update ENAS

lib/models/__init__.py

@@ -16,18 +16,10 @@ from .cell_searchs import CellStructure, CellArchitectures
 
 # Cell-based NAS Models
 def get_cell_based_tiny_net(config):
-  if config.name == 'DARTS-V1':
-    from .cell_searchs import TinyNetworkDartsV1
-    return TinyNetworkDartsV1(config.C, config.N, config.max_nodes, config.num_classes, config.space)
-  elif config.name == 'DARTS-V2':
-    from .cell_searchs import TinyNetworkDartsV2
-    return TinyNetworkDartsV2(config.C, config.N, config.max_nodes, config.num_classes, config.space)
-  elif config.name == 'GDAS':
-    from .cell_searchs import TinyNetworkGDAS
-    return TinyNetworkGDAS(config.C, config.N, config.max_nodes, config.num_classes, config.space)
-  elif config.name == 'SETN':
-    from .cell_searchs import TinyNetworkSETN
-    return TinyNetworkSETN(config.C, config.N, config.max_nodes, config.num_classes, config.space)
+  group_names = ['DARTS-V1', 'DARTS-V2', 'GDAS', 'SETN', 'ENAS']
+  from .cell_searchs import nas_super_nets
+  if config.name in group_names:
+    return nas_super_nets[config.name](config.C, config.N, config.max_nodes, config.num_classes, config.space)
   elif config.name == 'infer.tiny':
     from .cell_infers import TinyNetwork
     return TinyNetwork(config.C, config.N, config.genotype, config.num_classes)

lib/models/cell_searchs/__init__.py

@@ -2,4 +2,11 @@ from .search_model_darts_v1 import TinyNetworkDartsV1
 from .search_model_darts_v2 import TinyNetworkDartsV2
 from .search_model_gdas     import TinyNetworkGDAS
 from .search_model_setn     import TinyNetworkSETN
+from .search_model_enas     import TinyNetworkENAS
 from .genotypes             import Structure as CellStructure, architectures as CellArchitectures
+
+nas_super_nets = {'DARTS-V1': TinyNetworkDartsV1,
+                  'DARTS-V2': TinyNetworkDartsV2,
+                  'GDAS'    : TinyNetworkGDAS,
+                  'SETN'    : TinyNetworkSETN,
+                  'ENAS'    : TinyNetworkENAS}

lib/models/cell_searchs/_test_module.py

@@ -0,0 +1,9 @@
+import torch
+from search_model_enas_utils import Controller
+
+def main():
+  controller = Controller(6, 4)
+  predictions = controller()
+
+if __name__ == '__main__':
+  main()

lib/models/cell_searchs/search_model_enas.py

@@ -0,0 +1,94 @@
+##################################################
+# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2019 #
+##########################################################################
+# Efficient Neural Architecture Search via Parameters Sharing, ICML 2018 #
+##########################################################################
+import torch
+import torch.nn as nn
+from copy import deepcopy
+from ..cell_operations import ResNetBasicblock
+from .search_cells     import SearchCell
+from .genotypes        import Structure
+from .search_model_enas_utils import Controller
+
+
+class TinyNetworkENAS(nn.Module):
+
+  def __init__(self, C, N, max_nodes, num_classes, search_space):
+    super(TinyNetworkENAS, self).__init__()
+    self._C        = C
+    self._layerN   = N
+    self.max_nodes = max_nodes
+    self.stem = nn.Sequential(
+                    nn.Conv2d(3, C, kernel_size=3, padding=1, bias=False),
+                    nn.BatchNorm2d(C))
+  
+    layer_channels   = [C    ] * N + [C*2 ] + [C*2  ] * N + [C*4 ] + [C*4  ] * N    
+    layer_reductions = [False] * N + [True] + [False] * N + [True] + [False] * N
+
+    C_prev, num_edge, edge2index = C, None, None
+    self.cells = nn.ModuleList()
+    for index, (C_curr, reduction) in enumerate(zip(layer_channels, layer_reductions)):
+      if reduction:
+        cell = ResNetBasicblock(C_prev, C_curr, 2)
+      else:
+        cell = SearchCell(C_prev, C_curr, 1, max_nodes, search_space)
+        if num_edge is None: num_edge, edge2index = cell.num_edges, cell.edge2index
+        else: assert num_edge == cell.num_edges and edge2index == cell.edge2index, 'invalid {:} vs. {:}.'.format(num_edge, cell.num_edges)
+      self.cells.append( cell )
+      C_prev = cell.out_dim
+    self.op_names   = deepcopy( search_space )
+    self._Layer     = len(self.cells)
+    self.edge2index = edge2index
+    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)
+    # to maintain the sampled architecture
+    self.sampled_arch = None
+
+  def update_arch(self, _arch):
+    if _arch is None:
+      self.sampled_arch = None
+    elif isinstance(_arch, Structure):
+      self.sampled_arch = _arch
+    elif isinstance(_arch, (list, tuple)):
+      genotypes = []
+      for i in range(1, self.max_nodes):
+        xlist = []
+        for j in range(i):
+          node_str = '{:}<-{:}'.format(i, j)
+          op_index = _arch[ self.edge2index[node_str] ]
+          op_name  = self.op_names[ op_index ]
+          xlist.append((op_name, j))
+        genotypes.append( tuple(xlist) )
+      self.sampled_arch = Structure(genotypes)
+    else:
+      raise ValueError('invalid type of input architecture : {:}'.format(_arch))
+    return self.sampled_arch
+    
+  def create_controller(self):
+    return Controller(len(self.edge2index), len(self.op_names))
+
+  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}, Max-Nodes={max_nodes}, N={_layerN}, L={_Layer})'.format(name=self.__class__.__name__, **self.__dict__))
+
+  def forward(self, inputs):
+
+    feature = self.stem(inputs)
+    for i, cell in enumerate(self.cells):
+      if isinstance(cell, SearchCell):
+        feature = cell.forward_dynamic(feature, self.sampled_arch)
+      else: feature = cell(feature)
+
+    out = self.lastact(feature)
+    out = self.global_pooling( out )
+    out = out.view(out.size(0), -1)
+    logits = self.classifier(out)
+
+    return out, logits

