SPSR5

This is SPSR4, but the multiplexers have access to the output of the transformations
for making their decision.

codes/models/archs/SPSR_arch.py

@@ -5,7 +5,7 @@ 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
+from models.archs.SwitchedResidualGenerator_arch import ConvBasisMultiplexer, ConfigurableSwitchComputer, ReferencingConvMultiplexer, ReferenceImageBranch, AdaInConvBlock, ProcessingBranchWithStochasticity, EmbeddingMultiplexer, QueryKeyMultiplexer
 from switched_conv_util import save_attention_to_image_rgb
 from switched_conv import compute_attention_specificity
 import functools
@@ -540,3 +540,134 @@ class Spsr4(nn.Module):
             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, init_temperature=10):
+        super(Spsr5, self).__init__()
+        n_upscale = int(math.log(upscale, 2))
+
+        # switch options
+        transformation_filters = nf
+        self.transformation_counts = xforms
+        multiplx_fn = functools.partial(QueryKeyMultiplexer, 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, feed_transforms_into_multiplexer=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=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)
+
+        # 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, 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=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, 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=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):
+        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 % 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,
+               "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

codes/models/archs/SwitchedResidualGenerator_arch.py

@@ -175,7 +175,7 @@ class ReferencingConvMultiplexer(nn.Module):
 
 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):
+                 init_temp=20, add_scalable_noise_to_transforms=False, feed_transforms_into_multiplexer=False):
         super(ConfigurableSwitchComputer, self).__init__()
 
         tc = transform_count
@@ -187,6 +187,7 @@ class ConfigurableSwitchComputer(nn.Module):
             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
@@ -232,6 +233,8 @@ class ConfigurableSwitchComputer(nn.Module):
 
         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 memory_checkpointing_enabled:
             m = checkpoint(self.multiplexer, *att_in)
         else:
@@ -384,6 +387,9 @@ class BackboneEncoder(nn.Module):
 
         return combined
 
+# 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.
 
 # Mutiplexer that combines a structured embedding with a contextual switch input to guide alterations to that input.
 #
@@ -429,13 +435,65 @@ class EmbeddingMultiplexer(nn.Module):
         x = self.cbl3(x)
         return x
 
+
+class QueryKeyMultiplexer(nn.Module):
+    def __init__(self, nf, multiplexer_channels, 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.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.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)
+
+        k = k.view(b, t, f, h, w)  # Not sure if this is necessary..
+        q = q.view(b, 1, f, h, w).repeat(1, t, 1, 1, 1)
+        v = q * k
+        v = v.view(b * t, f, h, w)
+
+        v = self.cbl1(v)
+        v = self.cbl2(v)
+
+        return v.view(b, t, h, w)
+
+
 if __name__ == '__main__':
     bb = BackboneEncoder(64)
-    emb = EmbeddingMultiplexer(64, 10)
+    emb = QueryKeyMultiplexer(64, 10)
     x = torch.randn(4,3,64,64)
-    r = torch.randn(4,4,64,64)
+    r = torch.randn(4,3,128,128)
     xu = torch.randn(4,64,64,64)
     cp = torch.zeros((4,2), dtype=torch.long)
 
+    trans = [torch.randn(4,64,64,64) for t in range(10)]
+
     b = bb(x, r, cp)
-    emb(xu, b)
+    emb(xu, b, trans)

codes/models/networks.py

@@ -55,6 +55,10 @@ def define_G(opt, net_key='network_G', scale=None):
         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'],
+                                 init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10)
     elif which_model == "backbone_encoder":
         netG = SwitchedGen_arch.BackboneEncoder(pretrained_backbone=opt_net['pretrained_spinenet'])
     else: