diff --git a/codes/models/archs/SPSR_arch.py b/codes/models/archs/SPSR_arch.py index 1c8e7d7f..3b8f2d67 100644 --- a/codes/models/archs/SPSR_arch.py +++ b/codes/models/archs/SPSR_arch.py @@ -1,17 +1,20 @@ +import functools +import os + import math import torch import torch.nn as nn import torch.nn.functional as F -from models.archs import SPSR_util as B -from .RRDBNet_arch import RRDB -from models.archs.arch_util import ConvGnLelu, UpconvBlock, ConjoinBlock, ConvGnSilu, MultiConvBlock, ReferenceJoinBlock -from models.archs.SwitchedResidualGenerator_arch import ConvBasisMultiplexer, ConfigurableSwitchComputer, ReferencingConvMultiplexer, ReferenceImageBranch, AdaInConvBlock, ProcessingBranchWithStochasticity, EmbeddingMultiplexer, QueryKeyMultiplexer, QueryKeyPyramidMultiplexer -from switched_conv_util import save_attention_to_image_rgb -from switched_conv import compute_attention_specificity -from torch.utils.checkpoint import checkpoint -import functools -import os import torchvision +from torch.utils.checkpoint import checkpoint + +from models.archs import SPSR_util as B +from models.archs.SwitchedResidualGenerator_arch import ConfigurableSwitchComputer, ReferenceImageBranch, \ + QueryKeyMultiplexer, QueryKeyPyramidMultiplexer +from models.archs.arch_util import ConvGnLelu, UpconvBlock, MultiConvBlock, ReferenceJoinBlock +from switched_conv import compute_attention_specificity +from switched_conv_util import save_attention_to_image_rgb +from .RRDBNet_arch import RRDB class ImageGradient(nn.Module): @@ -239,444 +242,6 @@ class SPSRNet(nn.Module): return x_out_branch, x_out, x_grad -class SwitchedSpsr(nn.Module): - def __init__(self, in_nc, out_nc, nf, xforms=8, upscale=4, init_temperature=10): - super(SwitchedSpsr, self).__init__() - n_upscale = int(math.log(upscale, 2)) - - # switch options - transformation_filters = nf - switch_filters = nf - switch_reductions = 3 - switch_processing_layers = 2 - self.transformation_counts = xforms - multiplx_fn = functools.partial(ConvBasisMultiplexer, transformation_filters, switch_filters, switch_reductions, - switch_processing_layers, self.transformation_counts, use_exp2=True) - 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, - 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.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(ConvBasisMultiplexer, nf * 2, nf * 2, switch_reductions, - switch_processing_layers, self.transformation_counts // 2, use_exp2=True) - 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) - # 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. - # Note: "_branch_pretrain" is a special tag used to denote parameters that get pretrained before the rest. - 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(ConvGnLelu, transformation_filters * 2, transformation_filters * 2, norm=False, bias=False, weight_init_factor=.1) - 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.upsample = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=False, bias=False) for _ in range(n_upscale)]) - self.upsample_grad = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=False, bias=False) for _ in range(n_upscale)]) - 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): - x_grad = self.get_g_nopadding(x) - x = self.model_fea_conv(x) - - x1, a1 = self.sw1(x, True) - x2, a2 = self.sw2(x1, True) - x_fea = self.feature_lr_conv(x2) - 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, att_in=torch.cat([x1, x_b_fea], dim=1), output_attention_weights=True) - x_grad = self.grad_lr_conv(x_grad) - x_grad = self.grad_hr_conv(x_grad) - x_out_branch = self.upsample_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, att_in=x_fea, identity=x_fea, output_attention_weights=True) - x_out = self.final_lr_conv(x__branch_pretrain_cat) - x_out = self.upsample(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 - - -class RefJoiner(nn.Module): - def __init__(self, nf): - super(RefJoiner, self).__init__() - self.lin1 = nn.Linear(nf * 8, nf * 4) - self.lin2 = nn.Linear(nf * 4, nf) - self.join = ReferenceJoinBlock(nf, residual_weight_init_factor=.1) - - def forward(self, x, ref): - ref = self.lin1(ref) - ref = self.lin2(ref) - b, _, h, w = x.shape - ref = ref.view(b, -1, 1, 1) - return self.join(x, ref.repeat((1, 1, h, w))) - - -class ModuleWithRef(nn.Module): - def __init__(self, nf, mcnv, *args): - super(ModuleWithRef, self).__init__() - self.join = ReferenceJoinBlock(nf, residual_weight_init_factor=.2) - self.multi = mcnv(*args) - - def forward(self, x, ref): - out, _ = self.join(x, ref) - return self.multi(out) - - -class SwitchedSpsrWithRef2(nn.Module): - def __init__(self, in_nc, out_nc, nf, xforms=8, upscale=4, init_temperature=10): - super(SwitchedSpsrWithRef2, self).__init__() - n_upscale = int(math.log(upscale, 2)) - - # switch options - transformation_filters = nf - switch_filters = nf - self.transformation_counts = xforms - multiplx_fn = functools.partial(ConvBasisMultiplexer, transformation_filters, switch_filters, 3, - 2, use_exp2=True) - 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, - weight_init_factor=.1) - - self.reference_processor = ReferenceImageBranch(transformation_filters) - - # Feature branch - self.model_fea_conv = ConvGnLelu(in_nc, nf, kernel_size=3, norm=False, activation=False) - self.noise_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.1) - self.ref_join1 = RefJoiner(nf) - 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=False) - self.ref_join2 = RefJoiner(nf) - 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=False) - self.feature_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=False) - self.feature_lr_conv2 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False, bias=False) - - # Grad branch. Note - groupnorm on this branch is REALLY bad. Avoid it like the plague. - self.get_g_nopadding = ImageGradientNoPadding() - self.grad_conv = ConvGnLelu(in_nc, nf, kernel_size=3, norm=False, activation=False, bias=False) - self.noise_ref_join_grad = ReferenceJoinBlock(nf, residual_weight_init_factor=.1) - self.ref_join3 = RefJoiner(nf) - self.grad_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.3, final_norm=False) - self.sw_grad = ConfigurableSwitchComputer(transformation_filters, multiplx_fn, - 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=False) - self.grad_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=True) - self.grad_lr_conv2 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=True) - self.upsample_grad = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=True, bias=False) for _ in range(n_upscale)]) - self.grad_branch_output_conv = ConvGnLelu(nf, out_nc, kernel_size=1, norm=False, activation=False, bias=True) - - # Join branch (grad+fea) - self.ref_join4 = RefJoiner(nf) - self.noise_ref_join_conjoin = ReferenceJoinBlock(nf, residual_weight_init_factor=.1) - self.conjoin_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.3) - self.conjoin_sw = 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=False) - self.final_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=True) - self.upsample = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=True, bias=True) for _ in range(n_upscale)]) - self.final_hr_conv1 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False, bias=True) - 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.conjoin_sw] - self.attentions = None - self.init_temperature = init_temperature - self.final_temperature_step = 10000 - - def forward(self, x, ref, center_coord): - # The attention_maps debugger outputs . Save that here. - self.lr = x.detach().cpu() - - ref_stds = [] - noise_stds = [] - - x_grad = self.get_g_nopadding(x) - ref = self.reference_processor(ref, center_coord) - - x = self.model_fea_conv(x) - x1 = x - x1, rstd = self.ref_join1(x1, ref) - x1, a1 = self.sw1(x1, True, identity=x) - ref_stds.append(rstd) - - x2 = x1 - x2, nstd = self.noise_ref_join(x2, torch.randn_like(x2)) - x2, rstd = self.ref_join2(x2, ref) - x2, a2 = self.sw2(x2, True, identity=x1) - noise_stds.append(nstd) - ref_stds.append(rstd) - - x_grad = self.grad_conv(x_grad) - x_grad_identity = x_grad - x_grad, nstd = self.noise_ref_join_grad(x_grad, torch.randn_like(x_grad)) - x_grad, rstd = self.ref_join3(x_grad, ref) - x_grad, grad_fea_std = self.grad_ref_join(x_grad, x1) - x_grad, a3 = self.sw_grad(x_grad, True, identity=x_grad_identity) - x_grad = self.grad_lr_conv(x_grad) - x_grad = self.grad_lr_conv2(x_grad) - x_grad_out = self.upsample_grad(x_grad) - x_grad_out = self.grad_branch_output_conv(x_grad_out) - noise_stds.append(nstd) - ref_stds.append(rstd) - - x_out = x2 - x_out, nstd = self.noise_ref_join_conjoin(x_out, torch.randn_like(x_out)) - x_out, rstd = self.ref_join4(x_out, ref) - x_out, fea_grad_std = self.conjoin_ref_join(x_out, x_grad) - x_out, a4 = self.conjoin_sw(x_out, True, identity=x2) - x_out = self.final_lr_conv(x_out) - x_out = self.upsample(x_out) - x_out = self.final_hr_conv1(x_out) - x_out = self.final_hr_conv2(x_out) - noise_stds.append(nstd) - ref_stds.append(rstd) - - self.attentions = [a1, a2, a3, a4] - self.noise_stds = torch.stack(noise_stds).mean().detach().cpu() - self.ref_stds = torch.stack(ref_stds).mean().detach().cpu() - self.grad_fea_std = grad_fea_std.detach().cpu() - self.fea_grad_std = fea_grad_std.detach().cpu() - return x_grad_out, 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 % 500 == 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.transformation_counts, prefix % (step, i), step, output_mag=False) for i in range(len(self.attentions))] - torchvision.utils.save_image(self.lr, os.path.join(experiments_path, "attention_maps", "amap_%i_base_image.png" % (step,))) - - 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] - val = {"switch_temperature": temp, - "reference_branch_std_dev": self.ref_stds, - "noise_branch_std_dev": self.noise_stds, - "grad_branch_feat_intg_std_dev": self.grad_fea_std, - "conjoin_branch_grad_intg_std_dev": self.fea_grad_std} - for i in range(len(means)): - val["switch_%i_specificity" % (i,)] = means[i] - val["switch_%i_histogram" % (i,)] = hists[i] - return val - - -class Spsr4(nn.Module): - def __init__(self, in_nc, out_nc, nf, xforms=8, upscale=4, init_temperature=10): - super(Spsr4, self).__init__() - n_upscale = int(math.log(upscale, 2)) - - # switch options - transformation_filters = nf - self.transformation_counts = xforms - multiplx_fn = functools.partial(EmbeddingMultiplexer, transformation_filters) - 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, - weight_init_factor=.1) - - # Feature branch - self.model_fea_conv = ConvGnLelu(in_nc, nf, kernel_size=3, norm=False, activation=False) - self.noise_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.1) - 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=False) - 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=False) - self.feature_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=False) - self.feature_lr_conv2 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False, bias=False) - - # Grad branch. Note - groupnorm on this branch is REALLY bad. Avoid it like the plague. - self.get_g_nopadding = ImageGradientNoPadding() - self.grad_conv = ConvGnLelu(in_nc, nf, kernel_size=3, norm=False, activation=False, bias=False) - self.noise_ref_join_grad = ReferenceJoinBlock(nf, residual_weight_init_factor=.1) - self.grad_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.3, final_norm=False) - self.sw_grad = ConfigurableSwitchComputer(transformation_filters, multiplx_fn, - 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=False) - self.grad_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=True) - self.grad_lr_conv2 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=True) - self.upsample_grad = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=True, bias=False) for _ in range(n_upscale)]) - self.grad_branch_output_conv = ConvGnLelu(nf, out_nc, kernel_size=1, norm=False, activation=False, bias=True) - - # Join branch (grad+fea) - self.noise_ref_join_conjoin = ReferenceJoinBlock(nf, residual_weight_init_factor=.1) - self.conjoin_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.3) - self.conjoin_sw = 