lib/models/cell_searchs/search_model_enas_utils.py

@@ -0,0 +1,55 @@
+##################################################
+# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2019 #
+##########################################################################
+# Efficient Neural Architecture Search via Parameters Sharing, ICML 2018 #
+##########################################################################
+import torch
+import torch.nn as nn
+from torch.distributions.categorical import Categorical
+
+class Controller(nn.Module):
+  # we refer to https://github.com/TDeVries/enas_pytorch/blob/master/models/controller.py
+  def __init__(self, num_edge, num_ops, lstm_size=32, lstm_num_layers=2, tanh_constant=2.5, temperature=5.0):
+    super(Controller, self).__init__()
+    # assign the attributes
+    self.num_edge  = num_edge
+    self.num_ops   = num_ops
+    self.lstm_size = lstm_size
+    self.lstm_N    = lstm_num_layers
+    self.tanh_constant = tanh_constant
+    self.temperature   = temperature
+    # create parameters
+    self.register_parameter('input_vars', nn.Parameter(torch.Tensor(1, 1, lstm_size)))
+    self.w_lstm = nn.LSTM(input_size=self.lstm_size, hidden_size=self.lstm_size, num_layers=self.lstm_N)
+    self.w_embd = nn.Embedding(self.num_ops, self.lstm_size)
+    self.w_pred = nn.Linear(self.lstm_size, self.num_ops)
+
+    nn.init.uniform_(self.input_vars         , -0.1, 0.1)
+    nn.init.uniform_(self.w_lstm.weight_hh_l0, -0.1, 0.1)
+    nn.init.uniform_(self.w_lstm.weight_ih_l0, -0.1, 0.1)
+    nn.init.uniform_(self.w_embd.weight      , -0.1, 0.1)
+    nn.init.uniform_(self.w_pred.weight      , -0.1, 0.1)
+
+  def forward(self):
+
+    inputs, h0 = self.input_vars, None
+    log_probs, entropys, sampled_arch = [], [], []
+    for iedge in range(self.num_edge):
+      outputs, h0 = self.w_lstm(inputs, h0)
+      
+      logits = self.w_pred(outputs)
+      logits = logits / self.temperature
+      logits = self.tanh_constant * torch.tanh(logits)
+      # distribution
+      op_distribution = Categorical(logits=logits)
+      op_index    = op_distribution.sample()
+      sampled_arch.append( op_index.item() )
+
+      op_log_prob = op_distribution.log_prob(op_index)
+      log_probs.append( op_log_prob.view(-1) )
+      op_entropy  = op_distribution.entropy()
+      entropys.append( op_entropy.view(-1) )
+      
+      # obtain the input embedding for the next step
+      inputs = self.w_embd(op_index)
+    return torch.sum(torch.cat(log_probs)), torch.sum(torch.cat(entropys)), sampled_arch