-
Notifications
You must be signed in to change notification settings - Fork 0
/
decoder.py
37 lines (30 loc) · 937 Bytes
/
decoder.py
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
from hyperparamters import *
def get_decoder(
input_shape,
num_layers=num_layers,
patch_size=patch_size,
hidden_size=hidden_size,
mlp_dim=mlp_dim,
dropout=.02,
num_heads=num_heads,
name='Decoder'
):
in_channels = 3
patch_dim = in_channels * patch_size ** 2
h = 128 // patch_size
ip = Input(shape=input_shape)
y = ip
for n in range(4):
y, _ = vit_layers.TransformerBlock(
num_heads=num_heads,
mlp_dim=mlp_dim,
dropout=dropout,
name=f"T_Dec_Block_{n}"
)(y)
y = LayerNormalization(
epsilon=1e-6, name="T_LNorm"
)(y)
y = Dense(patch_dim)(y)
y = Rearrange('b (h w) (p1 p2 c) -> b (h p1) (w p2) c', h=h, w=h, p1=patch_size, p2=patch_size, c=in_channels)(y)
model = Model(inputs=ip, outputs=y, name=name)
return model