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=False) - self.final_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=True) - self.upsample = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=True, bias=True) for _ in range(n_upscale)]) - self.final_hr_conv1 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False, bias=True) - 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.conjoin_sw] - self.attentions = None - self.init_temperature = init_temperature - self.final_temperature_step = 10000 - - def forward(self, x, embedding): - # The attention_maps debugger outputs . Save that here. - self.lr = x.detach().cpu() - - noise_stds = [] - - x_grad = self.get_g_nopadding(x) - - x = self.model_fea_conv(x) - x1 = x - x1, a1 = self.sw1(x1, True, identity=x, att_in=(x1, embedding)) - - x2 = x1 - x2, nstd = self.noise_ref_join(x2, torch.randn_like(x2)) - x2, a2 = self.sw2(x2, True, identity=x1, att_in=(x2, embedding)) - noise_stds.append(nstd) - - x_grad = self.grad_conv(x_grad) - x_grad_identity = x_grad - x_grad, nstd = self.noise_ref_join_grad(x_grad, torch.randn_like(x_grad)) - x_grad, grad_fea_std = self.grad_ref_join(x_grad, x1) - x_grad, a3 = self.sw_grad(x_grad, True, identity=x_grad_identity, att_in=(x_grad, embedding)) - x_grad = self.grad_lr_conv(x_grad) - x_grad = self.grad_lr_conv2(x_grad) - x_grad_out = self.upsample_grad(x_grad) - x_grad_out = self.grad_branch_output_conv(x_grad_out) - noise_stds.append(nstd) - - x_out = x2 - x_out, nstd = self.noise_ref_join_conjoin(x_out, torch.randn_like(x_out)) - x_out, fea_grad_std = self.conjoin_ref_join(x_out, x_grad) - x_out, a4 = self.conjoin_sw(x_out, True, identity=x2, att_in=(x_out, embedding)) - x_out = self.final_lr_conv(x_out) - x_out = self.upsample(x_out) - x_out = self.final_hr_conv1(x_out) - x_out = self.final_hr_conv2(x_out) - noise_stds.append(nstd) - - self.attentions = [a1, a2, a3, a4] - self.noise_stds = torch.stack(noise_stds).mean().detach().cpu() - self.grad_fea_std = grad_fea_std.detach().cpu() - self.fea_grad_std = fea_grad_std.detach().cpu() - return x_grad_out, 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 % 500 == 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.transformation_counts, prefix % (step, i), step, output_mag=False) for i in range(len(self.attentions))] - torchvision.utils.save_image(self.lr, os.path.join(experiments_path, "attention_maps", "amap_%i_base_image.png" % (step,))) - - 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] - val = {"switch_temperature": temp, - "noise_branch_std_dev": self.noise_stds, - "grad_branch_feat_intg_std_dev": self.grad_fea_std, - "conjoin_branch_grad_intg_std_dev": self.fea_grad_std} - for i in range(len(means)): - val["switch_%i_specificity" % (i,)] = means[i] - val["switch_%i_histogram" % (i,)] = hists[i] - return val - - class Spsr5(nn.Module): def __init__(self, in_nc, out_nc, nf, xforms=8, upscale=4, multiplexer_reductions=2, init_temperature=10): super(Spsr5, self).__init__() @@ -813,6 +378,8 @@ class Spsr5(nn.Module): return val +# Variant of Spsr5 which uses multiplexer blocks that are not derived from an embedding. Also makes a few "best practices" +# adjustments learned over the past few weeks (no noise, kernel_size=7 class Spsr6(nn.Module): def __init__(self, in_nc, out_nc, nf, xforms=8, upscale=4, multiplexer_reductions=3, init_temperature=10): super(Spsr6, self).__init__() @@ -935,4 +502,139 @@ class Spsr6(nn.Module): val["switch_%i_histogram" % (i,)] = hists[i] return val +# Variant of Spsr7 which uses multiplexer blocks that feed off of a reference embedding. Also computes that embedding. +class Spsr7(nn.Module): + def __init__(self, in_nc, out_nc, nf, xforms=8, upscale=4, multiplexer_reductions=3, init_temperature=10): + super(Spsr7, self).