def __init__(self):
super(CifarNet, self).__init__()
self.pool = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)
self.dropout_050 = nn.Dropout(0.50)
self.conv_block1 = omth_blocks.Conv_Block(input=3,
filters=[32, 32],
kernel_sizes=[3, 3],
stride=[1, 1],
padding=[1, 0],
batch_norm=None,
dropout=[0, 0.25])
self.conv_block2 = omth_blocks.Conv_Block(input=32,
filters=[64, 64],
kernel_sizes=[3, 3],
stride=[1, 1],
padding=[1, 0],
batch_norm=None,
dropout=[0, 0.25])
self.fc_layer = omth_blocks.fc_layer(2304, [512, 10],
activation=[nn.ReLU(), None],
batch_norm=False)
def __init__(self, input, group):
super().__init__()
self.group = group
self.channel = input
self.attn_estimation = omth_blocks.Conv_Block(
input * group,
filters=[group * input, group],
kernel_sizes=[3, 1],
stride=[1, 1],
padding=[1, 0],
groups=[group, group],
)
def create_global_conv(self):
global_conv = []
for i, channel in enumerate(self.channels):
if i == 0:
input_channel = self.group * self.args.img_channel
else:
input_channel = self.channels[i - 1]
global_conv.append(
omth_blocks.Conv_Block(input_channel,
filters=channel,
kernel_sizes=self.kernel_sizes[i],
padding=self.global_padding[i],
stride=self.strides[i],
batch_norm=self.batch_norm))
self.global_conv = nn.Sequential(*global_conv)
def __init__(self, input, group, comb_method="add"):
super().__init__()
self.group = group
self.channel = input
self.comb_method = comb_method.lower()
assert self.comb_method in ["add", "conv"]
if self.comb_method == "conv":
self.comb_conv = omth_blocks.Conv_Block(
input * group * 2,
filters=[input * group * 2, input * group],
kernel_sizes=[3, 1],
stride=[1, 1],
padding=[1, 0],
groups=[group, group],
)