forked from mrq/DL-Art-School
Clean up unused archs
This commit is contained in:
James Betker
parent
e8613041c0
commit
55475d2ac1
|
|
@ -473,127 +473,3 @@ class SwitchedSpsrWithRef(nn.Module):
|
|||
val["switch_%i_specificity" % (i,)] = means[i]
|
||||
val["switch_%i_histogram" % (i,)] = hists[i]
|
||||
return val
|
||||
|
||||
|
||||
class SwitchedSpsrWithRef4x(nn.Module):
|
||||
def __init__(self, in_nc, out_nc, nf, xforms=8, upscale=4, init_temperature=10):
|
||||
super(SwitchedSpsrWithRef4x, self).__init__()
|
||||
n_upscale = int(math.log(upscale, 2))
|
||||
|
||||
# switch options
|
||||
transformation_filters = nf
|
||||
switch_filters = nf
|
||||
self.transformation_counts = xforms
|
||||
self.reference_processor = ReferenceImageBranch(transformation_filters)
|
||||
multiplx_fn = functools.partial(ReferencingConvMultiplexer, transformation_filters, switch_filters, self.transformation_counts)
|
||||
pretransform_fn = functools.partial(AdaInConvBlock, 512, transformation_filters, transformation_filters)
|
||||
transform_fn = functools.partial(MultiConvBlock, transformation_filters, int(transformation_filters * 1.5),
|
||||
transformation_filters, kernel_size=3, depth=3,
|
||||
weight_init_factor=.1)
|
||||
|
||||
# Feature branch
|
||||
self.model_fea_conv = ConvGnLelu(in_nc, nf, kernel_size=3, norm=False, activation=False)
|
||||
self.sw1 = ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
|
||||
pre_transform_block=pretransform_fn, transform_block=transform_fn,
|
||||
attention_norm=True,
|
||||
transform_count=self.transformation_counts, init_temp=init_temperature,
|
||||
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=self.transformation_counts, init_temp=init_temperature,
|
||||
add_scalable_noise_to_transforms=True)
|
||||
self.feature_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=False)
|
||||
self.stage1_up_fea = UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=False, bias=False)
|
||||
self.feature_hr_conv2 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False, bias=False)
|
||||
|
||||
# Grad branch
|
||||
self.get_g_nopadding = ImageGradientNoPadding()
|
||||
self.b_fea_conv = ConvGnLelu(in_nc, nf, kernel_size=3, norm=False, activation=False, bias=False)
|
||||
mplex_grad = functools.partial(ReferencingConvMultiplexer, nf * 2, nf * 2, self.transformation_counts // 2)
|
||||
self.sw_grad = ConfigurableSwitchComputer(transformation_filters, mplex_grad,
|
||||
pre_transform_block=pretransform_fn, transform_block=transform_fn,
|
||||
attention_norm=True,
|
||||
transform_count=self.transformation_counts // 2, init_temp=init_temperature,
|
||||
add_scalable_noise_to_transforms=True)
|
||||
self.stage1_up_grad = UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=False, bias=False)
|
||||
|
||||
# Upsampling
|
||||
self.grad_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=True, bias=False)
|
||||
self.grad_hr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False, bias=False)
|
||||
# Conv used to output grad branch shortcut.
|
||||
self.grad_branch_output_conv = ConvGnLelu(nf, out_nc, kernel_size=1, norm=False, activation=False, bias=False)