__init__() + n_upscale = int(math.log(upscale, 2)) + + # processing the input embedding + self.reference_embedding = ReferenceImageBranch(nf) + + # switch options + self.nf = nf + transformation_filters = nf + self.transformation_counts = xforms + multiplx_fn = functools.partial(QueryKeyMultiplexer, transformation_filters, embedding_channels=512, reductions=multiplexer_reductions) + 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=7, norm=False, activation=False) + self.sw1 = ConfigurableSwitchComputer(transformation_filters, 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) + self.sw1_out = nn.Sequential(ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True), + ConvGnLelu(nf, 3, kernel_size=1, norm=False, activation=False, bias=True)) + self.sw2 = ConfigurableSwitchComputer(transformation_filters, 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) + self.feature_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=True, activation=False) + self.feature_lr_conv2 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False, bias=False) + self.sw2_out = nn.Sequential(ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True), + ConvGnLelu(nf, 3, kernel_size=1, norm=False, activation=False, bias=True)) + + # Grad branch. Note - groupnorm on this branch is REALLY bad. Avoid it like the plague. + self.get_g_nopadding = ImageGradientNoPadding() + self.grad_conv = ConvGnLelu(in_nc, nf, kernel_size=7, norm=False, activation=False, bias=False) + self.grad_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.3, final_norm=False) + + self.sw_grad = ConfigurableSwitchComputer(transformation_filters, multiplx_fn, + pre_transform_block=None, transform_block=transform_fn, + attention_norm=True, + transform_count=self.transformation_counts // 2, init_temp=init_temperature, + add_scalable_noise_to_transforms=False, feed_transforms_into_multiplexer=True) + self.grad_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=True) + self.grad_lr_conv2 = ConvGnLelu(nf, nf, kernel_size=1, norm=False, activation=True, bias=True) + self.upsample_grad = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=True, bias=False) for _ in range(n_upscale)]) + self.grad_branch_output_conv = ConvGnLelu(nf, out_nc, kernel_size=1, norm=False, activation=False, bias=True) + + # Join branch (grad+fea) + self.noise_ref_join_conjoin = ReferenceJoinBlock(nf, residual_weight_init_factor=.1) + self.conjoin_ref_join = ReferenceJoinBlock(nf, residual_weight_init_factor=.3) + self.conjoin_sw = ConfigurableSwitchComputer(transformation_filters, 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) + self.final_lr_conv = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=True, bias=True) + self.upsample = nn.Sequential(*[UpconvBlock(nf, nf, block=ConvGnLelu, norm=False, activation=True, bias=True) for _ in range(n_upscale)]) + self.final_hr_conv1 = ConvGnLelu(nf, nf, kernel_size=3, norm=False, activation=False, bias=True) + self.final_hr_conv2 = ConvGnLelu(nf, out_nc, kernel_size=1, norm=False, activation=False, bias=False) + self.switches = [self.sw1, self.sw2, self.sw_grad, self.conjoin_sw] + self.attentions = None + self.init_temperature = init_temperature + self.final_temperature_step = 10000 + self.lr = None + + def forward(self, x, ref, ref_center): + # The attention_maps debugger outputs . Save that here. + self.lr = x.detach().cpu() + + x_grad = self.get_g_nopadding(x) + ref_code = self.reference_embedding(ref, ref_center) + ref_embedding = ref_code.view(-1, self.nf * 8, 1, 1).repeat(1, 1, x.shape[2] // 8, x.shape[3] // 8) + + x = self.model_fea_conv(x) + x1 = x + x1, a1 = self.sw1(x1, True, identity=x, att_in=(x1, ref_embedding)) + s1out = self.sw1_out(x1) + + x2 = x1 + x2, a2 = self.sw2(x2, True, identity=x1, att_in=(x2, ref_embedding)) + s2out = self.sw2_out(x2) + + x_grad = self.grad_conv(x_grad) + x_grad_identity = x_grad + x_grad, grad_fea_std = self.grad_ref_join(x_grad, x1) + x_grad, a3 = self.sw_grad(x_grad, True, identity=x_grad_identity, att_in=(x_grad, ref_embedding)) + x_grad = self.grad_lr_conv(x_grad) + x_grad = self.grad_lr_conv2(x_grad) + x_grad_out = self.upsample_grad(x_grad) + x_grad_out = self.grad_branch_output_conv(x_grad_out) + + x_out = x2 + x_out, fea_grad_std = self.conjoin_ref_join(x_out, x_grad) + x_out, a4 = self.conjoin_sw(x_out, True, identity=x2, att_in=(x_out, ref_embedding)) + x_out = self.final_lr_conv(x_out) + x_out = checkpoint(self.upsample, x_out) + x_out = checkpoint(self.final_hr_conv1, x_out) + x_out = self.final_hr_conv2(x_out) + + self.attentions = [a1, a2, a3, a4] + self.grad_fea_std = grad_fea_std.detach().cpu() + self.fea_grad_std = fea_grad_std.detach().cpu() + return x_grad_out, x_out, s1out, s2out + + 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 % 500 == 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.transformation_counts, prefix % (step, i), step, output_mag=False) for i in range(len(self.attentions))] + torchvision.utils.save_image(self.lr, os.path.join(experiments_path, "attention_maps", "amap_%i_base_image.png" % (step,))) + + 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] + val = {"switch_temperature": temp, + "grad_branch_feat_intg_std_dev": self.grad_fea_std, + "conjoin_branch_grad_intg_std_dev": self.fea_grad_std} + for i in range(len(means)): + val["switch_%i_specificity" % (i,)] = means[i] + val["switch_%i_histogram" % (i,)] = hists[i] + return val diff --git a/codes/models/archs/SwitchedResidualGenerator_arch.py b/codes/models/archs/SwitchedResidualGenerator_arch.py index 40ad3930..2f42b2e1 100644 --- a/codes/models/archs/SwitchedResidualGenerator_arch.py +++ b/codes/models/archs/SwitchedResidualGenerator_arch.py @@ -80,99 +80,6 @@ def gather_2d(input, index): return result -# 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.filter_conv = ConvGnSilu(4, base_filters, bias=True) - self.reduction_blocks = nn.ModuleList([HalvingProcessingBlock(base_filters * 2 ** i) for i in range(3)]) - reduction_filters = base_filters * 2 ** 3 - self.processing_blocks = nn.Sequential(OrderedDict([('block%i' % (i,), ConvGnSilu(reduction_filters, reduction_filters, bias=False)) for i in range(4)])) - - # 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.filter_conv(x) - reduction_identities = [] - for b in self.reduction_blocks: - reduction_identities.append(x) - x = b(x) - x = self.processing_blocks(x) - return gather_2d(x, center_point // 8) - - -class AdaInConvBlock(nn.Module): - def __init__(self, reference_size, in_nc, out_nc, conv_block=ConvGnLelu): - super(AdaInConvBlock, self).__init__() - self.filter_conv = conv_block(in_nc, out_nc, activation=True, norm=False, bias=False) - self.ref_proc = nn.Linear(reference_size, reference_size) - self.ref_red = nn.Linear(reference_size, out_nc * 2) - self.feature_norm = torch.nn.InstanceNorm2d(out_nc) - self.style_norm = torch.nn.InstanceNorm1d(out_nc) - self.post_fuse_conv = conv_block(out_nc, out_nc, activation=False, norm=True, bias=True) - - def forward(self, x, ref): - x = self.feature_norm(self.filter_conv(x)) - ref = self.ref_proc(ref) - ref = self.ref_red(ref) - b, c = ref.shape - ref = self.style_norm(ref.view(b, 2, c // 2)) - x = x * ref[:, 0, :].unsqueeze(dim=2).unsqueeze(dim=3).expand(x.shape) + ref[:, 1, :].unsqueeze(dim=2).unsqueeze(dim=3).expand(x.shape) - return self.post_fuse_conv(x) - - -class ProcessingBranchWithStochasticity(nn.Module): - def __init__(self, nf_in, nf_out, noise_filters, depth): - super(ProcessingBranchWithStochasticity, self).__init__() - nf_gap = nf_out - nf_in - self.noise_filters = noise_filters - self.processor = MultiConvBlock(nf_in + noise_filters, nf_in + nf_gap // 2, nf_out, kernel_size=3, depth=depth, weight_init_factor = .1) - - def forward(self, x): - b, c, h, w = x.shape - noise = torch.randn((b, self.noise_filters, h, w), device=x.device) - return self.processor(torch.cat([x, noise], dim=1)) - - -# This is similar to ConvBasisMultiplexer, except that it takes a linear reference tensor as a second input to -# provide better results. It also has fixed parameterization in several places -class ReferencingConvMultiplexer(nn.Module): - def __init__(self, input_channels, base_filters, multiplexer_channels, use_gn=True): - super(ReferencingConvMultiplexer, self).