mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix scaling bug

Browse Source
This commit is contained in:
James Betker 2020-07-01 16:42:27 -06:00
parent 30653181ba
commit 6ac6c95177
1 changed files with 7 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
codes/models/archs/SwitchedResidualGenerator_arch.py
View File

@ -192,8 +192,10 @@ class ConfigurableSwitchComputer(nn.Module):
# And the switch itself, including learned scalars # And the switch itself, including learned scalars
self.switch = BareConvSwitch(initial_temperature=init_temp) self.switch = BareConvSwitch(initial_temperature=init_temp)
self.switch_scale = nn.Parameter(torch.full((1,), float(init_scalar)))
self.post_switch_conv = ConvBnLelu(base_filters, base_filters, bn=False, bias=False) self.post_switch_conv = ConvBnLelu(base_filters, base_filters, bn=False, bias=False)
self.scale = nn.Parameter(torch.full((1,), float(init_scalar))) # The post_switch_conv gets a near-zero scale. The network can decide to magnify it (or not) depending on its needs.
self.psc_scale = nn.Parameter(torch.full((1,), float(1e-3)))
self.bias = nn.Parameter(torch.zeros(1)) self.bias = nn.Parameter(torch.zeros(1))
def forward(self, x, output_attention_weights=False): def forward(self, x, output_attention_weights=False):
@ -211,9 +213,9 @@ class ConfigurableSwitchComputer(nn.Module):
m = F.interpolate(m, size=x.shape[2:], mode='nearest') m = F.interpolate(m, size=x.shape[2:], mode='nearest')
outputs, attention = self.switch(xformed, m, True) outputs, attention = self.switch(xformed, m, True)
outputs = identity + outputs outputs = identity + outputs * self.switch_scale
#outputs = identity + self.post_switch_conv(outputs) outputs = identity + self.post_switch_conv(outputs) * self.psc_scale
outputs = outputs * self.scale + self.bias outputs = outputs + self.bias
if output_attention_weights: if output_attention_weights:
return outputs, attention return outputs, attention
else: else:
@ -361,11 +363,7 @@ class ConfigurableSwitchedResidualGenerator2(nn.Module):
switches.append(ConfigurableSwitchComputer(transformation_filters, multiplx_fn, switches.append(ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
functools.partial(MultiConvBlock, transformation_filters, transformation_filters, transformation_filters, kernel_size=kernel, depth=layers), functools.partial(MultiConvBlock, transformation_filters, transformation_filters, transformation_filters, kernel_size=kernel, depth=layers),
trans_count, initial_temp, enable_negative_transforms=enable_negative_transforms, trans_count, 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=.01))
initialize_weights(switches, 1)
# 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