|
||||
|
||||
# Conjoin branch.
|
||||
transform_fn_cat = functools.partial(MultiConvBlock, transformation_filters * 2, int(transformation_filters * 1.5),
|
||||
transformation_filters, kernel_size=3, depth=4,
|
||||
weight_init_factor=.1)
|
||||
pretransform_fn_cat = functools.partial(AdaInConvBlock, 512, transformation_filters * 2, transformation_filters * 2)
|
||||
self._branch_pretrain_sw = ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
|
||||
pre_transform_block=pretransform_fn_cat, transform_block=transform_fn_cat,
|
||||
attention_norm=True,
|
||||
transform_count=self.transformation_counts, init_temp=init_temperature,
|
||||
add_scalable_noise_to_transforms=True)
|
||||
self.stage2_up_fea = UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=False, bias=False)
|
||||
self.stage2_up_grad = UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=False, bias=False)
|
||||
self.final_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=False)
|
||||
self.final_hr_conv1 = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=True, bias=False)
|
||||
self.final_hr_conv2 = ConvGnLelu(nf, out_nc, kernel_size=3, norm=False, activation=False, bias=False)
|
||||
self.switches = [self.sw1, self.sw2, self.sw_grad, self._branch_pretrain_sw]
|
||||
self.attentions = None
|
||||
self.init_temperature = init_temperature
|
||||
self.final_temperature_step = 10000
|
||||
|
||||
def forward(self, x, ref, center_coord):
|
||||
x_grad = self.get_g_nopadding(x)
|
||||
ref = self.reference_processor(ref, center_coord)
|
||||
x = self.model_fea_conv(x)
|
||||
|
||||
x1, a1 = self.sw1((x, ref), True)
|
||||
x2, a2 = self.sw2((x1, ref), True)
|
||||
x_fea = self.feature_lr_conv(x2)
|
||||
x_fea = self.stage1_up_fea(x_fea)
|
||||
x_fea = self.feature_hr_conv2(x_fea)
|
||||
|
||||
x_b_fea = self.b_fea_conv(x_grad)
|
||||
x_grad, a3 = self.sw_grad((x_b_fea, ref), att_in=(torch.cat([x1, x_b_fea], dim=1), ref), output_attention_weights=True)
|
||||
x_grad = self.grad_lr_conv(x_grad)
|
||||
x_grad = self.stage1_up_grad(x_grad)
|
||||
x_grad = self.grad_hr_conv(x_grad)
|
||||
x_out_branch = self.stage2_up_grad(x_grad)
|
||||
x_out_branch = self.grad_branch_output_conv(x_out_branch)
|
||||
|
||||
x__branch_pretrain_cat = torch.cat([x_grad, x_fea], dim=1)
|
||||
x__branch_pretrain_cat, a4 = self._branch_pretrain_sw((x__branch_pretrain_cat, ref), att_in=(x_fea, ref), identity=x_fea, output_attention_weights=True)
|
||||
x_out = self.final_lr_conv(x__branch_pretrain_cat)
|
||||
x_out = self.stage2_up_fea(x_out)
|
||||
x_out = self.final_hr_conv1(x_out)
|
||||
x_out = self.final_hr_conv2(x_out)
|
||||
|
||||
self.attentions = [a1, a2, a3, a4]
|
||||
|
||||
return x_out_branch, x_out, x_grad
|
||||
|
||||
def set_temperature(self, temp):
|
||||
[sw.set_temperature(temp) for sw in self.switches]
|
||||
|
||||
def update_for_step(self, step, experiments_path='.'):
|
||||
if self.attentions:
|
||||
temp = max(1, 1 + self.init_temperature *
|
||||
(self.final_temperature_step - step) / self.final_temperature_step)
|
||||
self.set_temperature(temp)
|
||||
if step % 200 == 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))]
|
||||
|
||||
def get_debug_values(self, step):
|
||||
temp = self.switches[0].switch.temperature
|
||||
mean_hists = [compute_attention_specificity(att, 2) for att in self.attentions]
|
||||
means = [i[0] for i in mean_hists]
|
||||
hists = [i[1].clone().detach().cpu().flatten() for i in mean_hists]
|
||||
val = {"switch_temperature": temp}
|
||||
for i in range(len(means)):
|
||||
val["switch_%i_specificity" % (i,)] = means[i]
|
||||
val["switch_%i_histogram" % (i,)] = hists[i]
|
||||
return val
|
||||
|
|
|
|||
|
|
@ -4,9 +4,7 @@ from switched_conv import BareConvSwitch, compute_attention_specificity, Attenti
|
|||
import torch.nn.functional as F
|
||||
import functools
|
||||
from collections import OrderedDict
|
||||
from models.archs.arch_util import ConvBnLelu, ConvGnSilu, ExpansionBlock, ExpansionBlock2, ConjoinBlock, ConvGnLelu
|
||||
from models.archs.RRDBNet_arch import ResidualDenseBlock_5C, RRDB
|
||||
from models.archs.spinenet_arch import SpineNet
|
||||
from models.archs.arch_util import ConvBnLelu, ConvGnSilu, ExpansionBlock, ExpansionBlock2, ConvGnLelu
|
||||
from switched_conv_util import save_attention_to_image_rgb
|
||||
import os
|
||||
|
||||
|
|
@ -81,17 +79,6 @@ class ConvBasisMultiplexer(nn.Module):
|
|||
return x
|
||||
|
||||
|
||||
class CachedBackboneWrapper:
|
||||
def __init__(self, backbone: nn.Module):
|
||||
self.backbone = backbone
|
||||
|
||||
def __call__(self, *args):
|
||||
self.cache = self.backbone(*args)
|
||||
return self.cache
|
||||
|
||||
def get_forward_result(self):
