745 lines
38 KiB
Python
745 lines
38 KiB
Python
import torch
|
|
from torch import nn
|
|
from switched_conv.switched_conv import BareConvSwitch, compute_attention_specificity, AttentionNorm
|
|
import torch.nn.functional as F
|
|
import functools
|
|
from collections import OrderedDict
|
|
from models.archs.arch_util import ConvBnLelu, ConvGnSilu, ExpansionBlock, ExpansionBlock2, ConvGnLelu, MultiConvBlock, \
|
|
SiLU, UpconvBlock, ReferenceJoinBlock
|
|
from switched_conv.switched_conv_util import save_attention_to_image_rgb
|
|
import os
|
|
from models.archs.spinenet_arch import SpineNet
|
|
import torchvision
|
|
from utils.util import checkpoint
|
|
|
|
# VGG-style layer with Conv(stride2)->BN->Activation->Conv->BN->Activation
|
|
# Doubles the input filter count.
|
|
class HalvingProcessingBlock(nn.Module):
|
|
def __init__(self, filters):
|
|
super(HalvingProcessingBlock, self).__init__()
|
|
self.bnconv1 = ConvGnSilu(filters, filters * 2, stride=2, norm=False, bias=False)
|
|
self.bnconv2 = ConvGnSilu(filters * 2, filters * 2, norm=True, bias=False)
|
|
|
|
def forward(self, x):
|
|
x = self.bnconv1(x)
|
|
return self.bnconv2(x)
|
|
|
|
|
|
# This is a classic u-net architecture with the goal of assigning each individual pixel an individual transform
|
|
# switching set.
|
|
class ConvBasisMultiplexer(nn.Module):
|
|
def __init__(self, input_channels, base_filters, reductions, processing_depth, multiplexer_channels, use_gn=True, use_exp2=False):
|
|
super(ConvBasisMultiplexer, self).__init__()
|
|
self.filter_conv = ConvGnSilu(input_channels, base_filters, bias=True)
|
|
self.reduction_blocks = nn.ModuleList([HalvingProcessingBlock(base_filters * 2 ** i) for i in range(reductions)])
|
|
reduction_filters = base_filters * 2 ** reductions
|
|
self.processing_blocks = nn.Sequential(OrderedDict([('block%i' % (i,), ConvGnSilu(reduction_filters, reduction_filters, bias=False)) for i in range(processing_depth)]))
|
|
if use_exp2:
|
|
self.expansion_blocks = nn.ModuleList([ExpansionBlock2(reduction_filters // (2 ** i)) for i in range(reductions)])
|
|
else:
|
|
self.expansion_blocks = nn.ModuleList([ExpansionBlock(reduction_filters // (2 ** i)) for i in range(reductions)])
|
|
|
|
gap = base_filters - multiplexer_channels
|
|
cbl1_out = ((base_filters - (gap // 2)) // 4) * 4 # Must be multiples of 4 to use with group norm.
|
|
self.cbl1 = ConvGnSilu(base_filters, cbl1_out, norm=use_gn, bias=False, num_groups=4)
|
|
cbl2_out = ((base_filters - (3 * gap // 4)) // 4) * 4
|
|
self.cbl2 = ConvGnSilu(cbl1_out, cbl2_out, norm=use_gn, bias=False, num_groups=4)
|
|
self.cbl3 = ConvGnSilu(cbl2_out, multiplexer_channels, bias=True, norm=False)
|
|
|
|
def forward(self, x):
|
|
x = self.filter_conv(x)
|
|
reduction_identities = []
|
|
for b in self.reduction_blocks:
|
|
reduction_identities.append(x)
|
|
x = b(x)
|
|
x = self.processing_blocks(x)
|
|
for i, b in enumerate(self.expansion_blocks):
|
|
x = b(x, reduction_identities[-i - 1])
|
|
|
|
x = self.cbl1(x)
|
|
x = self.cbl2(x)
|
|
x = self.cbl3(x)