__init__() - self.style_fuse = AdaInConvBlock(512, input_channels, base_filters, ConvGnSilu) - - self.reduction_blocks = nn.ModuleList([HalvingProcessingBlock(base_filters * 2 ** i) for i in range(3)]) - reduction_filters = base_filters * 2 ** 3 - self.processing_blocks = nn.Sequential(OrderedDict([('block%i' % (i,), ConvGnSilu(reduction_filters, reduction_filters, bias=False)) for i in range(2)])) - self.expansion_blocks = nn.ModuleList([ExpansionBlock2(reduction_filters // (2 ** i)) for i in range(3)]) - - 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, ref): - x = self.style_fuse(x, ref) - - 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 - - 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, feed_transforms_into_multiplexer=False): @@ -476,9 +383,28 @@ class BackboneResnet(nn.Module): return self.sequence(fea) -# Note to future self: -# Can I do a real transformer here? Such as by having the multiplexer be able to toggle off of transformations by -# their output? The embedding will be used as the "Query" to the "QueryxKey=Value" relationship. +# 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. # @@ -526,12 +452,12 @@ class EmbeddingMultiplexer(nn.Module): class QueryKeyMultiplexer(nn.Module): - def __init__(self, nf, multiplexer_channels, reductions=2): + 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(256, 256, 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( @@ -571,7 +497,7 @@ class QueryKeyMultiplexer(nn.Module): v = self.cbl2(v) return v.view(b, t, h, w) - + class QueryKeyPyramidMultiplexer(nn.Module): def __init__(self, nf, multiplexer_channels, reductions=3): @@ -615,6 +541,7 @@ class QueryKeyPyramidMultiplexer(nn.Module): return v.view(b, t, h, w) + if __name__ == '__main__': bb = BackboneEncoder(64) emb = QueryKeyMultiplexer(64, 10) diff --git a/codes/models/networks.py b/codes/models/networks.py index 1dfb8600..5f28cf72 100644 --- a/codes/models/networks.py +++ b/codes/models/networks.py @@ -56,18 +56,6 @@ def define_G(opt, net_key='network_G', scale=None): 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']) - elif which_model == "spsr_switched": - xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8 - netG = spsr.SwitchedSpsr(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'], - init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10) - elif which_model == "spsr_switched_with_ref2" or which_model == "spsr3": - xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8 - netG = spsr.SwitchedSpsrWithRef2(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'], - init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10) - elif which_model == "spsr4": - xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8 - netG = spsr.Spsr4(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'], - init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10) elif which_model == "spsr5": xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8 netG = spsr.Spsr5(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'], @@ -78,6 +66,11 @@ def define_G(opt, net_key='network_G', scale=None): netG = spsr.Spsr6(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'], multiplexer_reductions=opt_net['multiplexer_reductions'] if 'multiplexer_reductions' in opt_net.keys() else 3, init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10) + elif which_model == "spsr7": + xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8 + netG = spsr.Spsr7(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'], + multiplexer_reductions=opt_net['multiplexer_reductions'] if 'multiplexer_reductions' in opt_net.keys() else 3, + init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10) elif which_model == "ssgr1": xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8 netG = ssg.SSGr1(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'], diff --git a/codes/train.py b/codes/train.py index f88e327c..7c6f360e 100644 --- a/codes/train.py +++ b/codes/train.py @@ -32,7 +32,7 @@ def init_dist(backend='nccl', **kwargs): def main(): #### options parser = argparse.ArgumentParser() - parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_exd_imgset_spsr5_spine_no_pretrain.yml') + parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_exd_imgset_spsr7_multiloss.yml') parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args()