Added Custom Resnet file
Browse files- custom_resnet.py +69 -0
custom_resnet.py
ADDED
|
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch.nn as nn
|
| 2 |
+
import torch.nn.functional as F
|
| 3 |
+
|
| 4 |
+
def normalization(norm_type, embedding):
|
| 5 |
+
if norm_type=='batch':
|
| 6 |
+
return nn.BatchNorm2d(embedding)
|
| 7 |
+
elif norm_type=='layer':
|
| 8 |
+
return nn.GroupNorm(1, embedding)
|
| 9 |
+
else:
|
| 10 |
+
return nn.GroupNorm(4, embedding)
|
| 11 |
+
|
| 12 |
+
def custom_conv_layer(in_channels,
|
| 13 |
+
out_channels,
|
| 14 |
+
pool,
|
| 15 |
+
norm_type,
|
| 16 |
+
):
|
| 17 |
+
conv_layer = [
|
| 18 |
+
nn.Conv2d(in_channels, out_channels, kernel_size=(3, 3), padding=1, stride=1, bias=False)
|
| 19 |
+
]
|
| 20 |
+
if pool :
|
| 21 |
+
conv_layer.append(
|
| 22 |
+
nn.MaxPool2d(2, 2),
|
| 23 |
+
)
|
| 24 |
+
conv_layer.append(
|
| 25 |
+
normalization(norm_type, out_channels),
|
| 26 |
+
)
|
| 27 |
+
conv_layer.append(
|
| 28 |
+
nn.ReLU()
|
| 29 |
+
)
|
| 30 |
+
block = nn.Sequential(*conv_layer)
|
| 31 |
+
return block
|
| 32 |
+
|
| 33 |
+
class Net(nn.Module):
|
| 34 |
+
def __init__(self, normtype):
|
| 35 |
+
super(Net, self).__init__()
|
| 36 |
+
# prep layer
|
| 37 |
+
self.prep_layer = custom_conv_layer(3, 64, False, 'batch')
|
| 38 |
+
# layer 1
|
| 39 |
+
self.layer1_x = custom_conv_layer(64, 128, True, 'batch')
|
| 40 |
+
self.layer1_r1 = nn.Sequential(
|
| 41 |
+
custom_conv_layer(128, 128, False, 'batch'),
|
| 42 |
+
custom_conv_layer(128, 128, False, 'batch')
|
| 43 |
+
)
|
| 44 |
+
# layer 2
|
| 45 |
+
self.layer2 = custom_conv_layer(128, 256, True, 'batch')
|
| 46 |
+
# Layer 3
|
| 47 |
+
self.layer3_x = custom_conv_layer(256, 512, True, 'batch')
|
| 48 |
+
self.layer3_r3 = nn.Sequential(
|
| 49 |
+
custom_conv_layer(512, 512, False, 'batch'),
|
| 50 |
+
custom_conv_layer(512, 512, False, 'batch')
|
| 51 |
+
)
|
| 52 |
+
# MaxPooling with Kernel Size 4
|
| 53 |
+
self.pool = nn.MaxPool2d(4, 4)
|
| 54 |
+
# FC Layer
|
| 55 |
+
self.fc = nn.Linear(512, 10)
|
| 56 |
+
|
| 57 |
+
def forward(self, x):
|
| 58 |
+
x = self.prep_layer(x)
|
| 59 |
+
x1 = self.layer1_x(x)
|
| 60 |
+
r1 = self.layer1_r1(x1)
|
| 61 |
+
x = x1 + r1
|
| 62 |
+
x = self.layer2(x)
|
| 63 |
+
x3 = self.layer3_x(x)
|
| 64 |
+
r3 = self.layer3_r3(x3)
|
| 65 |
+
x = x3 + r3
|
| 66 |
+
x = self.pool(x)
|
| 67 |
+
x = x.view(-1, 512)
|
| 68 |
+
x = self.fc(x)
|
| 69 |
+
return F.softmax(x, dim=-1)
|