|
|
return x
|
|
|
|
|
|
# torch.gather() which operates across 2d images.
|
|
def gather_2d(input, index):
|
|
b, c, h, w = input.shape
|
|
nodim = input.view(b, c, h * w)
|
|
ind_nd = index[:, 0]*w + index[:, 1]
|
|
ind_nd = ind_nd.unsqueeze(1)
|
|
ind_nd = ind_nd.repeat((1, c))
|
|
ind_nd = ind_nd.unsqueeze(2)
|
|
result = torch.gather(nodim, dim=2, index=ind_nd)
|
|
result = result.squeeze()
|
|
if b == 1:
|
|
result = result.unsqueeze(0)
|
|
return result
|
|
|
|
|
|
class ConfigurableSwitchComputer(nn.Module):
|
|
def __init__(self, base_filters, multiplexer_net, pre_transform_block, transform_block, transform_count, attention_norm,
|
|
post_transform_block=None,
|
|
init_temp=20, add_scalable_noise_to_transforms=False, feed_transforms_into_multiplexer=False, post_switch_conv=True,
|
|
anorm_multiplier=16):
|
|
super(ConfigurableSwitchComputer, self).__init__()
|
|
|
|
tc = transform_count
|
|
self.multiplexer = multiplexer_net(tc)
|
|
|
|
if pre_transform_block:
|
|
self.pre_transform = pre_transform_block()
|
|
else:
|
|
self.pre_transform = None
|
|
self.transforms = nn.ModuleList([transform_block() for _ in range(transform_count)])
|
|
self.add_noise = add_scalable_noise_to_transforms
|
|
self.feed_transforms_into_multiplexer = feed_transforms_into_multiplexer
|
|
self.noise_scale = nn.Parameter(torch.full((1,), float(1e-3)))
|
|
|
|
# And the switch itself, including learned scalars
|
|
self.switch = BareConvSwitch(initial_temperature=init_temp, attention_norm=AttentionNorm(transform_count, accumulator_size=anorm_multiplier * transform_count) if attention_norm else None)
|
|
self.switch_scale = nn.Parameter(torch.full((1,), float(1)))
|
|
self.post_transform_block = post_transform_block
|
|
if post_switch_conv:
|
|
self.post_switch_conv = ConvBnLelu(base_filters, base_filters, norm=False, bias=True)
|
|
# The post_switch_conv gets a low scale initially. The network can decide to magnify it (or not)
|
|
# depending on its needs.
|
|
self.psc_scale = nn.Parameter(torch.full((1,), float(.1)))
|
|
else:
|
|
self.post_switch_conv = None
|
|
self.update_norm = True
|
|
|
|
def set_update_attention_norm(self, set_val):
|
|
self.update_norm = set_val
|
|
|
|
# Regarding inputs: it is acceptable to pass in a tuple/list as an input for (x), but the first element
|
|
# *must* be the actual parameter that gets fed through the network - it is assumed to be the identity.
|
|
def forward(self, x, att_in=None, identity=None, output_attention_weights=True, fixed_scale=1, do_checkpointing=False,
|
|
output_att_logits=False):
|
|
if isinstance(x, tuple):
|
|
x1 = x[0]
|
|
else:
|
|
x1 = x
|
|
|
|
if att_in is None:
|
|
att_in = x
|
|
|
|
if identity is None:
|
|
identity = x1
|
|
|
|
if self.add_noise:
|
|
rand_feature = torch.randn_like(x1) * self.noise_scale
|
|
if isinstance(x, tuple):
|
|
x = (x1 + rand_feature,) + x[1:]
|
|
else:
|
|
x = x1 + rand_feature
|
|
|
|
if not isinstance(x, tuple):
|
|
x = (x,)
|
|
if self.pre_transform:
|
|
x = self.pre_transform(*x)
|
|
if not isinstance(x, tuple):
|
|
x = (x,)
|
|
if do_checkpointing:
|
|
xformed = [checkpoint(t, *x) for t in self.transforms]
|
|
else:
|
|
xformed = [t(*x) for t in self.transforms]
|
|
|
|
if not isinstance(att_in, tuple):
|
|
att_in = (att_in,)
|
|
if self.feed_transforms_into_multiplexer:
|
|
att_in = att_in + (torch.stack(xformed, dim=1),)
|
|
if do_checkpointing:
|
|
m = checkpoint(self.multiplexer, *att_in)
|
|
else:
|
|
m = self.multiplexer(*att_in)
|
|
|
|
# It is assumed that [xformed] and [m] are collapsed into tensors at this point.
|
|
outputs, attention, att_logits = self.switch(xformed, m, True, self.update_norm, output_attention_logits=True)
|
|
if self.post_transform_block is not None:
|
|
outputs = self.post_transform_block(outputs)
|
|
|
|
outputs = identity + outputs * self.switch_scale * fixed_scale
|
|
if self.post_switch_conv is not None:
|
|
outputs = outputs + self.post_switch_conv(outputs) * self.psc_scale * fixed_scale
|
|
if output_attention_weights:
|
|
if output_att_logits:
|
|
return outputs, attention, att_logits
|
|
else:
|
|
return outputs, attention
|
|
else:
|
|
return outputs
|
|
|
|
def set_temperature(self, temp):
|
|
self.switch.set_attention_temperature(temp)
|
|
|
|
|
|
class ConfigurableSwitchedResidualGenerator2(nn.Module):
|
|
def __init__(self, switch_depth, 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(ConfigurableSwitchedResidualGenerator2, self).__init__()
|
|
switches = []
|
|
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)
|
|
self.final_conv = ConvBnLelu(transformation_filters, 3, norm=False, activation=False, bias=True)
|
|
for _ in range(switch_depth):
|
|
multiplx_fn = functools.partial(ConvBasisMultiplexer, transformation_filters, switch_filters, switch_reductions, switch_processing_layers, trans_counts)
|
|
pretransform_fn = functools.partial(ConvBnLelu, 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=trans_kernel_sizes, depth=trans_layers, weight_init_factor=.1)
|
|
switches.append(ConfigurableSwitchComputer(transformation_filters, multiplx_fn,
|
|
pre_transform_block=pretransform_fn, 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.switches = nn.ModuleList(switches)
|
|
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)
|
|
|
|
self.attentions = []
|
|
for i, sw in enumerate(self.switches):
|
|
x, att = sw.forward(x, 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)