|
||||
return self.cache
|
||||
|
||||
# torch.gather() which operates across 2d images.
|
||||
def gather_2d(input, index):
|
||||
b, c, h, w = input.shape
|
||||
|
|
@ -187,26 +174,6 @@ class ReferencingConvMultiplexer(nn.Module):
|
|||
return x
|
||||
|
||||
|
||||
class BackboneMultiplexer(nn.Module):
|
||||
def __init__(self, backbone: CachedBackboneWrapper, transform_count):
|
||||
super(BackboneMultiplexer, self).__init__()
|
||||
self.backbone = backbone
|
||||
self.proc = nn.Sequential(ConvGnSilu(256, 256, kernel_size=3, bias=True),
|
||||
ConvGnSilu(256, 256, kernel_size=3, bias=False))
|
||||
self.up1 = nn.Sequential(ConvGnSilu(256, 128, kernel_size=3, bias=False, norm=False, activation=False),
|
||||
ConvGnSilu(128, 128, kernel_size=3, bias=False))
|
||||
self.up2 = nn.Sequential(ConvGnSilu(128, 64, kernel_size=3, bias=False, norm=False, activation=False),
|
||||
ConvGnSilu(64, 64, kernel_size=3, bias=False))
|
||||
self.final = ConvGnSilu(64, transform_count, bias=False, norm=False, activation=False)
|
||||
|
||||
def forward(self, x):
|
||||
spine = self.backbone.get_forward_result()
|
||||
feat = self.proc(spine[0])
|
||||
feat = self.up1(F.interpolate(feat, scale_factor=2, mode="nearest"))
|
||||
feat = self.up2(F.interpolate(feat, scale_factor=2, mode="nearest"))
|
||||
return self.final(feat)
|
||||
|
||||
|
||||
class ConfigurableSwitchComputer(nn.Module):
|
||||
def __init__(self, base_filters, multiplexer_net, pre_transform_block, transform_block, transform_count, attention_norm,
|
||||
init_temp=20, add_scalable_noise_to_transforms=False):
|
||||
|
|
@ -364,204 +331,3 @@ class ConfigurableSwitchedResidualGenerator2(nn.Module):
|
|||
val["switch_%i_histogram" % (i,)] = hists[i]
|
||||
return val
|
||||
|
||||
|
||||
# Equivalent to SRG2 - Uses RDB blocks in between two switches.
|
||||
class ConfigurableSwitchedResidualGenerator4(nn.Module):
|
||||
def __init__(self, switch_filters, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes,
|
||||
trans_layers, transformation_filters, attention_norm, initial_temp=20, final_temperature_step=50000, heightened_temp_min=1,
|
||||
heightened_final_step=50000, upsample_factor=1,
|
||||
add_scalable_noise_to_transforms=False):
|
||||
super(ConfigurableSwitchedResidualGenerator4, self).__init__()
|
||||
self.initial_conv = ConvBnLelu(3, transformation_filters, norm=False, activation=False, bias=True)
|
||||
self.upconv1 = ConvBnLelu(transformation_filters, transformation_filters, norm=False, bias=True)
|
||||
self.upconv2 = ConvBnLelu(transformation_filters, transformation_filters, norm=False, bias=True)
|
||||
self.hr_conv = ConvBnLelu(transformation_filters, transformation_filters, norm=False, bias=True)
|
||||
|
||||
multiplx_fn = functools.partial(ConvBasisMultiplexer, transformation_filters, switch_filters, switch_reductions,
|
||||
switch_processing_layers, trans_counts)
|
||||
half_multiplx_fn = functools.partial(ConvBasisMultiplexer, transformation_filters, switch_filters, switch_reductions,
|
||||
switch_processing_layers, trans_counts // 2)
|
||||
transform_fn = functools.partial(MultiConvBlock, transformation_filters, int(transformation_filters * 1.5),
|
||||
transformation_filters, kernel_size=trans_kernel_sizes, depth=trans_layers,
|
||||
weight_init_factor=.1)
|
||||
