SRG2 revival

Big update to SRG2 architecture to pull in a lot of things that have been learned:
- Use group norm instead of batch norm
- Initialize the weights on the transformations low like is done in RRDB rather than using the scalar. Models live or die by their early stages, and this ones early stage is pretty weak
- Transform multiplexer to use u-net like architecture.
- Just use one set of configuration variables instead of a list - flat networks performed fine in this regard.

codes/models/archs/NestedSwitchGenerator.py

@@ -1,6 +1,7 @@
 import torch
 from torch import nn
-from models.archs.SwitchedResidualGenerator_arch import ConvBnLelu, ConvBnRelu, MultiConvBlock, initialize_weights
+from models.archs.arch_util import ConvBnLelu, ConvBnRelu
+from models.archs.SwitchedResidualGenerator_arch import MultiConvBlock
 from switched_conv import BareConvSwitch, compute_attention_specificity
 from switched_conv_util import save_attention_to_image
 from functools import partial

codes/models/archs/SwitchedResidualGenerator_arch.py

@@ -4,20 +4,20 @@ from switched_conv import BareConvSwitch, compute_attention_specificity
 import torch.nn.functional as F
 import functools
 from collections import OrderedDict
-from models.archs.arch_util import initialize_weights, ConvBnRelu, ConvBnLelu, ConvBnSilu
+from models.archs.arch_util import ConvBnLelu, ConvGnSilu
 from models.archs.RRDBNet_arch import ResidualDenseBlock_5C
 from models.archs.spinenet_arch import SpineNet
 from switched_conv_util import save_attention_to_image
 
 
 class MultiConvBlock(nn.Module):
-    def __init__(self, filters_in, filters_mid, filters_out, kernel_size, depth, scale_init=1, bn=False):
+    def __init__(self, filters_in, filters_mid, filters_out, kernel_size, depth, scale_init=1, bn=False, weight_init_factor=1):
         assert depth >= 2
         super(MultiConvBlock, self).__init__()
         self.noise_scale = nn.Parameter(torch.full((1,), fill_value=.01))
-        self.bnconvs = nn.ModuleList([ConvBnLelu(filters_in, filters_mid, kernel_size, bn=bn, bias=False)] +
+        self.bnconvs = nn.ModuleList([ConvBnLelu(filters_in, filters_mid, kernel_size, bn=bn, bias=False, weight_init_factor=weight_init_factor)] +
-                                     [ConvBnLelu(filters_mid, filters_mid, kernel_size, bn=bn, bias=False) for i in range(depth-2)] +
+                                     [ConvBnLelu(filters_mid, filters_mid, kernel_size, bn=bn, bias=False, weight_init_factor=weight_init_factor) for i in range(depth-2)] +
-                                     [ConvBnLelu(filters_mid, filters_out, kernel_size, lelu=False, bn=False, bias=False)])
+                                     [ConvBnLelu(filters_mid, filters_out, kernel_size, lelu=False, bn=False, bias=False, weight_init_factor=weight_init_factor)])
         self.scale = nn.Parameter(torch.full((1,), fill_value=scale_init))
         self.bias = nn.Parameter(torch.zeros(1))
 
@@ -35,43 +35,56 @@ class MultiConvBlock(nn.Module):
 class HalvingProcessingBlock(nn.Module):
     def __init__(self, filters):
         super(HalvingProcessingBlock, self).__init__()
-        self.bnconv1 = ConvBnLelu(filters, filters * 2, stride=2, bn=False, bias=False)
+        self.bnconv1 = ConvGnSilu(filters, filters * 2, stride=2, gn=False, bias=False)
-        self.bnconv2 = ConvBnLelu(filters * 2, filters * 2, bn=True, bias=False)
+        self.bnconv2 = ConvGnSilu(filters * 2, filters * 2, gn=True, bias=False)
 
     def forward(self, x):
         x = self.bnconv1(x)
         return self.bnconv2(x)
 
