update CVPR-2019-GDAS re-train NASNet-search-space searched models

lib/models/cell_infers/__init__.py

@@ -2,3 +2,4 @@
 # Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2019.01 #
 #####################################################
 from .tiny_network import TinyNetwork
+from .nasnet_cifar import NASNetonCIFAR

lib/models/cell_infers/cells.py

@@ -2,6 +2,7 @@
 # Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2019.01 #
 #####################################################
 
+import torch
 import torch.nn as nn
 from copy import deepcopy
 from ..cell_operations import OPS
@@ -50,3 +51,70 @@ class InferCell(nn.Module):
       node_feature = sum( self.layers[_il](nodes[_ii]) for _il, _ii in zip(node_layers, node_innods) )
       nodes.append( node_feature )
     return nodes[-1]
+
+
+
+# Learning Transferable Architectures for Scalable Image Recognition, CVPR 2018
+class NASNetInferCell(nn.Module):
+
+  def __init__(self, genotype, C_prev_prev, C_prev, C, reduction, reduction_prev, affine, track_running_stats):
+    super(NASNetInferCell, self).__init__()
+    self.reduction = reduction
+    if reduction_prev: self.preprocess0 = OPS['skip_connect'](C_prev_prev, C, 2, affine, track_running_stats)
+    else             : self.preprocess0 = OPS['nor_conv_1x1'](C_prev_prev, C, 1, affine, track_running_stats)
+    self.preprocess1 = OPS['nor_conv_1x1'](C_prev, C, 1, affine, track_running_stats)
+
+    if not reduction:
+      nodes, concats = genotype['normal'], genotype['normal_concat']
+    else:
+      nodes, concats = genotype['reduce'], genotype['reduce_concat']
+    self._multiplier = len(concats)
+    self._concats = concats
+    self._steps = len(nodes)
+    self._nodes = nodes
+    self.edges = nn.ModuleDict()
+    for i, node in enumerate(nodes):
+      for in_node in node:
+        name, j = in_node[0], in_node[1]
+        stride = 2 if reduction and j < 2 else 1
+        node_str = '{:}<-{:}'.format(i+2, j)
+        self.edges[node_str] = OPS[name](C, C, stride, affine, track_running_stats)
+
+  # [TODO] to support drop_prob in this function..
+  def forward(self, s0, s1, unused_drop_prob):
+    s0 = self.preprocess0(s0)
+    s1 = self.preprocess1(s1)
+
+    states = [s0, s1]
+    for i, node in enumerate(self._nodes):
+      clist = []
+      for in_node in node:
+        name, j = in_node[0], in_node[1]
+        node_str = '{:}<-{:}'.format(i+2, j)
+        op = self.edges[ node_str ]
+        clist.append( op(states[j]) )
+      states.append( sum(clist) )
+    return torch.cat([states[x] for x in self._concats], dim=1)
+
+
+class AuxiliaryHeadCIFAR(nn.Module):
+
+  def __init__(self, C, num_classes):
+    """assuming input size 8x8"""
+    super(AuxiliaryHeadCIFAR, self).__init__()
+    self.features = nn.Sequential(
+      nn.ReLU(inplace=True),
+      nn.AvgPool2d(5, stride=3, padding=0, count_include_pad=False), # image size = 2 x 2
+      nn.Conv2d(C, 128, 1, bias=False),
+      nn.BatchNorm2d(128),
+      nn.ReLU(inplace=True),
+      nn.Conv2d(128, 768, 2, bias=False),
+      nn.BatchNorm2d(768),
+      nn.ReLU(inplace=True)
+    )
+    self.classifier = nn.Linear(768, num_classes)
+
+  def forward(self, x):
+    x = self.features(x)
+    x = self.classifier(x.view(x.size(0),-1))
+    return x

lib/models/cell_infers/nasnet_cifar.py

@@ -0,0 +1,71 @@
+#####################################################
+# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2019.01 #
+#####################################################
+import torch
+import torch.nn as nn
+from copy import deepcopy
+from .cells import NASNetInferCell as InferCell, AuxiliaryHeadCIFAR
+
+
+# The macro structure is based on NASNet
+class NASNetonCIFAR(nn.Module):
+
+  def __init__(self, C, N, stem_multiplier, num_classes, genotype, auxiliary, affine=True, track_running_stats=True):
+    super(NASNetonCIFAR, self).__init__()
+    self._C        = C
+    self._layerN   = N
+    self.stem = nn.Sequential(
+                    nn.Conv2d(3, C*stem_multiplier, kernel_size=3, padding=1, bias=False),
+                    nn.BatchNorm2d(C*stem_multiplier))
+  
+    # config for each layer
+    layer_channels   = [C    ] * N + [C*2 ] + [C*2  ] * (N-1) + [C*4 ] + [C*4  ] * (N-1)
+    layer_reductions = [False] * N + [True] + [False] * (N-1) + [True] + [False] * (N-1)
+
+    C_prev_prev, C_prev, C_curr, reduction_prev = C*stem_multiplier, C*stem_multiplier, C, False
+    self.auxiliary_index = None
+    self.auxiliary_head  = None
+    self.cells = nn.ModuleList()
+    for index, (C_curr, reduction) in enumerate(zip(layer_channels, layer_reductions)):
+      cell = InferCell(genotype, C_prev_prev, C_prev, C_curr, reduction, reduction_prev, affine, track_running_stats)
+      self.cells.append( cell )
+      C_prev_prev, C_prev, reduction_prev = C_prev, cell._multiplier*C_curr, reduction
+      if reduction and C_curr == C*4 and auxiliary:
+        self.auxiliary_head = AuxiliaryHeadCIFAR(C_prev, num_classes)
+        self.auxiliary_index = index
+    self._Layer     = len(self.cells)
+    self.lastact    = nn.Sequential(nn.BatchNorm2d(C_prev), nn.ReLU(inplace=True))
+    self.global_pooling = nn.AdaptiveAvgPool2d(1)
+    self.classifier = nn.Linear(C_prev, num_classes)
+    self.drop_path_prob = -1
+
+  def update_drop_path(self, drop_path_prob):
+    self.drop_path_prob = drop_path_prob
+
+  def auxiliary_param(self):
+    if self.auxiliary_head is None: return []
+    else: return list( self.auxiliary_head.parameters() )
+
+  def get_message(self):
+    string = self.extra_repr()
+    for i, cell in enumerate(self.cells):
+      string += '\n {:02d}/{:02d} :: {:}'.format(i, len(self.cells), cell.extra_repr())
+    return string
+
+  def extra_repr(self):
+    return ('{name}(C={_C}, N={_layerN}, L={_Layer})'.format(name=self.__class__.__name__, **self.__dict__))
+
+  def forward(self, inputs):
+    stem_feature, logits_aux = self.stem(inputs), None
+    cell_results = [stem_feature, stem_feature]
+    for i, cell in enumerate(self.cells):
+      cell_feature = cell(cell_results[-2], cell_results[-1], self.drop_path_prob)
+      cell_results.append( cell_feature )
+      if self.auxiliary_index is not None and i == self.auxiliary_index and self.training:
+        logits_aux = self.auxiliary_head( cell_results[-1] )
+    out = self.lastact(cell_results[-1])
+    out = self.global_pooling( out )
+    out = out.view(out.size(0), -1)
+    logits = self.classifier(out)
+    if logits_aux is None: return out, logits
+    else: return out, [logits, logits_aux]