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

SRG2 architectural changes

Browse Source
This commit is contained in:
James Betker 2020-07-06 22:22:29 -06:00
parent 9a1c3241f5
commit 3c31bea1ac
1 changed files with 18 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
codes/models/archs/SwitchedResidualGenerator_arch.py
View File

@ -115,6 +115,7 @@ class ConvBasisMultiplexer(nn.Module):
self.processing_blocks, self.output_filter_count = create_sequential_growing_processing_block(reduction_filters, growth, processing_depth) self.processing_blocks, self.output_filter_count = create_sequential_growing_processing_block(reduction_filters, growth, processing_depth)
gap = self.output_filter_count - multiplexer_channels gap = self.output_filter_count - multiplexer_channels
# Hey silly - if you're going to interpolate later, do it here instead. Then add some processing layers to let the model adjust it properly.
self.cbl1 = ConvBnSilu(self.output_filter_count, self.output_filter_count - (gap // 2), bn=use_bn, bias=False) self.cbl1 = ConvBnSilu(self.output_filter_count, self.output_filter_count - (gap // 2), bn=use_bn, bias=False)
self.cbl2 = ConvBnSilu(self.output_filter_count - (gap // 2), self.output_filter_count - (3 * gap // 4), bn=use_bn, bias=False) self.cbl2 = ConvBnSilu(self.output_filter_count - (gap // 2), self.output_filter_count - (3 * gap // 4), bn=use_bn, bias=False)
self.cbl3 = ConvBnSilu(self.output_filter_count - (3 * gap // 4), multiplexer_channels, bias=True) self.cbl3 = ConvBnSilu(self.output_filter_count - (3 * gap // 4), multiplexer_channels, bias=True)
@ -152,18 +153,19 @@ class ConfigurableSwitchedResidualGenerator2(nn.Module):
add_scalable_noise_to_transforms=False): add_scalable_noise_to_transforms=False):
super(ConfigurableSwitchedResidualGenerator2, self).__init__() super(ConfigurableSwitchedResidualGenerator2, self).__init__()
switches = [] switches = []
self.initial_conv = ConvBnLelu(3, transformation_filters, bn=False) self.initial_conv = ConvBnLelu(3, transformation_filters, bn=False, lelu=False, bias=True)
self.proc_conv = ConvBnLelu(transformation_filters, transformation_filters, bn=False) self.sw_conv = ConvBnLelu(transformation_filters, transformation_filters, lelu=False, bias=True)
self.final_conv = ConvBnLelu(transformation_filters, 3, bn=False, lelu=False) self.upconv1 = ConvBnLelu(transformation_filters, transformation_filters, bn=False, biasd=True)
self.upconv2 = ConvBnLelu(transformation_filters, transformation_filters, bn=False, bias=True)
self.hr_conv = ConvBnLelu(transformation_filters, transformation_filters, bn=False, bias=True)
self.final_conv = ConvBnLelu(transformation_filters, 3, bn=False, lelu=False, bias=True)
for filters, growth, sw_reduce, sw_proc, trans_count, kernel, layers in zip(switch_filters, switch_growths, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes, trans_layers): for filters, growth, sw_reduce, sw_proc, trans_count, kernel, layers in zip(switch_filters, switch_growths, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes, trans_layers):
multiplx_fn = functools.partial(ConvBasisMultiplexer, transformation_filters, filters, growth, sw_reduce, sw_proc, trans_count) multiplx_fn = functools.partial(ConvBasisMultiplexer, transformation_filters, filters, growth, sw_reduce, sw_proc, trans_count)
switches.append(ConfigurableSwitchComputer(transformation_filters, multiplx_fn, switches.append(ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
pre_transform_block=functools.partial(ConvBnLelu, transformation_filters, transformation_filters, bn=False, bias=False), pre_transform_block=functools.partial(ConvBnLelu, transformation_filters, transformation_filters, bn=False, bias=False),
transform_block=functools.partial(MultiConvBlock, transformation_filters, transformation_filters, transformation_filters, kernel_size=kernel, depth=layers), transform_block=functools.partial(MultiConvBlock, transformation_filters, transformation_filters + growth, transformation_filters, kernel_size=kernel, depth=layers),
transform_count=trans_count, init_temp=initial_temp, enable_negative_transforms=enable_negative_transforms, transform_count=trans_count, init_temp=initial_temp, enable_negative_transforms=enable_negative_transforms,
add_scalable_noise_to_transforms=add_scalable_noise_to_transforms, init_scalar=1)) add_scalable_noise_to_transforms=add_scalable_noise_to_transforms, init_scalar=.2))
# Initialize the transforms with a lesser weight, since they are repeatedly added on to the resultant image.
initialize_weights([s.transforms for s in switches], .2 / len(switches))
self.switches = nn.ModuleList(switches) self.switches = nn.ModuleList(switches)
self.transformation_counts = trans_counts self.transformation_counts = trans_counts
@ -178,16 +180,18 @@ class ConfigurableSwitchedResidualGenerator2(nn.Module):
x = self.initial_conv(x) x = self.initial_conv(x)
self.attentions = [] self.attentions = []
swx = x
for i, sw in enumerate(self.switches): for i, sw in enumerate(self.switches):
x, att = sw.forward(x, True) swx, att = sw.forward(swx, True)
self.attentions.append(att) self.attentions.append(att)
x = swx + self.sw_conv(x)
if self.upsample_factor > 1: assert x == 2 or x == 4
x = F.interpolate(x, scale_factor=self.upsample_factor, mode="nearest") x = self.upconv1(F.interpolate(x, scale_factor=2, mode="nearest"))
if self.upsample_factor > 2:
x = self.proc_conv(x) x = F.interpolate(x, scale_factor=2, mode="nearest")
x = self.final_conv(x) x = self.upconv2(x)
return x, return self.final_conv(self.hr_conv(x)),
def set_temperature(self, temp): def set_temperature(self, temp):
[sw.set_temperature(temp) for sw in self.switches] [sw.set_temperature(temp) for sw in self.switches]