|
|
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
|
|
|
|
|
|
# This class encapsulates an encoder based on an object detection network backbone whose purpose is to generated a
|
|
# structured embedding encoding what is in an image patch. This embedding can then be used to perform structured
|
|
# alterations to the underlying image.
|
|
#
|
|
# Caveat: Since this uses a pre-defined (and potentially pre-trained) SpineNet backbone, it has a minimum-supported
|
|
# image size, which is 128x128. In order to use 64x64 patches, you must set interpolate_first=True. though this will
|
|
# degrade quality.
|
|
class BackboneEncoder(nn.Module):
|
|
def __init__(self, interpolate_first=True, pretrained_backbone=None):
|
|
super(BackboneEncoder, self).__init__()
|
|
self.interpolate_first = interpolate_first
|
|
|
|
# Uses dual spinenets, one for the input patch and the other for the reference image.
|
|
self.patch_spine = SpineNet('49', in_channels=3, use_input_norm=True)
|
|
self.ref_spine = SpineNet('49', in_channels=3, use_input_norm=True)
|
|
|
|
self.merge_process1 = ConvGnSilu(512, 512, kernel_size=1, activation=True, norm=False, bias=True)
|
|
self.merge_process2 = ConvGnSilu(512, 384, kernel_size=1, activation=True, norm=True, bias=False)
|
|
self.merge_process3 = ConvGnSilu(384, 256, kernel_size=1, activation=False, norm=False, bias=True)
|
|
|
|
if pretrained_backbone is not None:
|
|
loaded_params = torch.load(pretrained_backbone)
|
|
self.ref_spine.load_state_dict(loaded_params['state_dict'], strict=True)
|
|
self.patch_spine.load_state_dict(loaded_params['state_dict'], strict=True)
|
|
|
|
# Returned embedding will have been reduced in size by a factor of 8 (4 if interpolate_first=True).
|
|
# Output channels are always 256.
|
|
# ex, 64x64 input with interpolate_first=True will result in tensor of shape [bx256x16x16]
|
|
def forward(self, x, ref, ref_center_point):
|
|
if self.interpolate_first:
|
|
x = F.interpolate(x, scale_factor=2, mode="bicubic")
|
|
# Don't interpolate ref - assume it is fed in at the proper resolution.
|
|
# ref = F.interpolate(ref, scale_factor=2, mode="bicubic")
|
|
|
|
# [ref] will have a 'mask' channel which we cannot use with pretrained spinenet.
|
|
ref = ref[:, :3, :, :]
|
|
ref_emb = self.ref_spine(ref)[0]
|
|
ref_code = gather_2d(ref_emb, ref_center_point // 8) # Divide by 8 to bring the center point to the correct location.
|
|
|
|
patch = self.patch_spine(x)[0]
|
|
ref_code_expanded = ref_code.view(-1, 256, 1, 1).repeat(1, 1, patch.shape[2], patch.shape[3])
|
|
combined = self.merge_process1(torch.cat([patch, ref_code_expanded], dim=1))
|
|
combined = self.merge_process2(combined)
|
|
combined = self.merge_process3(combined)
|
|
|
|
return combined
|
|
|
|
|
|
class BackboneEncoderNoRef(nn.Module):
|
|
def __init__(self, interpolate_first=True, pretrained_backbone=None):
|
|
super(BackboneEncoderNoRef, self).__init__()
|
|
self.interpolate_first = interpolate_first
|
|
|
|
self.patch_spine = SpineNet('49', in_channels=3, use_input_norm=True)
|
|
|
|
if pretrained_backbone is not None:
|
|
loaded_params = torch.load(pretrained_backbone)
|
|
self.patch_spine.load_state_dict(loaded_params['state_dict'], strict=True)
|
|
|
|
# Returned embedding will have been reduced in size by a factor of 8 (4 if interpolate_first=True).
|
|
# Output channels are always 256.
|
|
# ex, 64x64 input with interpolate_first=True will result in tensor of shape [bx256x16x16]
|
|
def forward(self, x):
|
|
if self.interpolate_first:
|
|
x = F.interpolate(x, scale_factor=2, mode="bicubic")
|
|
|
|
patch = self.patch_spine(x)[0]
|
|
return patch
|
|
|
|
|
|
class BackboneSpinenetNoHead(nn.Module):
|
|
def __init__(self):
|
|
super(BackboneSpinenetNoHead, self).__init__()