 
-# Creates a nested series of convolutional blocks. Each block processes the input data in-place and adds
+class ExpansionBlock(nn.Module):
-# filter_growth filters. Return is (nn.Sequential, ending_filters)
+    def __init__(self, filters):
-def create_sequential_growing_processing_block(filters_init, filter_growth, num_convs):
+        super(ExpansionBlock, self).__init__()
-    convs = []
+        self.decimate = ConvGnSilu(filters, filters // 2, kernel_size=1, bias=False, silu=False, gn=False)
-    current_filters = filters_init
+        self.conjoin = ConvGnSilu(filters, filters // 2, kernel_size=3, bias=True, silu=False, gn=True)
-    for i in range(num_convs):
+        self.process = ConvGnSilu(filters // 2, filters // 2, kernel_size=3, bias=False, silu=True, gn=True)
-        convs.append(ConvBnSilu(current_filters, current_filters + filter_growth, bn=True, bias=False))
+
-        current_filters += filter_growth
+    def forward(self, input, passthrough):
-    return nn.Sequential(*convs), current_filters
+        x = F.interpolate(input, scale_factor=2, mode="nearest")
+        x = self.decimate(x)
+        x = self.conjoin(torch.cat([x, passthrough], dim=1))
+        return self.process(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, growth, reductions, processing_depth, multiplexer_channels, use_bn=True):
+    def __init__(self, input_channels, base_filters, reductions, processing_depth, multiplexer_channels, use_gn=True):
         super(ConvBasisMultiplexer, self).__init__()
-        self.filter_conv = ConvBnSilu(input_channels, base_filters, bias=True)
+        self.filter_conv = ConvGnSilu(input_channels, base_filters, bias=True)
-        self.reduction_blocks = nn.Sequential(OrderedDict([('block%i:' % (i,), HalvingProcessingBlock(base_filters * 2 ** i)) for i in range(reductions)]))
+        self.reduction_blocks = nn.ModuleList([HalvingProcessingBlock(base_filters * 2 ** i) for i in range(reductions)])
         reduction_filters = base_filters * 2 ** reductions
-        self.processing_blocks, self.output_filter_count = create_sequential_growing_processing_block(reduction_filters, growth, processing_depth)
+        self.processing_blocks = nn.Sequential(OrderedDict([('block%i' % (i,), ConvGnSilu(reduction_filters, reduction_filters, bias=False)) for i in range(processing_depth)]))
+        self.expansion_blocks = nn.ModuleList([ExpansionBlock(reduction_filters // (2 ** i)) for i in range(reductions)])
 
-        gap = self.output_filter_count - multiplexer_channels
+        gap = base_filters - multiplexer_channels
-        # Hey silly - if you're going to interpolate later, do it here instead. Then add some processing layers to let the model adjust it properly.
+        cbl1_out = ((base_filters - (gap // 2)) // 4) * 4   # Must be multiples of 4 to use with group norm.
-        self.cbl1 = ConvBnSilu(self.output_filter_count, self.output_filter_count - (gap // 2), bn=use_bn, bias=False)
+        self.cbl1 = ConvGnSilu(base_filters, cbl1_out, gn=use_gn, bias=False, num_groups=4)
-        self.cbl2 = ConvBnSilu(self.output_filter_count - (gap // 2), self.output_filter_count - (3 * gap // 4), bn=use_bn, bias=False)
+        cbl2_out = ((base_filters - (3 * gap // 4)) // 4) * 4
-        self.cbl3 = ConvBnSilu(self.output_filter_count - (3 * gap // 4), multiplexer_channels, bias=True)
+        self.cbl2 = ConvGnSilu(cbl1_out, cbl2_out, gn=use_gn, bias=False, num_groups=4)
+        self.cbl3 = ConvGnSilu(cbl2_out, multiplexer_channels, bias=True, gn=False)
 
     def forward(self, x):
         x = self.filter_conv(x)
-        x = self.reduction_blocks(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)
@@ -94,13 +107,13 @@ class BackboneMultiplexer(nn.Module):
     def __init__(self, backbone: CachedBackboneWrapper, transform_count):
         super(BackboneMultiplexer, self).__init__()
         self.backbone = backbone
-        self.proc = nn.Sequential(ConvBnSilu(256, 256, kernel_size=3, bias=True),
+        self.proc = nn.Sequential(ConvGnSilu(256, 256, kernel_size=3, bias=True),
-                                  ConvBnSilu(256, 256, kernel_size=3, bias=False))
+                                  ConvGnSilu(256, 256, kernel_size=3, bias=False))
-        self.up1 = nn.Sequential(ConvBnSilu(256, 128, kernel_size=3, bias=False, bn=False, silu=False),
+        self.up1 = nn.Sequential(ConvGnSilu(256, 128, kernel_size=3, bias=False, gn=False, silu=False),
-                                 ConvBnSilu(128, 128, kernel_size=3, bias=False))
+                                 ConvGnSilu(128, 128, kernel_size=3, bias=False))
-        self.up2 = nn.Sequential(ConvBnSilu(128, 64, kernel_size=3, bias=False, bn=False, silu=False),
+        self.up2 = nn.Sequential(ConvGnSilu(128, 64, kernel_size=3, bias=False, gn=False, silu=False),
-                                 ConvBnSilu(64, 64, kernel_size=3, bias=False))
+                                 ConvGnSilu(64, 64, kernel_size=3, bias=False))
-        self.final = ConvBnSilu(64, transform_count, bias=False, bn=False, silu=False)
+        self.final = ConvGnSilu(64, transform_count, bias=False, gn=False, silu=False)
 
