmonic Construction Module
|
20 |
+
# derive P = total downsampling factor from encoder strides
|
21 |
+
stem_ds = 2 * 2
|
22 |
+
P = stem_ds * math.prod(config.downsample_strides)
|
23 |
+
# basis size T = P // 2
|
24 |
+
T = P // 2
|
25 |
+
self.T = T
|
26 |
+
self.hcm = HarmonicConstructionModule(
|
27 |
+
bottleneck_channels=config.bottleneck_channels,
|
28 |
+
basis_size=config.basis_size
|
29 |
+
)
|
30 |
+
|
31 |
+
# 3) Decoder
|
32 |
+
self.decoder = CosAEDecoder(config)
|
33 |
+
|
34 |
+
# initialize weights, etc.
|
35 |
+
self.post_init()
|
36 |
+
|
37 |
+
def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
|
38 |
+
"""
|
39 |
+
Args:
|
40 |
+
pixel_values: [B, C_in, H, W] (C_in = 3 or 9 if using FFT)
|
41 |
+
Returns:
|
42 |
+
recon: [B, 3, H, W] reconstructed image
|
43 |
+
"""
|
44 |
+
# Encode to get amplitudes & phases
|
45 |
+
bottleneck = self.encoder(pixel_values) # [B, 2c, H', W']
|
46 |
+
amp, ph = torch.chunk(bottleneck, 2, dim=1) # each [B, c, H', W']
|
47 |
+
|
48 |
+
# Build harmonics
|
49 |
+
harmonics = self.hcm(amp, ph) # [B, c, H, W]
|
50 |
+
|
51 |
+
# Decode to reconstruct
|
52 |
+
recon = self.decoder(harmonics) # [B, 3, H, W]
|
53 |
+
return recon
|
cosae/modules.py
ADDED
@@ -0,0 +1,267 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
import torch.nn.functional as F
|
5 |
+
|
6 |
+
from transformers import PretrainedConfig, PreTrainedModel
|
7 |
+
from .config import CosAEConfig
|
8 |
+
|
9 |
+
"""This code has partially been generated by ChatGPT"""
|
10 |
+
|
11 |
+
class ResBlock(nn.Module):
|
12 |
+
def __init__(self, in_ch, out_ch, norm_type="gn", activation="gelu"):
|
13 |
+
super().__init__()
|
14 |
+
Norm = nn.GroupNorm if norm_type == "gn" else nn.LayerNorm
|
15 |
+
act = nn.GELU if activation == "gelu" else nn.SiLU
|
16 |
+
|
17 |
+
self.conv1 = nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1, bias=False)
|
18 |
+
self.norm1 = Norm(8, out_ch) if norm_type == "gn" else Norm(out_ch)
|
19 |
+
self.act1 = act()
|
20 |
+
|
21 |
+
self.conv2 = nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1, bias=False)
|
22 |
+
self.norm2 = Norm(8, out_ch) if norm_type == "gn" else Norm(out_ch)
|
23 |
+
self.act2 = act()
|
24 |
+
|
25 |
+
if in_ch != out_ch:
|
26 |
+
self.skip = nn.Conv2d(in_ch, out_ch, kernel_size=1, bias=False)
|
27 |
+
else:
|
28 |
+
self.skip = nn.Identity()
|
29 |
+
|
30 |
+
def forward(self, x):
|
31 |
+
identity = self.skip(x)
|
32 |
+
out = self.conv1(x)
|
33 |
+
out = self.norm1(out)
|
34 |
+
out = self.act1(out)
|
35 |
+
out = self.conv2(out)
|
36 |
+
out = self.norm2(out)
|
37 |
+
out = out + identity
|
38 |
+
return self.act2(out)
|
39 |
+
|
40 |
+
|
41 |
+
class CosAEEncoder(PreTrainedModel):
|
42 |
+
config_class = CosAEConfig
|
43 |
+
base_model_prefix = "encoder"
|
44 |
+
|
45 |
+
def __init__(self, config: CosAEConfig):
|
46 |
+
super().__init__(config)
|
47 |
+
c = config
|
48 |
+
# Stem
|
49 |
+
self.stem = nn.Sequential(
|
50 |
+
nn.Conv2d(c.in_channels, c.hidden_dims[0], kernel_size=7, stride=2, padding=3, bias=False),
|
51 |
+
nn.GroupNorm(8, c.hidden_dims[0]) if c.norm_type == "gn" else nn.LayerNorm([c.hidden_dims[0], 128, 128]),
|
52 |
+
nn.GELU() if c.activation == "gelu" else nn.SiLU(),
|
53 |
+
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),
|
54 |
+
)
|
55 |
+
# Downsampling stages
|
56 |
+
dims = c.hidden_dims
|
57 |
+
self.stages = nn.ModuleList()
|
58 |
+
in_ch = dims[0]
|
59 |
+
for i, out_ch in enumerate(dims[1:]):
|
60 |
+
blocks = []
|
61 |
+
for _ in range(c.num_res_blocks):
|
62 |
+
blocks.append(ResBlock(in_ch, out_ch, norm_type=c.norm_type, activation=c.activation))
|
63 |
+
in_ch = out_ch
|
64 |
+
# downsample conv
|
65 |
+
blocks.append(
|
66 |
+
nn.Sequential(
|
67 |
+
nn.Conv2d(in_ch, out_ch, kernel_size=3, stride=c.downsample_strides[i], padding=1, bias=False),
|
68 |
+
nn.GroupNorm(8, out_ch) if c.norm_type == "gn" else nn.LayerNorm([out_ch, -1, -1]),
|
69 |
+
nn.GELU() if c.activation == "gelu" else nn.SiLU(),
|
70 |
+
)
|
71 |
+
)
|
72 |
+
self.stages.append(nn.Sequential(*blocks))
|
73 |
+
|
74 |
+
# Optional global attention
|
75 |
+
if c.use_encoder_attention:
|
76 |
+
encoder_layer = nn.TransformerEncoderLayer(
|
77 |
+
d_model=dims[-1],
|
78 |
+
nhead=c.encoder_attention_heads,
|
79 |
+
dim_feedforward=dims[-1] * 4,
|
80 |
+
activation=c.activation,
|
81 |
+
batch_first=True,
|
82 |
+
)
|
83 |
+
self.attn = nn.TransformerEncoder(encoder_layer, num_layers=c.encoder_attention_layers)
|
84 |
+
else:
|
85 |
+
self.attn = None
|
86 |
+
|
87 |
+
# Head: project to 2 * bottleneck_channels
|
88 |
+
self.head = nn.Conv2d(dims[-1], 2 * c.bottleneck_channels, kernel_size=1)
|
89 |
+
|
90 |
+
# Initialize weights
|
91 |
+
self.post_init()
|
92 |
+
|
93 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
94 |
+
"""
|
95 |
+
Args:
|
96 |
+
x: [B, C_in, H, W]
|
97 |
+
Returns:
|
98 |
+
bottleneck: [B, 2c, H/P, W/P]
|
99 |
+
"""
|
100 |
+
# Stem
|
101 |
+
x = self.stem(x)
|
102 |
+
# Stages
|
103 |
+
for stage in self.stages:
|
104 |
+
x = stage(x)
|
105 |
+
# x: [B, dims[-1], H/P, W/P]
|
106 |
+
# Optional attention
|
107 |
+
if self.attn is not None:
|
108 |
+
B, C, H, W = x.shape
|
109 |
+
seq = x.flatten(2).transpose(1, 2) # [B, H*W, C]
|
110 |
+
seq = self.attn(seq)
|
111 |
+
x = seq.transpose(1, 2).view(B, C, H, W)
|
112 |
+
# Head
|
113 |
+
bottleneck = self.head(x)
|
114 |
+
return bottleneck
|
115 |
+
|
116 |
+
class HarmonicConstructionModule(nn.Module):
|
117 |
+
"""
|
118 |
+
Given:
|
119 |
+
- amplitudes: Tensor of shape [B, c, H', W']
|
120 |
+
- phases: Tensor of shape [B, c, H', W']
|
121 |
+
