kxngluca/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

Re-add biases, also add new init

Browse Source 
A/B testing where we lost our GAN competitiveness.
This commit is contained in:
James Betker 2020-07-04 22:24:42 -06:00
parent b03741f30e
commit 86cda86e94
1 changed files with 19 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
codes/models/archs/SRG1_arch.py
View File

@ -8,11 +8,11 @@ from switched_conv_util import save_attention_to_image
class ConvBnLelu(nn.Module): class ConvBnLelu(nn.Module):
def __init__(self, filters_in, filters_out, kernel_size=3, stride=1, lelu=True, bn=True, bias=True): def __init__(self, filters_in, filters_out, kernel_size=3, stride=1, lelu=True, bn=True):
super(ConvBnLelu, self).__init__() super(ConvBnLelu, self).__init__()
padding_map = {1: 0, 3: 1, 5: 2, 7: 3} padding_map = {1: 0, 3: 1, 5: 2, 7: 3}
assert kernel_size in padding_map.keys() assert kernel_size in padding_map.keys()
self.conv = nn.Conv2d(filters_in, filters_out, kernel_size, stride, padding_map[kernel_size], bias=bias) self.conv = nn.Conv2d(filters_in, filters_out, kernel_size, stride, padding_map[kernel_size])
if bn: if bn:
self.bn = nn.BatchNorm2d(filters_out) self.bn = nn.BatchNorm2d(filters_out)
else: else:
@ -22,6 +22,15 @@ class ConvBnLelu(nn.Module):
else: else:
self.lelu = None self.lelu = None
# Init params.
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, a=.1, mode='fan_out',
nonlinearity='leaky_relu' if self.lelu else 'linear')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def forward(self, x): def forward(self, x):
x = self.conv(x) x = self.conv(x)
if self.bn: if self.bn:
@ -33,14 +42,13 @@ class ConvBnLelu(nn.Module):
class ResidualBranch(nn.Module): class ResidualBranch(nn.Module):
def __init__(self, filters_in, filters_mid, filters_out, kernel_size, depth, bn=False): def __init__(self, filters_in, filters_mid, filters_out, kernel_size, depth):
assert depth >= 2 assert depth >= 2
super(ResidualBranch, self).__init__() super(ResidualBranch, self).__init__()
self.noise_scale = nn.Parameter(torch.full((1,), fill_value=.01)) self.noise_scale = nn.Parameter(torch.full((1,), fill_value=.01))
self.bnconvs = nn.ModuleList([ConvBnLelu(filters_in, filters_mid, kernel_size, bn=bn, bias=False)] + self.bnconvs = nn.ModuleList([ConvBnLelu(filters_in, filters_mid, kernel_size, bn=False)] +
[ConvBnLelu(filters_mid, filters_mid, kernel_size, bn=bn, bias=False) for i in range(depth-2)] + [ConvBnLelu(filters_mid, filters_mid, kernel_size, bn=False) for i in range(depth-2)] +
[ConvBnLelu(filters_mid, filters_out, kernel_size, lelu=False, bn=False, bias=False)]) [ConvBnLelu(filters_mid, filters_out, kernel_size, lelu=False, bn=False)])
self.scale = nn.Parameter(torch.ones(1)) self.scale = nn.Parameter(torch.ones(1))
self.bias = nn.Parameter(torch.zeros(1)) self.bias = nn.Parameter(torch.zeros(1))
@ -58,8 +66,9 @@ class ResidualBranch(nn.Module):
class HalvingProcessingBlock(nn.Module): class HalvingProcessingBlock(nn.Module):
def __init__(self, filters): def __init__(self, filters):
super(HalvingProcessingBlock, self).__init__() super(HalvingProcessingBlock, self).__init__()
self.bnconv1 = ConvBnLelu(filters, filters * 2, stride=2, bn=False, bias=False) self.bnconv1 = ConvBnLelu(filters, filters * 2, stride=2, bn=False)
self.bnconv2 = ConvBnLelu(filters * 2, filters * 2, bn=True, bias=False) self.bnconv2 = ConvBnLelu(filters * 2, filters * 2, bn=True)
def forward(self, x): def forward(self, x):
x = self.bnconv1(x) x = self.bnconv1(x)
return self.bnconv2(x) return self.bnconv2(x)
@ -71,7 +80,7 @@ def create_sequential_growing_processing_block(filters_init, filter_growth, num_
convs = [] convs = []
current_filters = filters_init current_filters = filters_init
for i in range(num_convs): for i in range(num_convs):
convs.append(ConvBnRelu(current_filters, current_filters + filter_growth, bn=True, bias=False)) convs.append(ConvBnLelu(current_filters, current_filters + filter_growth, bn=True))
current_filters += filter_growth current_filters += filter_growth
return nn.Sequential(*convs), current_filters return nn.Sequential(*convs), current_filters