forked from Lightning-Universe/lightning-bolts
-
Notifications
You must be signed in to change notification settings - Fork 0
/
pixel_cnn.py
61 lines (48 loc) · 1.89 KB
/
pixel_cnn.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
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
"""
PixelCNN
Implemented by: William Falcon
Reference: https://arxiv.org/pdf/1905.09272.pdf (page 15)
Accessed: May 14, 2020
"""
from torch import nn
from torch.nn import functional as F
class PixelCNN(nn.Module):
"""Implementation of `Pixel CNN <https://arxiv.org/abs/1606.05328>`_.
Paper authors: Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves,
Koray Kavukcuoglu
Implemented by:
- William Falcon
Example::
>>> from pl_bolts.models.vision import PixelCNN
>>> import torch
...
>>> model = PixelCNN(input_channels=3)
>>> x = torch.rand(5, 3, 64, 64)
>>> out = model(x)
...
>>> out.shape
torch.Size([5, 3, 64, 64])
"""
def __init__(self, input_channels: int, hidden_channels: int = 256, num_blocks=5):
super().__init__()
self.input_channels = input_channels
self.hidden_channels = hidden_channels
self.blocks = nn.ModuleList([self.conv_block(input_channels) for _ in range(num_blocks)])
def conv_block(self, input_channels):
c1 = nn.Conv2d(in_channels=input_channels, out_channels=self.hidden_channels, kernel_size=(1, 1))
act1 = nn.ReLU()
c2 = nn.Conv2d(in_channels=self.hidden_channels, out_channels=self.hidden_channels, kernel_size=(1, 3))
pad = nn.ConstantPad2d((0, 0, 1, 0, 0, 0, 0, 0), 1)
c3 = nn.Conv2d(
in_channels=self.hidden_channels, out_channels=self.hidden_channels, kernel_size=(2, 1), padding=(0, 1)
)
act2 = nn.ReLU()
c4 = nn.Conv2d(in_channels=self.hidden_channels, out_channels=input_channels, kernel_size=(1, 1))
block = nn.Sequential(c1, act1, c2, pad, c3, act2, c4)
return block
def forward(self, z):
c = z
for conv_block in self.blocks:
c = c + conv_block(c)
c = F.relu(c)
return c