a-One-Fan/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

More attention fixes for switched_spsr

Browse Source
This commit is contained in:
James Betker 2020-08-07 21:11:50 -06:00
parent d02509ef97
commit 4e972144ae
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
codes/models/archs/SPSR_arch.py
View File

@ -439,9 +439,9 @@ class SwitchedSpsr(nn.Module):
switch_filters = nf
switch_reductions = 3
switch_processing_layers = 2
trans_counts = 8
self.transformation_counts = 8
multiplx_fn = functools.partial(ConvBasisMultiplexer, transformation_filters, switch_filters, switch_reductions,
switch_processing_layers, trans_counts)
switch_processing_layers, self.transformation_counts)
pretransform_fn = functools.partial(ConvGnLelu, transformation_filters, transformation_filters, norm=False, bias=False, weight_init_factor=.1)
transform_fn = functools.partial(MultiConvBlock, transformation_filters, int(transformation_filters * 1.5),
transformation_filters, kernel_size=3, depth=3,
@ -452,12 +452,12 @@ class SwitchedSpsr(nn.Module):
self.sw1 = ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
pre_transform_block=pretransform_fn, transform_block=transform_fn,
attention_norm=True,
transform_count=trans_counts, init_temp=10,
transform_count=self.transformation_counts, init_temp=10,
add_scalable_noise_to_transforms=True)
self.sw2 = ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
pre_transform_block=pretransform_fn, transform_block=transform_fn,
attention_norm=True,
transform_count=trans_counts, init_temp=10,
transform_count=self.transformation_counts, init_temp=10,
add_scalable_noise_to_transforms=True)
self.feature_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False)
self.model_upsampler = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=False, bias=False) for _ in range(n_upscale)])
@ -470,7 +470,7 @@ class SwitchedSpsr(nn.Module):
self.sw_grad = ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
pre_transform_block=pretransform_fn, transform_block=transform_fn,
attention_norm=True,
transform_count=trans_counts, init_temp=10,
transform_count=self.transformation_counts, init_temp=10,
add_scalable_noise_to_transforms=True)
# Upsampling
self.grad_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=False)
@ -487,7 +487,7 @@ class SwitchedSpsr(nn.Module):
self._branch_pretrain_sw = ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
pre_transform_block=pretransform_fn, transform_block=transform_fn,
attention_norm=True,
transform_count=trans_counts, init_temp=10,
transform_count=self.transformation_counts, init_temp=10,
add_scalable_noise_to_transforms=True)
self._branch_pretrain_HR_conv0 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=False)
self._branch_pretrain_HR_conv1 = ConvGnLelu(nf, out_nc, kernel_size=3, norm=False, activation=False, bias=False)
@ -531,7 +531,7 @@ class SwitchedSpsr(nn.Module):
temp = max(1, 1 + self.init_temperature *
(self.final_temperature_step - step) / self.final_temperature_step)
self.set_temperature(temp)
if step % 50 == 0:
if step % 10 == 0:
output_path = os.path.join(experiments_path, "attention_maps", "a%i")
prefix = "attention_map_%i_%%i.png" % (step,)
[save_attention_to_image_rgb(output_path % (i,), self.attentions[i], self.transformation_counts, prefix, step) for i in range(len(self.attentions))]