|
|
# Uses dual spinenets, one for the input patch and the other for the reference image.
|
|
self.patch_spine = SpineNet('49', in_channels=3, use_input_norm=False, double_reduce_early=False)
|
|
self.ref_spine = SpineNet('49', in_channels=4, use_input_norm=False, double_reduce_early=False)
|
|
|
|
self.merge_process1 = ConvGnSilu(512, 512, kernel_size=1, activation=True, norm=False, bias=True)
|
|
self.merge_process2 = ConvGnSilu(512, 384, kernel_size=1, activation=True, norm=True, bias=False)
|
|
self.merge_process3 = ConvGnSilu(384, 256, kernel_size=1, activation=False, norm=False, bias=True)
|
|
|
|
def forward(self, x, ref, ref_center_point):
|
|
ref_emb = self.ref_spine(ref)[0]
|
|
ref_code = gather_2d(ref_emb, ref_center_point // 4) # Divide by 8 to bring the center point to the correct location.
|
|
|
|
patch = self.patch_spine(x)[0]
|
|
ref_code_expanded = ref_code.view(-1, 256, 1, 1).repeat(1, 1, patch.shape[2], patch.shape[3])
|
|
combined = self.merge_process1(torch.cat([patch, ref_code_expanded], dim=1))
|
|
combined = self.merge_process2(combined)
|
|
combined = self.merge_process3(combined)
|
|
return combined
|
|
|
|
|
|
class ResBlock(nn.Module):
|
|
def __init__(self, nf, downsample):
|
|
super(ResBlock, self).__init__()
|
|
nf_int = nf * 2
|
|
nf_out = nf * 2 if downsample else nf
|
|
stride = 2 if downsample else 1
|
|
self.c1 = ConvGnSilu(nf, nf_int, kernel_size=3, bias=False, activation=True, norm=True)
|
|
self.c2 = ConvGnSilu(nf_int, nf_int, stride=stride, kernel_size=3, bias=False, activation=True, norm=True)
|
|
self.c3 = ConvGnSilu(nf_int, nf_out, kernel_size=3, bias=False, activation=False, norm=True)
|
|
if downsample:
|
|
self.downsample = ConvGnSilu(nf, nf_out, kernel_size=1, stride=stride, bias=False, activation=False, norm=True)
|
|
else:
|
|
self.downsample = None
|
|
self.act = SiLU()
|
|
|
|
def forward(self, x):
|
|
identity = x
|
|
branch = self.c1(x)
|
|
branch = self.c2(branch)
|
|
branch = self.c3(branch)
|
|
|
|
if self.downsample:
|
|
identity = self.downsample(identity)
|
|
return self.act(identity + branch)
|
|
|
|
|
|
class BackboneResnet(nn.Module):
|
|
def __init__(self):
|
|
super(BackboneResnet, self).__init__()
|
|
self.initial_conv = ConvGnSilu(3, 64, kernel_size=7, bias=True, activation=False, norm=False)
|
|
self.sequence = nn.Sequential(
|
|
ResBlock(64, downsample=False),
|
|
ResBlock(64, downsample=True),
|
|
ResBlock(128, downsample=False),
|
|
ResBlock(128, downsample=True),
|
|
ResBlock(256, downsample=False),
|
|
ResBlock(256, downsample=False))
|
|
|
|
def forward(self, x):
|
|
fea = self.initial_conv(x)
|
|
return self.sequence(fea)
|
|
|
|
|
|
# Computes a linear latent by performing processing on the reference image and returning the filters of a single point,
|
|
# which should be centered on the image patch being processed.
|
|
#
|
|
# Output is base_filters * 8.
|
|
class ReferenceImageBranch(nn.Module):
|
|
def __init__(self, base_filters=64):
|
|
super(ReferenceImageBranch, self).__init__()
|
|
self.features = nn.Sequential(ConvGnSilu(4, base_filters, kernel_size=7, bias=True),
|
|
HalvingProcessingBlock(base_filters),
|
|
ConvGnSilu(base_filters*2, base_filters*2, activation=True, norm=True, bias=False),
|
|
HalvingProcessingBlock(base_filters*2),
|
|
ConvGnSilu(base_filters*4, base_filters*4, activation=True, norm=True, bias=False),
|
|
HalvingProcessingBlock(base_filters*4),
|
|
ConvGnSilu(base_filters*8, base_filters*8, activation=True, norm=True, bias=False),
|
|
ConvGnSilu(base_filters*8, base_filters*8, activation=True, norm=True, bias=False))
|
|
|
|
# center_point is a [b,2] long tensor describing the center point of where the patch was taken from the reference
|
|
# image.
|
|
def forward(self, x, center_point):
|
|
x = self.features(x)
|
|
return gather_2d(x, center_point // 8) # Divide by 8 to scale the center_point down.
