Add bypass to ChainedEmbeddingGen

codes/models/archs/ChainedEmbeddingGen.py

@@ -1,4 +1,7 @@
+import os
+
 import torch
+import torchvision
 from torch import nn
 
 from models.archs.SPSR_arch import ImageGradientNoPadding
@@ -46,7 +49,7 @@ class BasicEmbeddingPyramid(nn.Module):
         for i in range(2):
             p = self.expanders[i](p, identities[-(i+1)])
         x = self.final_process(torch.cat([x, p], dim=1))
-        return x
+        return x, p
 
 
 class ChainedEmbeddingGen(nn.Module):
@@ -61,7 +64,7 @@ class ChainedEmbeddingGen(nn.Module):
         fea = self.initial_conv(x)
         emb = checkpoint(self.spine, fea)
         for block in self.blocks:
-            fea = fea + checkpoint(block, fea, *emb)
+            fea = fea + checkpoint(block, fea, *emb)[0]
         return checkpoint(self.upsample, fea),
 
 
@@ -100,7 +103,7 @@ class ChainedEmbeddingGenWithStructure(nn.Module):
         emb = checkpoint(self.spine, fea)
         grad = fea
         for i, block in enumerate(self.blocks):
-            fea = fea + checkpoint(block, fea, *emb)
+            fea = fea + checkpoint(block, fea, *emb)[0]
             if i < 3:
                 structure_br = checkpoint(self.structure_joins[i], grad, fea)
                 grad = grad + checkpoint(self.structure_blocks[i], structure_br)
@@ -109,3 +112,76 @@ class ChainedEmbeddingGenWithStructure(nn.Module):
 
     def get_debug_values(self, step, net_name):
         return { 'ref_join_std': self.ref_join_std }
+
+
+# This is a structural block that learns to mute regions of a residual transformation given a signal.
+class OptionalPassthroughBlock(nn.Module):
+    def __init__(self, nf, initial_bias=10):
+        super(OptionalPassthroughBlock, self).__init__()
+        self.switch_process = nn.Sequential(ConvGnLelu(nf, nf // 2, 1, activation=False, norm=False, bias=False),
+                                       ConvGnLelu(nf // 2, nf // 4, 1, activation=False, norm=False, bias=False),
+                                       ConvGnLelu(nf // 4, 1, 1, activation=False, norm=False, bias=False))
+        self.bias = nn.Parameter(torch.tensor(initial_bias, dtype=torch.float), requires_grad=True)
+        self.activation = nn.Sigmoid()
+
+    def forward(self, x, switch_signal):
+        switch = self.switch_process(switch_signal)
+        bypass_map = self.activation(self.bias + switch)
+        return x * bypass_map, bypass_map
+
+
+class StructuredChainedEmbeddingGenWithBypass(nn.Module):
+    def __init__(self, depth=10, recurrent=False, recurrent_nf=3, recurrent_stride=2, bypass_bias=10):
+        super(StructuredChainedEmbeddingGenWithBypass, self).__init__()
+        self.recurrent = recurrent
+        self.initial_conv = ConvGnLelu(3, 64, kernel_size=7, bias=True, norm=False, activation=False)
+        if recurrent:
+            self.recurrent_nf = recurrent_nf
+            self.recurrent_stride = recurrent_stride
+            self.recurrent_process = ConvGnLelu(recurrent_nf, 64, kernel_size=3, stride=recurrent_stride, norm=False, bias=True, activation=False)
+            self.recurrent_join = ReferenceJoinBlock(64, residual_weight_init_factor=.01, final_norm=False, kernel_size=1, depth=3, join=False)
+        self.spine = SpineNet(arch='49', output_level=[3, 4], double_reduce_early=False)
+        self.blocks = nn.ModuleList([BasicEmbeddingPyramid() for i in range(depth)])
+        self.bypasses = nn.ModuleList([OptionalPassthroughBlock(64, initial_bias=bypass_bias) for i in range(depth)])
+        self.structure_joins = nn.ModuleList([ConjoinBlock(64) for i in range(3)])
+        self.structure_blocks = nn.ModuleList([ConvGnLelu(64, 64, kernel_size=3, bias=False, norm=False, activation=False, weight_init_factor=.1) for i in range(3)])
+        self.structure_upsample = FinalUpsampleBlock2x(64)
+        self.grad_extract = ImageGradientNoPadding()
+        self.upsample = FinalUpsampleBlock2x(64)
+        self.ref_join_std = 0
+        self.bypass_maps = []
+
+    def forward(self, x, recurrent=None):
+        fea = self.initial_conv(x)
+        if self.recurrent:
+            if recurrent is None:
+                if self.recurrent_nf == 3:
+                    recurrent = torch.zeros_like(x)
+                    if self.recurrent_stride != 1:
+                        recurrent = torch.nn.functional.interpolate(recurrent, scale_factor=self.recurrent_stride, mode='nearest')
+                else:
+                    recurrent = torch.zeros_like(fea)
+            rec = self.recurrent_process(recurrent)
+            fea, recstd = self.recurrent_join(fea, rec)
+            self.ref_join_std = recstd.item()
+        emb = checkpoint(self.spine, fea)
+        grad = fea
+        self.bypass_maps = []
+        for i, block in enumerate(self.blocks):
+            residual, context = checkpoint(block, fea, *emb)
+            residual, bypass_map = checkpoint(self.bypasses[i], residual, context)
+            fea = fea + residual
+            self.bypass_maps.append(bypass_map.detach())
+            if i < 3:
+                structure_br = checkpoint(self.structure_joins[i], grad, fea)
+                grad = grad + checkpoint(self.structure_blocks[i], structure_br)
+        out = checkpoint(self.upsample, fea)
+        return out, self.grad_extract(checkpoint(self.structure_upsample, grad)), self.grad_extract(out), fea
+
+    def visual_dbg(self, step, path):
+        for i, bm in enumerate(self.bypass_maps):
+            torchvision.utils.save_image(bm.cpu(), os.path.join(path, "%i_bypass_%i.png" % (step, i+1)))
+
+    def get_debug_values(self, step, net_name):
+        biases = [b.bias.item() for b in self.bypasses]
+        return { 'ref_join_std': self.ref_join_std, 'bypass_biases': sum(biases) / len(biases) }