self.rdb1 = RRDB(transformation_filters)
|
||||
self.sw1 = ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
|
||||
pre_transform_block=None, transform_block=transform_fn,
|
||||
attention_norm=attention_norm,
|
||||
transform_count=trans_counts, init_temp=initial_temp,
|
||||
add_scalable_noise_to_transforms=add_scalable_noise_to_transforms)
|
||||
self.rdb2 = RRDB(transformation_filters)
|
||||
self.sw2 = ConfigurableSwitchComputer(transformation_filters, half_multiplx_fn,
|
||||
pre_transform_block=None, transform_block=transform_fn,
|
||||
attention_norm=attention_norm,
|
||||
transform_count=trans_counts // 2, init_temp=initial_temp,
|
||||
add_scalable_noise_to_transforms=add_scalable_noise_to_transforms)
|
||||
self.rdb3 = RRDB(transformation_filters)
|
||||
self.sw3 = ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
|
||||
pre_transform_block=None, transform_block=transform_fn,
|
||||
attention_norm=attention_norm,
|
||||
transform_count=trans_counts, init_temp=initial_temp,
|
||||
add_scalable_noise_to_transforms=add_scalable_noise_to_transforms)
|
||||
self.rdb4 = RRDB(transformation_filters)
|
||||
self.switches = [self.sw1, self.sw2, self.sw3]
|
||||
|
||||
self.final_conv = ConvBnLelu(transformation_filters, 3, norm=False, activation=False, bias=True)
|
||||
self.transformation_counts = trans_counts
|
||||
self.init_temperature = initial_temp
|
||||
self.final_temperature_step = final_temperature_step
|
||||
self.heightened_temp_min = heightened_temp_min
|
||||
self.heightened_final_step = heightened_final_step
|
||||
self.attentions = None
|
||||
self.upsample_factor = upsample_factor
|
||||
assert self.upsample_factor == 2 or self.upsample_factor == 4
|
||||
|
||||
def forward(self, x):
|
||||
# This is a common bug when evaluating SRG2 generators. It needs to be configured properly in eval mode. Just fail.
|
||||
if not self.train:
|
||||
assert self.switches[0].switch.temperature == 1
|
||||
|
||||
x = self.initial_conv(x)
|
||||
|
||||
x = self.rdb1(x)
|
||||
x, a1 = self.sw1(x, True)
|
||||
x = self.rdb2(x)
|
||||
x, a2 = self.sw2(x, True)
|
||||
x = self.rdb3(x)
|
||||
x, a3 = self.sw3(x, True)
|
||||
x = self.rdb4(x)
|
||||
self.attentions = [a1, a2, a3]
|
||||
|
||||
x = self.upconv1(F.interpolate(x, scale_factor=2, mode="nearest"))
|
||||
if self.upsample_factor > 2:
|
||||
x = F.interpolate(x, scale_factor=2, mode="nearest")
|
||||
x = self.upconv2(x)
|
||||
x = self.final_conv(self.hr_conv(x))
|
||||
return x, x
|
||||
|
||||
def set_temperature(self, temp):
|
||||
[sw.set_temperature(temp) for sw in self.switches]
|
||||
|
||||
def update_for_step(self, step, experiments_path='.'):
|
||||
if self.attentions:
|
||||
temp = max(1,
|
||||
1 + self.init_temperature * (self.final_temperature_step - step) / self.final_temperature_step)
|
||||
if temp == 1 and self.heightened_final_step and step > self.final_temperature_step and \
|
||||
self.heightened_final_step != 1:
|
||||
# Once the temperature passes (1) it enters an inverted curve to match the linear curve from above.
|
||||
# without this, the attention specificity "spikes" incredibly fast in the last few iterations.
|
||||
h_steps_total = self.heightened_final_step - self.final_temperature_step
|
||||
h_steps_current = min(step - self.final_temperature_step, h_steps_total)
|
||||
# The "gap" will represent the steps that need to be traveled as a linear function.
|
||||
h_gap = 1 / self.heightened_temp_min
|
||||
temp = h_gap * h_steps_current / h_steps_total
|
||||
# Invert temperature to represent reality on this side of the curve
|
||||
temp = 1 / temp
|
||||
self.set_temperature(temp)
|
||||
if step % 50 == 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))]
|
||||
|
||||
def get_debug_values(self, step):
|
||||
temp = self.switches[0].switch.temperature
|
||||
mean_hists = [compute_attention_specificity(att, 2) for att in self.attentions]
|
||||
means = [i[0] for i in mean_hists]
|
||||
hists = [i[1].clone().detach().cpu().flatten() for i in mean_hists]
|
||||
val = {"switch_temperature": temp}
|
||||
for i in range(len(means)):
|
||||
val["switch_%i_specificity" % (i,)] = means[i]
|
||||
val["switch_%i_histogram" % (i,)] = hists[i]
|
||||
return val
|
||||
|
||||
class Interpolate(nn.Module):
|
||||
def __init__(self, factor):
|
||||
super(Interpolate, self).__init__()
|
||||
self.factor = factor
|
||||
|
||||
def forward(self, x):
|
||||
return F.interpolate(x, scale_factor=self.factor)
|
||||
|
||||
|
||||
class ConfigurableSwitchedResidualGenerator3(nn.Module):
|
||||
def __init__(self, base_filters, trans_count, initial_temp=20, final_temperature_step=50000,
|
||||
heightened_temp_min=1,
|
||||
heightened_final_step=50000, upsample_factor=4):
|
||||
super(ConfigurableSwitchedResidualGenerator3, self).__init__()
|
||||
self.initial_conv = ConvBnLelu(3, base_filters, norm=False, activation=False, bias=True)
|
||||
self.sw_conv = ConvBnLelu(base_filters, base_filters, activation=False, bias=True)
|
||||
self.upconv1 = ConvBnLelu(base_filters, base_filters, norm=False, bias=True)
|
||||
self.upconv2 = ConvBnLelu(base_filters, base_filters, norm=False, bias=True)
|
||||
self.hr_conv = ConvBnLelu(base_filters, base_filters, norm=False, bias=True)
|
||||
self.final_conv = ConvBnLelu(base_filters, 3, norm=False, activation=False, bias=True)
|
||||
|
||||
self.backbone = SpineNet('49', in_channels=3, use_input_norm=True)
|
||||
for p in self.backbone.parameters(recurse=True):
|
||||
p.requires_grad = False
|
||||
self.backbone_wrapper = CachedBackboneWrapper(self.backbone)
|
||||
multiplx_fn = functools.partial(BackboneMultiplexer, self.backbone_wrapper)
|
||||
pretransform_fn = functools.partial(nn.Sequential, ConvBnLelu(base_filters, base_filters, kernel_size=3, norm=False, activation=False, bias=False))
|
||||
transform_fn = functools.partial(MultiConvBlock, base_filters, int(base_filters * 1.5), base_filters, kernel_size=3, depth=4)
|
||||
self.switch = ConfigurableSwitchComputer(base_filters, multiplx_fn, pretransform_fn, transform_fn, trans_count, init_temp=initial_temp,
|
||||
add_scalable_noise_to_transforms=True, init_scalar=.1)
|
||||
|
||||
self.transformation_counts = trans_count
|
||||
self.init_temperature = initial_temp
|
||||
self.final_temperature_step = final_temperature_step
|
||||
self.heightened_temp_min = heightened_temp_min
|
||||
self.heightened_final_step = heightened_final_step
|
||||
self.attentions = None
|
||||
self.upsample_factor = upsample_factor
|
||||
self.backbone_forward = None
|
||||
|
||||
def get_forward_results(self):
|
||||
return self.backbone_forward
|
||||
|
||||
def forward(self, x):
|
||||
self.backbone_forward = self.backbone_wrapper(F.interpolate(x, scale_factor=2, mode="nearest"))
|
||||
|
||||
x = self.initial_conv(x)
|
||||
|
||||
self.attentions = []
|
||||
x, att = self.switch(x, output_attention_weights=True)
|
||||
self.attentions.append(att)
|
||||
|
||||
x = self.upconv1(F.interpolate(x, scale_factor=2, mode="nearest"))
|
||||
if self.upsample_factor > 2:
|
||||
x = F.interpolate(x, scale_factor=2, mode="nearest")
|
||||
x = self.upconv2(x)
|
||||
return self.final_conv(self.hr_conv(x)),
|
||||
|
||||
def set_temperature(self, temp):
|
||||
self.switch.set_temperature(temp)
|
||||
|
||||
def update_for_step(self, step, experiments_path='.'):
|
||||
if self.attentions:
|
||||
temp = max(1,
|
||||
1 + self.init_temperature * (self.final_temperature_step - step) / self.final_temperature_step)
|
||||
if temp == 1 and self.heightened_final_step and step > self.final_temperature_step and \
|
||||
self.heightened_final_step != 1:
|
||||
# Once the temperature passes (1) it enters an inverted curve to match the linear curve from above.
|
||||
# without this, the attention specificity "spikes" incredibly fast in the last few iterations.
|
||||
h_steps_total = self.heightened_final_step - self.final_temperature_step
|
||||
h_steps_current = min(step - self.final_temperature_step, h_steps_total)