and learnable frequencies (u, v) of shape [c, 2],
|
122 |
+
this module builds a [B, c, H'*T, W'*T] tensor of harmonics:
|
123 |
+
H[b,k,i*T + x, j*T + y]
|
124 |
+
= A[b,k,i,j] * cos( 2π/T * (u[k]*x + v[k]*y) - Φ[b,k,i,j] )
|
125 |
+
"""
|
126 |
+
def __init__(self, bottleneck_channels: int, basis_size: int):
|
127 |
+
"""
|
128 |
+
Args:
|
129 |
+
bottleneck_channels: c, number of freq components
|
130 |
+
basis_size: T, size of each cosine basis (e.g. 32 or 64)
|
131 |
+
"""
|
132 |
+
super().__init__()
|
133 |
+
self.c = bottleneck_channels
|
134 |
+
self.T = basis_size
|
135 |
+
|
136 |
+
# Learnable frequencies in [0, T/2)
|
137 |
+
self.freqs = nn.Parameter(
|
138 |
+
torch.rand(self.c, 2) * (self.T / 2)
|
139 |
+
) # shape [c,2] for (u,v)
|
140 |
+
|
141 |
+
# Precompute the x,y grid of size [T, T]
|
142 |
+
x = torch.arange(self.T, dtype=torch.float32)
|
143 |
+
y = torch.arange(self.T, dtype=torch.float32)
|
144 |
+
xs, ys = torch.meshgrid(x, y, indexing="ij") # both shape [T,T]
|
145 |
+
|
146 |
+
# Register as buffers so they move with .to(device)
|
147 |
+
self.register_buffer("xs", xs) # [T,T]
|
148 |
+
self.register_buffer("ys", ys) # [T,T]
|
149 |
+
|
150 |
+
def forward(self, amplitude: torch.Tensor, phase: torch.Tensor) -> torch.Tensor:
|
151 |
+
"""
|
152 |
+
Args:
|
153 |
+
amplitude: [B, c, H', W']
|
154 |
+
phase: [B, c, H', W']
|
155 |
+
Returns:
|
156 |
+
harmonics: [B, c, H'*T, W'*T]
|
157 |
+
"""
|
158 |
+
B, c, Hp, Wp = amplitude.shape
|
159 |
+
assert c == self.c, "Channel mismatch"
|
160 |
+
|
161 |
+
# 1) compute spatial_phase for each freq: [c, T, T]
|
162 |
+
# 2π/T * (u[k]*xs + v[k]*ys)
|
163 |
+
u = self.freqs[:, 0].view(c, 1, 1) # [c,1,1]
|
164 |
+
v = self.freqs[:, 1].view(c, 1, 1) # [c,1,1]
|
165 |
+
spatial_phase = (2 * math.pi / self.T) * (u * self.xs + v * self.ys)
|
166 |
+
# reshape for broadcasting to [1,c,1,1,T,T]
|
167 |
+
spatial_phase = spatial_phase.view(1, c, 1, 1, self.T, self.T)
|
168 |
+
|
169 |
+
# 2) prepare amplitude & phase maps:
|
170 |
+
# [B, c, Hp, Wp] → [B, c, Hp, Wp, 1, 1]
|
171 |
+
A = amplitude.view(B, c, Hp, Wp, 1, 1)
|
172 |
+
Φ = phase.view(B, c, Hp, Wp, 1, 1)
|
173 |
+
|
174 |
+
# 3) compute argument and harmonic:
|
175 |
+
# arg = spatial_phase - Φ
|
176 |
+
# H = A * cos(arg)
|
177 |
+
arg = spatial_phase - Φ # [B, c, Hp, Wp, T, T]
|
178 |
+
H = A * torch.cos(arg) # same shape
|
179 |
+
|
180 |
+
# 4) tile out to full spatial size [B, c, Hp*T, Wp*T]
|
181 |
+
# first permute to [B, c, Hp, T, Wp, T] then reshape
|
182 |
+
H = H.permute(0, 1, 2, 4, 3, 5) # [B, c, Hp, T, Wp, T]
|
183 |
+
H = H.reshape(B, c, Hp * self.T, Wp * self.T)
|
184 |
+
|
185 |
+
return H
|
186 |
+
|
187 |
+
class CosAEDecoder(PreTrainedModel):
|
188 |
+
config_class = CosAEConfig
|
189 |
+
base_model_prefix = "decoder"
|
190 |
+
|
191 |
+
def __init__(self, config: CosAEConfig):
|
192 |
+
super().__init__(config)
|
193 |
+
c = config
|
194 |
+
|
195 |
+
# 1×1 projection from HCM channels → decoder hidden dim
|
196 |
+
self.proj = nn.Conv2d(
|
197 |
+
c.bottleneck_channels,
|
198 |
+
c.decoder_hidden_dim,
|
199 |
+
kernel_size=1,
|
200 |
+
bias=False
|
201 |
+
)
|
202 |
+
# normalization + activation after proj
|
203 |
+
Norm = nn.GroupNorm if c.norm_type == "gn" else nn.LayerNorm
|
204 |
+
self.norm0 = Norm(8, c.decoder_hidden_dim) if c.norm_type=="gn" else Norm([c.decoder_hidden_dim, -1, -1])
|
205 |
+
self.act0 = nn.GELU() if c.activation=="gelu" else nn.SiLU()
|
206 |
+
|
207 |
+
# upsampling blocks
|
208 |
+
self.upsamples = nn.ModuleList()
|
209 |
+
for scale in c.decoder_upsample_strides:
|
210 |
+
block = nn.Sequential(
|
211 |
+
nn.Upsample(scale_factor=scale, mode="bilinear", align_corners=False),
|
212 |
+
nn.Conv2d(c.decoder_hidden_dim, c.decoder_hidden_dim, kernel_size=3, padding=1, bias=False),
|
213 |
+
Norm(8, c.decoder_hidden_dim) if c.norm_type=="gn" else Norm([c.decoder_hidden_dim, -1, -1]),
|
214 |
+
nn.GELU() if c.activation=="gelu" else nn.SiLU(),
|
215 |
+
)
|
216 |
+
self.upsamples.append(block)
|
217 |
+
|
218 |
+
# optional global attention in decoder
|
219 |
+
if c.use_decoder_attention:
|
220 |
+
enc_layer = nn.TransformerEncoderLayer(
|
221 |
+
d_model=c.decoder_hidden_dim,
|
222 |
+
nhead=c.decoder_attention_heads,
|
223 |
+
dim_feedforward=c.decoder_hidden_dim * 4,
|
224 |
+
activation=c.activation,
|
225 |
+
batch_first=True,
|
226 |
+
)
|
227 |
+
self.attn = nn.TransformerEncoder(enc_layer, num_layers=c.decoder_attention_layers)
|
228 |
+
else:
|
229 |
+
self.attn = None
|
230 |
+
|
231 |
+
# final conv to RGB
|
232 |
+
self.final_conv = nn.Conv2d(
|
233 |
+
c.decoder_hidden_dim,
|
234 |
+
3,
|
235 |
+
kernel_size=3,
|
236 |
+
padding=1
|
237 |
+
)
|
238 |
+
|
239 |
+
# initialize weights
|
240 |
+
self.post_init()
|
241 |
+
|
242 |
+
def forward(self, harmonics: torch.Tensor) -> torch.Tensor:
|
243 |
+
"""
|
244 |
+
Args:
|
245 |
+
harmonics: Tensor from HCM, shape [B, c, H*, W*]
|
246 |
+
Returns:
|
247 |
+
recon: Reconstructed image, shape [B, 3, H, W]
|
248 |
+
"""
|
249 |
+
x = self.proj(harmonics) # [B, hidden_dim, H*, W*]
|
250 |
+
x = self.norm0(x)
|
251 |
+
x = self.act0(x)
|
252 |
+
|
253 |
+
# upsample to higher resolution
|
254 |
+
for up in self.upsamples:
|
255 |
+
x = up(x) # doubles H*, W* each block
|
256 |
+
|
257 |
+
# optional global attention
|
258 |
+
if self.attn is not None:
|
259 |
+
B, C, H, W = x.shape
|
260 |
+
seq = x.flatten(2).transpose(1, 2) # [B, H*W, C]
|
261 |
+
seq = self.attn(seq)
|
262 |
+
x = seq.transpose(1, 2).view(B, C, H, W)
|
263 |
+
|
264 |
+
# final RGB projection
|
265 |
+
recon = self.final_conv(x)
|
266 |
+
return recon
|
267 |
+
|