|
|
|
|
|
|
# Mutiplexer that combines a structured embedding with a contextual switch input to guide alterations to that input.
|
|
#
|
|
# Implemented as basically a u-net which reduces the input into the same structural space as the embedding, combines the
|
|
# two, then expands back into the original feature space.
|
|
class EmbeddingMultiplexer(nn.Module):
|
|
# Note: reductions=2 if the encoder is using interpolated input, otherwise reductions=3.
|
|
def __init__(self, nf, multiplexer_channels, reductions=2):
|
|
super(EmbeddingMultiplexer, self).__init__()
|
|
self.embedding_process = MultiConvBlock(256, 256, 256, kernel_size=3, depth=3, norm=True)
|
|
|
|
self.filter_conv = ConvGnSilu(nf, nf, activation=True, norm=False, bias=True)
|
|
self.reduction_blocks = nn.ModuleList([HalvingProcessingBlock(nf * 2 ** i) for i in range(reductions)])
|
|
reduction_filters = nf * 2 ** reductions
|
|
self.processing_blocks = nn.Sequential(
|
|
ConvGnSilu(reduction_filters + 256, reduction_filters + 256, kernel_size=1, activation=True, norm=False, bias=True),
|
|
ConvGnSilu(reduction_filters + 256, reduction_filters + 128, kernel_size=1, activation=True, norm=True, bias=False),
|
|
ConvGnSilu(reduction_filters + 128, reduction_filters, kernel_size=3, activation=True, norm=True, bias=False),
|
|
ConvGnSilu(reduction_filters, reduction_filters, kernel_size=3, activation=True, norm=True, bias=False))
|
|
self.expansion_blocks = nn.ModuleList([ExpansionBlock2(reduction_filters // (2 ** i)) for i in range(reductions)])
|
|
|
|
gap = nf - multiplexer_channels
|
|
cbl1_out = ((nf - (gap // 2)) // 4) * 4 # Must be multiples of 4 to use with group norm.
|
|
self.cbl1 = ConvGnSilu(nf, cbl1_out, norm=True, bias=False, num_groups=4)
|
|
cbl2_out = ((nf - (3 * gap // 4)) // 4) * 4
|
|
self.cbl2 = ConvGnSilu(cbl1_out, cbl2_out, norm=True, bias=False, num_groups=4)
|
|
self.cbl3 = ConvGnSilu(cbl2_out, multiplexer_channels, bias=True, norm=False)
|
|
|
|
def forward(self, x, embedding):
|
|
x = self.filter_conv(x)
|
|
embedding = self.embedding_process(embedding)
|
|
|
|
reduction_identities = []
|
|
for b in self.reduction_blocks:
|
|
reduction_identities.append(x)
|
|
x = b(x)
|
|
x = self.processing_blocks(torch.cat([x, embedding], dim=1))
|
|
for i, b in enumerate(self.expansion_blocks):
|
|
x = b(x, reduction_identities[-i - 1])
|
|
|
|
x = self.cbl1(x)
|
|
x = self.cbl2(x)
|
|
x = self.cbl3(x)
|
|
return x
|
|
|
|
|
|
class QueryKeyMultiplexer(nn.Module):
|
|
def __init__(self, nf, multiplexer_channels, embedding_channels=256, reductions=2):
|
|
super(QueryKeyMultiplexer, self).__init__()
|
|
|
|
# Blocks used to create the query
|
|
self.input_process = ConvGnSilu(nf, nf, activation=True, norm=False, bias=True)
|
|
self.embedding_process = ConvGnSilu(embedding_channels, 256, activation=True, norm=False, bias=True)
|
|
self.reduction_blocks = nn.ModuleList([HalvingProcessingBlock(nf * 2 ** i) for i in range(reductions)])
|
|
reduction_filters = nf * 2 ** reductions
|
|
self.processing_blocks = nn.Sequential(
|
|
ConvGnSilu(reduction_filters + 256, reduction_filters + 256, kernel_size=1, activation=True, norm=False, bias=True),
|
|
ConvGnSilu(reduction_filters + 256, reduction_filters + 128, kernel_size=1, activation=True, norm=True, bias=False),
|
|
ConvGnSilu(reduction_filters + 128, reduction_filters, kernel_size=3, activation=True, norm=True, bias=False),
|
|
ConvGnSilu(reduction_filters, reduction_filters, kernel_size=3, activation=True, norm=True, bias=False))
|
|
self.expansion_blocks = nn.ModuleList([ExpansionBlock2(reduction_filters // (2 ** i)) for i in range(reductions)])
|
|
|
|
# Blocks used to create the key
|
|
self.key_process = ConvGnSilu(nf, nf, kernel_size=1, activation=True, norm=False, bias=True)