     def forward(self, x):
         spine = self.backbone.get_forward_result()
@@ -112,13 +125,10 @@ class BackboneMultiplexer(nn.Module):
 
 class ConfigurableSwitchComputer(nn.Module):
     def __init__(self, base_filters, multiplexer_net, pre_transform_block, transform_block, transform_count, init_temp=20,
-                 enable_negative_transforms=False, add_scalable_noise_to_transforms=False, init_scalar=1):
+                 add_scalable_noise_to_transforms=False):
         super(ConfigurableSwitchComputer, self).__init__()
-        self.enable_negative_transforms = enable_negative_transforms
 
         tc = transform_count
-        if self.enable_negative_transforms:
-            tc = transform_count * 2
         self.multiplexer = multiplexer_net(tc)
 
         self.pre_transform = pre_transform_block()
@@ -128,11 +138,11 @@ class ConfigurableSwitchComputer(nn.Module):
 
         # And the switch itself, including learned scalars
         self.switch = BareConvSwitch(initial_temperature=init_temp)
-        self.switch_scale = nn.Parameter(torch.full((1,), float(init_scalar)))
+        self.switch_scale = nn.Parameter(torch.full((1,), float(1)))
-        self.post_switch_conv = ConvBnLelu(base_filters, base_filters, bn=False, bias=False)
+        self.post_switch_conv = ConvBnLelu(base_filters, base_filters, bn=False, bias=True)
-        # The post_switch_conv gets a near-zero scale. The network can decide to magnify it (or not) depending on its needs.
+        # The post_switch_conv gets a low scale initially. The network can decide to magnify it (or not)
-        self.psc_scale = nn.Parameter(torch.full((1,), float(1e-3)))
+        # depending on its needs.
-        self.bias = nn.Parameter(torch.zeros(1))
+        self.psc_scale = nn.Parameter(torch.full((1,), float(.1)))
 
     def forward(self, x, output_attention_weights=False):
         identity = x
@@ -142,17 +152,12 @@ class ConfigurableSwitchComputer(nn.Module):
 
         x = self.pre_transform(x)
         xformed = [t.forward(x) for t in self.transforms]
-        if self.enable_negative_transforms:
-            xformed.extend([-t for t in xformed])
 
         m = self.multiplexer(identity)
-        # Interpolate the multiplexer across the entire shape of the image.
-        m = F.interpolate(m, size=xformed[0].shape[2:], mode='nearest')
 
         outputs, attention = self.switch(xformed, m, True)
         outputs = identity + outputs * self.switch_scale
-        outputs = identity + self.post_switch_conv(outputs) * self.psc_scale
+        outputs = outputs + self.post_switch_conv(outputs) * self.psc_scale
-        outputs = outputs + self.bias
         if output_attention_weights:
             return outputs, attention
         else:
@@ -163,25 +168,25 @@ class ConfigurableSwitchComputer(nn.Module):
 
 
 class ConfigurableSwitchedResidualGenerator2(nn.Module):
-    def __init__(self, switch_filters, switch_growths, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes,
+    def __init__(self, switch_depth, switch_filters, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes,
                  trans_layers, transformation_filters, initial_temp=20, final_temperature_step=50000, heightened_temp_min=1,
-                 heightened_final_step=50000, upsample_factor=1, enable_negative_transforms=False,
+                 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, bn=False, lelu=False, bias=True)
-        self.sw_conv = ConvBnLelu(transformation_filters, transformation_filters, lelu=False, bias=True)
         self.upconv1 = ConvBnLelu(transformation_filters, transformation_filters, bn=False, bias=True)
         self.upconv2 = ConvBnLelu(transformation_filters, transformation_filters, bn=False, bias=True)
         self.hr_conv = ConvBnLelu(transformation_filters, transformation_filters, bn=False, bias=True)
         self.final_conv = ConvBnLelu(transformation_filters, 3, bn=False, lelu=False, bias=True)
-        for filters, growth, sw_reduce, sw_proc, trans_count, kernel, layers in zip(switch_filters, switch_growths, switch_reductions, switch_processing_layers, trans_counts, trans_kernel_sizes, trans_layers):
+        for _ in range(switch_depth):
-            multiplx_fn = functools.partial(ConvBasisMultiplexer, transformation_filters, filters, growth, sw_reduce, sw_proc, trans_count)
+            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, bn=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=functools.partial(ConvBnLelu, transformation_filters, transformation_filters, bn=False, bias=False),
+                                                       pre_transform_block=pretransform_fn, transform_block=transform_fn,
-                                                       transform_block=functools.partial(MultiConvBlock, transformation_filters, transformation_filters + growth, transformation_filters, kernel_size=kernel, depth=layers),
+                                                       transform_count=trans_counts, init_temp=initial_temp,
-                                                       transform_count=trans_count, init_temp=initial_temp, enable_negative_transforms=enable_negative_transforms,
+                                                       add_scalable_noise_to_transforms=add_scalable_noise_to_transforms))
-                                                       add_scalable_noise_to_transforms=add_scalable_noise_to_transforms, init_scalar=.1))
 
         self.switches = nn.ModuleList(switches)
         self.transformation_counts = trans_counts
@@ -268,7 +273,7 @@ class ConfigurableSwitchedResidualGenerator3(nn.Module):
         pretransform_fn = functools.partial(nn.Sequential, ConvBnLelu(base_filters, base_filters, kernel_size=3, bn=False, lelu=False, bias=False))
         transform_fn = functools.partial(MultiConvBlock, base_filters, int(base_filters * 1.5), base_filters, kernel_size=3, depth=4)
         self.switch = ConfigurableSwitchComputer(base_filters, multiplx_fn, pretransform_fn, transform_fn, trans_count, init_temp=initial_temp,
-                                            enable_negative_transforms=False, add_scalable_noise_to_transforms=True, init_scalar=.1)
+                                            add_scalable_noise_to_transforms=True, init_scalar=.1)
 
         self.transformation_counts = trans_count
         self.init_temperature = initial_temp

codes/models/archs/arch_util.py

@@ -219,7 +219,7 @@ class ConvBnRelu(nn.Module):
 ''' Convenience class with Conv->BN->SiLU. Includes weight initialization and auto-padding for standard
     kernel sizes. '''
 class ConvBnSilu(nn.Module):
-    def __init__(self, filters_in, filters_out, kernel_size=3, stride=1, silu=True, bn=True, bias=True):
+    def __init__(self, filters_in, filters_out, kernel_size=3, stride=1, silu=True, bn=True, bias=True, weight_init_factor=1):
         super(ConvBnSilu, self).__init__()
         padding_map = {1: 0, 3: 1, 5: 2, 7: 3}
         assert kernel_size in padding_map.keys()
@@ -237,6 +237,9 @@ class ConvBnSilu(nn.Module):
         for m in self.modules():
             if isinstance(m, nn.Conv2d):
                 nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu' if self.silu else 'linear')
+                m.weight.data *= weight_init_factor
+                if m.bias is not None:
+                    m.bias.data.zero_()
             elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                 nn.init.constant_(m.weight, 1)
                 nn.init.constant_(m.bias, 0)
@@ -254,7 +257,7 @@ class ConvBnSilu(nn.Module):
 ''' Convenience class with Conv->BN->LeakyReLU. Includes weight initialization and auto-padding for standard
     kernel sizes. '''
 class ConvBnLelu(nn.Module):
-    def __init__(self, filters_in, filters_out, kernel_size=3, stride=1, lelu=True, bn=True, bias=True):
+    def __init__(self, filters_in, filters_out, kernel_size=3, stride=1, lelu=True, bn=True, bias=True, weight_init_factor=1):
         super(ConvBnLelu, self).__init__()
         padding_map = {1: 0, 3: 1, 5: 2, 7: 3}
         assert kernel_size in padding_map.keys()
@@ -273,6 +276,9 @@ class ConvBnLelu(nn.Module):
             if isinstance(m, nn.Conv2d):
                 nn.init.kaiming_normal_(m.weight, a=.1, mode='fan_out',
                                         nonlinearity='leaky_relu' if self.lelu else 'linear')
+                m.weight.data *= weight_init_factor
+                if m.bias is not None:
+                    m.bias.data.zero_()
             elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                 nn.init.constant_(m.weight, 1)
                 nn.init.constant_(m.bias, 0)
@@ -321,3 +327,40 @@ class ConvGnLelu(nn.Module):
             return self.lelu(x)
         else:
             return x
+
+''' Convenience class with Conv->BN->SiLU. Includes weight initialization and auto-padding for standard
+    kernel sizes. '''
+class ConvGnSilu(nn.Module):
+    def __init__(self, filters_in, filters_out, kernel_size=3, stride=1, silu=True, gn=True, bias=True, num_groups=8, weight_init_factor=1):
+        super(ConvGnSilu, self).__init__()
+        padding_map = {1: 0, 3: 1, 5: 2, 7: 3}
+        assert kernel_size in padding_map.keys()
+        self.conv = nn.Conv2d(filters_in, filters_out, kernel_size, stride, padding_map[kernel_size], bias=bias)
+        if gn:
+            self.gn = nn.GroupNorm(num_groups, filters_out)
+        else:
+            self.gn = None
+        if silu:
+            self.silu = SiLU()
+        else:
+            self.silu = None
+
+        # Init params.
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu' if self.silu else 'linear')
+                m.weight.data *= weight_init_factor
+                if m.bias is not None:
+                    m.bias.data.zero_()
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        x = self.conv(x)
+        if self.gn:
+            x = self.gn(x)
+        if self.silu:
+            return self.silu(x)
+        else:
+            return x

codes/models/networks.py

@@ -59,7 +59,7 @@ def define_G(opt, net_key='network_G'):
                                                                       heightened_temp_min=opt_net['heightened_temp_min'], heightened_final_step=opt_net['heightened_final_step'],
                                                                       upsample_factor=scale, add_scalable_noise_to_transforms=opt_net['add_noise'])
     elif which_model == "ConfigurableSwitchedResidualGenerator2":
-        netG = SwitchedGen_arch.ConfigurableSwitchedResidualGenerator2(switch_filters=opt_net['switch_filters'], switch_growths=opt_net['switch_growths'],
+        netG = SwitchedGen_arch.ConfigurableSwitchedResidualGenerator2(switch_depth=opt_net['switch_depth'], switch_filters=opt_net['switch_filters'],
                                                                       switch_reductions=opt_net['switch_reductions'],
                                                                       switch_processing_layers=opt_net['switch_processing_layers'], trans_counts=opt_net['trans_counts'],
                                                                       trans_kernel_sizes=opt_net['trans_kernel_sizes'], trans_layers=opt_net['trans_layers'],

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_div2k_pixgan_rrdb.yml')
+    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_div2k_pixgan_srg2.yml')
     parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none',
                         help='job launcher')
     parser.add_argument('--local_rank', type=int, default=0)

codes/utils/numeric_stability.py

@@ -97,20 +97,19 @@ if __name__ == "__main__":
                    torch.randn(1, 3, 64, 64),
                    device='cuda')
     '''
-    '''
     test_stability(functools.partial(srg.ConfigurableSwitchedResidualGenerator2,
-                                     switch_filters=[32,32,32,32],
+                                     switch_depth=4,
-                                     switch_growths=[16,16,16,16],
+                                     switch_filters=64,
-                                     switch_reductions=[4,3,2,1],
+                                     switch_reductions=4,
-                                     switch_processing_layers=[3,3,4,5],
+                                     switch_processing_layers=2,
-                                     trans_counts=[16,16,16,16,16],
+                                     trans_counts=8,
-                                     trans_kernel_sizes=[3,3,3,3,3],
+                                     trans_kernel_sizes=3,
-                                     trans_layers=[3,3,3,3,3],
+                                     trans_layers=4,
                                      transformation_filters=64,
-                                     initial_temp=10),
+                                     upsample_factor=4),
                    torch.randn(1, 3, 64, 64),
                    device='cuda')
-    '''
+
     '''
     test_stability(functools.partial(srg1.ConfigurableSwitchedResidualGenerator,
                                      switch_filters=[32,32,32,32],
@@ -125,7 +124,9 @@ if __name__ == "__main__":
                    torch.randn(1, 3, 64, 64),
                    device='cuda')
                    '''
+    '''
     test_stability(functools.partial(srg.ConfigurableSwitchedResidualGenerator3,
                                      64, 16),
                    torch.randn(1, 3, 64, 64),
                    device='cuda')
+    '''