codes/models/networks.py

@@ -18,7 +18,8 @@ import models.archs.discriminator_vgg_arch as SRGAN_arch
 import models.archs.feature_arch as feature_arch
 import models.archs.panet.panet as panet
 import models.archs.rcan as rcan
-from models.archs.ChainedEmbeddingGen import ChainedEmbeddingGen, ChainedEmbeddingGenWithStructure
+from models.archs.ChainedEmbeddingGen import ChainedEmbeddingGen, ChainedEmbeddingGenWithStructure, \
+    StructuredChainedEmbeddingGenWithBypass
 
 logger = logging.getLogger('base')
 
@@ -130,6 +131,12 @@ def define_G(opt, net_key='network_G', scale=None):
         recnf = opt_net['recurrent_nf'] if 'recurrent_nf' in opt_net.keys() else 3
         recstd = opt_net['recurrent_stride'] if 'recurrent_stride' in opt_net.keys() else 2
         netG = ChainedEmbeddingGenWithStructure(depth=opt_net['depth'], recurrent=rec, recurrent_nf=recnf, recurrent_stride=recstd)
+    elif which_model == 'chained_gen_structured_with_bypass':
+        rec = opt_net['recurrent'] if 'recurrent' in opt_net.keys() else False
+        recnf = opt_net['recurrent_nf'] if 'recurrent_nf' in opt_net.keys() else 3
+        recstd = opt_net['recurrent_stride'] if 'recurrent_stride' in opt_net.keys() else 2
+        bypass_bias = opt_net['bypass_bias'] if 'bypass_bias' in opt_net.keys() else 0
+        netG = StructuredChainedEmbeddingGenWithBypass(depth=opt_net['depth'], recurrent=rec, recurrent_nf=recnf, recurrent_stride=recstd, bypass_bias=bypass_bias)
     elif which_model == "flownet2":
         from models.flownet2.models import FlowNet2
         ld = torch.load(opt_net['load_path'])