|
|
|
|
# Postprocessing blocks.
|
|
self.query_key_combine = ConvGnSilu(nf*2, nf, kernel_size=1, activation=True, norm=False, bias=False)
|
|
self.cbl1 = ConvGnSilu(nf, nf // 2, kernel_size=1, norm=True, bias=False, num_groups=4)
|
|
self.cbl2 = ConvGnSilu(nf // 2, 1, kernel_size=1, norm=False, bias=False)
|
|
|
|
def forward(self, x, embedding, transformations):
|
|
q = self.input_process(x)
|
|
embedding = self.embedding_process(embedding)
|
|
reduction_identities = []
|
|
for b in self.reduction_blocks:
|
|
reduction_identities.append(q)
|
|
q = b(q)
|
|
q = self.processing_blocks(torch.cat([q, embedding], dim=1))
|
|
for i, b in enumerate(self.expansion_blocks):
|
|
q = b(q, reduction_identities[-i - 1])
|
|
|
|
b, t, f, h, w = transformations.shape
|
|
k = transformations.view(b * t, f, h, w)
|
|
k = self.key_process(k)
|
|
|
|
q = q.view(b, 1, f, h, w).repeat(1, t, 1, 1, 1).view(b * t, f, h, w)
|
|
v = self.query_key_combine(torch.cat([q, k], dim=1))
|
|
|
|
v = self.cbl1(v)
|
|
v = self.cbl2(v)
|
|
|
|
return v.view(b, t, h, w)
|
|
|
|
|
|
class QueryKeyPyramidMultiplexer(nn.Module):
|
|
def __init__(self, nf, multiplexer_channels, reductions=3):
|
|
super(QueryKeyPyramidMultiplexer, self).__init__()
|
|
|
|
# Blocks used to create the query
|
|
self.input_process = ConvGnSilu(nf, nf, activation=True, norm=False, bias=True)
|
|
self.reduction_blocks = nn.ModuleList([HalvingProcessingBlock(nf * 2 ** i) for i in range(reductions)])
|
|
reduction_filters = nf * 2 ** reductions
|
|
self.processing_blocks = nn.Sequential(OrderedDict([('block%i' % (i,), ConvGnSilu(reduction_filters, reduction_filters, kernel_size=1, norm=True, bias=False)) for i in range(3)]))
|
|
self.expansion_blocks = nn.ModuleList([ExpansionBlock2(reduction_filters // (2 ** i)) for i in range(reductions)])
|
|
|
|
# Blocks used to create the key
|
|
self.key_process = ConvGnSilu(nf, nf, kernel_size=1, activation=True, norm=False, bias=True)
|
|
|
|
# Postprocessing blocks.
|
|
self.query_key_combine = ConvGnSilu(nf*2, nf, kernel_size=3, activation=True, norm=False, bias=False)
|
|
self.cbl0 = ConvGnSilu(nf, nf, kernel_size=3, activation=True, norm=True, bias=False)
|
|
self.cbl1 = ConvGnSilu(nf, nf // 2, kernel_size=1, norm=True, bias=False, num_groups=4)
|
|
self.cbl2 = ConvGnSilu(nf // 2, 1, kernel_size=1, norm=False, bias=False)
|
|
|
|
def forward(self, x, transformations):
|
|
q = self.input_process(x)
|
|
reduction_identities = []
|
|
for b in self.reduction_blocks:
|
|
reduction_identities.append(q)
|
|
q = b(q)
|
|
q = self.processing_blocks(q)
|
|
for i, b in enumerate(self.expansion_blocks):
|
|
q = b(q, reduction_identities[-i - 1])
|
|
|
|
b, t, f, h, w = transformations.shape
|
|
k = transformations.view(b * t, f, h, w)
|
|
k = self.key_process(k)
|
|
|
|
q = q.view(b, 1, f, h, w).repeat(1, t, 1, 1, 1).view(b * t, f, h, w)
|
|
v = self.query_key_combine(torch.cat([q, k], dim=1))
|
|
v = self.cbl0(v)
|
|
v = self.cbl1(v)
|
|
v = self.cbl2(v)
|
|
|
|
return v.view(b, t, h, w)
|
|
|
|
|
|
# Base class for models that utilize ConfigurableSwitchComputer. Provides basis functionality like logging
|
|
# switch temperature, distribution and images, as well as managing attention norms.
|
|
class SwitchModelBase(nn.Module):
|
|
def __init__(self, init_temperature=10, final_temperature_step=10000):
|
|
super(SwitchModelBase, self).__init__()
|
|
self.switches = [] # The implementing class is expected to set this to a list of all ConfigurableSwitchComputers.
|
|
self.attentions = [] # The implementing class is expected to set this in forward() to the output of the attention blocks.
|
|
self.lr = None # The implementing class is expected to set this to the input image fed into the generator. If not
|
|
# set, the attention logger will not output an image reference.
|
|
self.init_temperature = init_temperature
|
|
self.final_temperature_step = final_temperature_step
|
|
|
|
def set_temperature(self, temp):
|
|
[sw.set_temperature(temp) for sw in self.switches]
|
|
|
|
def update_for_step(self, step, experiments_path='.'):
|
|
# All-reduce the attention norm.
|
|
for sw in self.switches:
|
|
sw.switch.reduce_norm_params()
|
|
|
|
temp = max(1, 1 + self.init_temperature *
|
|
(self.final_temperature_step - step) / self.final_temperature_step)
|
|
self.set_temperature(temp)
|
|
if step % 100 == 0:
|
|
output_path = os.path.join(experiments_path, "attention_maps")
|
|
prefix = "amap_%i_a%i_%%i.png"
|
|
[save_attention_to_image_rgb(output_path, self.attentions[i], self.attentions[i].shape[3], prefix % (step, i), step,
|
|
output_mag=False) for i in range(len(self.attentions))]
|
|
if self.lr is not None:
|
|
torchvision.utils.save_image(self.lr[:, :3], os.path.join(experiments_path, "attention_maps",
|
|
"amap_%i_base_image.png" % (step,)))
|
|
|
|
# This is a bit awkward. We want this plot to show up in TB as a histogram, but we are getting an intensity
|
|
# plot out of the attention norm tensor. So we need to convert it back into a list of indexes, then feed into TB.
|
|
def compute_anorm_histogram(self):
|
|
intensities = [sw.switch.attention_norm.compute_buffer_norm().clone().detach().cpu() for sw in self.switches]
|
|
result = []
|
|
for intensity in intensities:
|
|
intensity = intensity * 10
|
|
bins = torch.tensor(list(range(len(intensity))))
|
|
intensity = intensity.long()
|
|
result.append(bins.repeat_interleave(intensity, 0))
|
|
return result
|
|
|
|
def get_debug_values(self, step, net_name):
|
|
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]
|
|
anorms = self.compute_anorm_histogram()
|
|
val = {"switch_temperature": temp}
|
|
for i in range(len(means)):
|
|
val["switch_%i_specificity" % (i,)] = means[i]
|
|
val["switch_%i_histogram" % (i,)] = hists[i]
|
|
val["switch_%i_attention_norm_histogram" % (i,)] = anorms[i]
|
|
return val
|
|
|
|
|
|
from models.archs.spinenet_arch import make_res_layer, BasicBlock
|
|
class BigMultiplexer(nn.Module):
|
|
def __init__(self, in_nc, nf, multiplexer_channels):
|
|
super(BigMultiplexer, self).__init__()
|
|
|
|
self.spine = SpineNet(arch='96', output_level=[3], double_reduce_early=False)
|
|
self.spine_red_proc = ConvGnSilu(256, nf, kernel_size=1, activation=False, norm=False, bias=False)
|
|
self.fea_tail = ConvGnSilu(in_nc, nf, kernel_size=7, bias=True, norm=False, activation=False)
|
|
self.tail_proc = make_res_layer(BasicBlock, nf, nf, 2)
|
|
self.tail_join = ReferenceJoinBlock(nf)
|
|
|
|
self.reduce = nn.Sequential(ConvGnSilu(nf, nf // 2, kernel_size=1, activation=True, norm=True, bias=False),
|
|
ConvGnSilu(nf // 2, multiplexer_channels, kernel_size=1, activation=False, norm=False, bias=False))
|
|
|
|
def forward(self, x, transformations):
|
|
s = self.spine(x)[0]
|
|
tail = self.fea_tail(x)
|
|
tail = self.tail_proc(tail)
|
|
q = F.interpolate(s, scale_factor=2, mode='nearest')
|
|
q = self.spine_red_proc(q)
|
|
q, _ = self.tail_join(q, tail)
|
|
return self.reduce(q)
|
|
|
|
|
|
class TheBigSwitch(SwitchModelBase):
|
|
def __init__(self, in_nc, nf, xforms=16, upscale=2, init_temperature=10):
|
|
super(TheBigSwitch, self).__init__(init_temperature, 10000)
|
|
self.nf = nf
|
|
self.transformation_counts = xforms
|
|
|
|
self.model_fea_conv = ConvGnLelu(in_nc, nf, kernel_size=7, norm=False, activation=False)
|
|
|
|
multiplx_fn = functools.partial(BigMultiplexer, in_nc, nf)
|
|
transform_fn = functools.partial(MultiConvBlock, nf, int(nf * 1.5), nf, kernel_size=3, depth=4, weight_init_factor=.1)
|
|
self.switch = ConfigurableSwitchComputer(nf, multiplx_fn,
|
|
pre_transform_block=None, transform_block=transform_fn,
|
|
attention_norm=True,
|
|
transform_count=self.transformation_counts, init_temp=init_temperature,
|
|
add_scalable_noise_to_transforms=False, feed_transforms_into_multiplexer=True,
|
|
anorm_multiplier=128)
|
|
self.switches = [self.switch]
|
|
|
|
self.final_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=True)
|
|
self.upsample = UpconvBlock(nf, nf // 2, block=ConvGnLelu, norm=False, activation=True, bias=True)
|
|
self.final_hr_conv1 = ConvGnLelu(nf // 2, nf // 2, kernel_size=3, norm=False, activation=False, bias=True)
|
|
self.final_hr_conv2 = ConvGnLelu(nf // 2, 3, kernel_size=3, norm=False, activation=False, bias=False)
|
|
|
|
def forward(self, x, save_attentions=True):