|
||||
# The "gap" will represent the steps that need to be traveled as a linear function.
|
||||
h_gap = 1 / self.heightened_temp_min
|
||||
temp = h_gap * h_steps_current / h_steps_total
|
||||
# Invert temperature to represent reality on this side of the curve
|
||||
temp = 1 / temp
|
||||
self.set_temperature(temp)
|
||||
if step % 50 == 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))]
|
||||
|
||||
def get_debug_values(self, step):
|
||||
temp = self.switch.switch.temperature
|
||||
mean_hists = [compute_attention_specificity(att, 2) for att in self.attentions]
|
||||
means = [i[0] for i in mean_hists]
|
||||
hists = [i[1].clone().detach().cpu().flatten() for i in mean_hists]
|
||||
val = {"switch_temperature": temp}
|
||||
for i in range(len(means)):
|
||||
val["switch_%i_specificity" % (i,)] = means[i]
|
||||
val["switch_%i_histogram" % (i,)] = hists[i]
|
||||
return val
|
||||
|
|
|
|||
|
|
@ -50,21 +50,6 @@ def define_G(opt, net_key='network_G', scale=None):
|
|||
initial_temp=opt_net['temperature'], final_temperature_step=opt_net['temperature_final_step'],
|
||||
heightened_temp_min=opt_net['heightened_temp_min'], heightened_final_step=opt_net['heightened_final_step'],
|
||||
upsample_factor=scale, add_scalable_noise_to_transforms=opt_net['add_noise'])
|
||||
elif which_model == "ConfigurableSwitchedResidualGenerator4":
|
||||
netG = SwitchedGen_arch.ConfigurableSwitchedResidualGenerator4(switch_filters=opt_net['switch_filters'],
|
||||
switch_reductions=opt_net['switch_reductions'],
|
||||
switch_processing_layers=opt_net['switch_processing_layers'], trans_counts=opt_net['trans_counts'],
|
||||
trans_kernel_sizes=opt_net['trans_kernel_sizes'], trans_layers=opt_net['trans_layers'],
|
||||
transformation_filters=opt_net['transformation_filters'], attention_norm=opt_net['attention_norm'],
|
||||
initial_temp=opt_net['temperature'], final_temperature_step=opt_net['temperature_final_step'],
|
||||
heightened_temp_min=opt_net['heightened_temp_min'], heightened_final_step=opt_net['heightened_final_step'],
|
||||
upsample_factor=scale, add_scalable_noise_to_transforms=opt_net['add_noise'])
|
||||
elif which_model == 'spsr_net':
|
||||
netG = spsr.SPSRNet(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'], nf=opt_net['nf'],
|
||||
nb=opt_net['nb'], gc=opt_net['gc'], upscale=opt_net['scale'], norm_type=opt_net['norm_type'],
|
||||
act_type='leakyrelu', mode=opt_net['mode'], upsample_mode='upconv', bl_inc=opt_net['bl_inc'])
|
||||
if opt['is_train']:
|
||||
arch_util.initialize_weights(netG, scale=.1)
|
||||
elif which_model == 'spsr_net_improved':
|
||||
netG = spsr.SPSRNetSimplified(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'], nf=opt_net['nf'],
|
||||
nb=opt_net['nb'], upscale=opt_net['scale'])
|
||||
|
|
@ -78,19 +63,8 @@ def define_G(opt, net_key='network_G', scale=None):
|
|||
init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10)
|
||||
elif which_model == "spsr_switched_with_ref4x":
|
||||
xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8
|
||||
netG = spsr.SwitchedSpsrWithRef4x(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'],
|
||||
netG = spsr.SwitchedSpsrWithRef4x(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms,
|
||||
init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10)
|
||||
|
||||
# image corruption
|
||||
elif which_model == 'HighToLowResNet':
|
||||
netG = HighToLowResNet.HighToLowResNet(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'],
|
||||
nf=opt_net['nf'], nb=opt_net['nb'], downscale=opt_net['scale'])
|
||||
elif which_model == 'FlatProcessorNet':
|
||||
'''netG = FlatProcessorNet_arch.FlatProcessorNet(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'],
|
||||
nf=opt_net['nf'], downscale=opt_net['scale'], reduce_anneal_blocks=opt_net['ra_blocks'],
|
||||
assembler_blocks=opt_net['assembler_blocks'])'''
|
||||
netG = FlatProcessorNetNew_arch.fixup_resnet34(num_filters=opt_net['nf'])\
|
||||
|
||||
else:
|
||||
raise NotImplementedError('Generator model [{:s}] not recognized'.format(which_model))
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue
Block a user