|
|
# The attention_maps debugger outputs <x>. Save that here.
|
|
self.lr = x.detach().cpu()
|
|
|
|
# If we're not saving attention, we also shouldn't be updating the attention norm. This is because the attention
|
|
# norm should only be getting updates with new data, not recurrent generator sampling.
|
|
for sw in self.switches:
|
|
sw.set_update_attention_norm(save_attentions)
|
|
|
|
x1 = self.model_fea_conv(x)
|
|
x1, a1 = self.switch(x1, att_in=x, do_checkpointing=True)
|
|
x_out = checkpoint(self.final_lr_conv, x1)
|
|
x_out = checkpoint(self.upsample, x_out)
|
|
x_out = checkpoint(self.final_hr_conv2, x_out)
|
|
|
|
if save_attentions:
|
|
self.attentions = [a1]
|
|
return x_out,
|
|
|
|
|
|
class ArtistMultiplexer(nn.Module):
|
|
def __init__(self, in_nc, nf, multiplexer_channels):
|
|
super(ArtistMultiplexer, self).__init__()
|
|
|
|
self.spine = SpineNet(arch='96', output_level=[3], double_reduce_early=False)
|
|
self.spine_red_proc = ConvGnSilu(256, nf, kernel_size=1, activation=False, norm=False, bias=False)
|
|
self.fea_tail = ConvGnSilu(in_nc, nf, kernel_size=7, bias=True, norm=False, activation=False)
|
|
self.tail_proc = make_res_layer(BasicBlock, nf, nf, 2)
|
|
self.tail_join = ReferenceJoinBlock(nf)
|
|
|
|
self.reduce = ConvGnSilu(nf, nf // 2, kernel_size=1, activation=True, norm=True, bias=False)
|
|
self.last_process = ConvGnSilu(nf // 2, nf // 2, kernel_size=1, activation=True, norm=False, bias=False)
|
|
self.to_attention = ConvGnSilu(nf // 2, multiplexer_channels, kernel_size=1, activation=False, norm=False, bias=False)
|
|
|
|
def forward(self, x, transformations):
|
|
s = self.spine(x)[0]
|
|
tail = self.fea_tail(x)
|
|
tail = self.tail_proc(tail)
|
|
q = F.interpolate(s, scale_factor=2, mode='nearest')
|
|
q = self.spine_red_proc(q)
|
|
q, _ = self.tail_join(q, tail)
|
|
q = self.reduce(q)
|
|
q = F.interpolate(q, scale_factor=2, mode='nearest')
|
|
return self.to_attention(self.last_process(q))
|
|
|
|
|
|
class ArtistGen(SwitchModelBase):
|
|
def __init__(self, in_nc, nf, xforms=16, upscale=2, init_temperature=10):
|
|
super(ArtistGen, self).__init__(init_temperature, 10000)
|
|
self.nf = nf
|
|
self.transformation_counts = xforms
|
|
|
|
multiplx_fn = functools.partial(ArtistMultiplexer, in_nc, nf)
|
|
transform_fn = functools.partial(MultiConvBlock, in_nc, int(in_nc * 2), in_nc, kernel_size=3, depth=4, weight_init_factor=.1)
|
|
self.switch = ConfigurableSwitchComputer(nf, multiplx_fn,
|
|
pre_transform_block=None, transform_block=transform_fn,
|
|
attention_norm=True,
|
|
transform_count=self.transformation_counts, init_temp=init_temperature,
|
|
add_scalable_noise_to_transforms=False, feed_transforms_into_multiplexer=True,
|
|
anorm_multiplier=128, post_switch_conv=False)
|
|
self.switches = [self.switch]
|
|
|
|
def forward(self, x, save_attentions=True):
|
|
# The attention_maps debugger outputs <x>. Save that here.
|
|
self.lr = x.detach().cpu()
|
|
|
|
# If we're not saving attention, we also shouldn't be updating the attention norm. This is because the attention
|
|
# norm should only be getting updates with new data, not recurrent generator sampling.
|
|
for sw in self.switches:
|
|
sw.set_update_attention_norm(save_attentions)
|
|
|
|
up = F.interpolate(x, scale_factor=2, mode="bicubic")
|
|
out, a1, att_logits = self.switch(up, att_in=x, do_checkpointing=True, output_att_logits=True)
|
|
|
|
if save_attentions:
|
|
self.attentions = [a1]
|
|
return out, att_logits.permute(0,3,1,2)
|
|
|
|
if __name__ == '__main__':
|
|
tbs = TheBigSwitch(3, 64)
|
|
x = torch.randn(4,3,64,64)
|
|
b = tbs(x) |