xautodl/lib/models/cell_searchs/search_model_gdas_nasnet.py

###########################################################################
# Searching for A Robust Neural Architecture in Four GPU Hours, CVPR 2019 #
###########################################################################
import torch
import torch.nn as nn
from copy import deepcopy
from .search_cells     import NASNetSearchCell as SearchCell
from .genotypes        import Structure


# The macro structure is based on NASNet
class NASNetworkGDAS(nn.Module):

  def __init__(self, C, N, steps, multiplier, stem_multiplier, num_classes, search_space, affine, track_running_stats):
    super(NASNetworkGDAS, self).__init__()
    self._C        = C
    self._layerN   = N
    self._steps    = steps
    self._multiplier = multiplier
    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)

    num_edge, edge2index = None, None
    C_prev_prev, C_prev, C_curr, reduction_prev = C*stem_multiplier, C*stem_multiplier, C, False

    self.cells = nn.ModuleList()
    for index, (C_curr, reduction) in enumerate(zip(layer_channels, layer_reductions)):
      cell = SearchCell(search_space, steps, multiplier, C_prev_prev, C_prev, C_curr, reduction, reduction_prev, affine, track_running_stats)
      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_prev, C_prev, reduction_prev = C_prev, multiplier*C_curr, reduction
    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)
    self.arch_normal_parameters = nn.Parameter( 1e-3*torch.randn(num_edge, len(search_space)) )
    self.arch_reduce_parameters = nn.Parameter( 1e-3*torch.randn(num_edge, len(search_space)) )
    self.tau        = 10

  def get_weights(self):
    xlist = list( self.stem.parameters() ) + list( self.cells.parameters() )
    xlist+= list( self.lastact.parameters() ) + list( self.global_pooling.parameters() )
    xlist+= list( self.classifier.parameters() )
    return xlist

  def set_tau(self, tau):
    self.tau = tau

  def get_tau(self):
    return self.tau

  def get_alphas(self):
    return [self.arch_normal_parameters, self.arch_reduce_parameters]

  def show_alphas(self):
    with torch.no_grad():
      A = 'arch-normal-parameters :\n{:}'.format( nn.functional.softmax(self.arch_normal_parameters, dim=-1).cpu() )
      B = 'arch-reduce-parameters :\n{:}'.format( nn.functional.softmax(self.arch_reduce_parameters, dim=-1).cpu() )
    return '{:}\n{:}'.format(A, B)

  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}, steps={_steps}, multiplier={_multiplier}, L={_Layer})'.format(name=self.__class__.__name__, **self.__dict__))

  def genotype(self):
    def _parse(weights):
      gene = []
      for i in range(self._steps):
        edges = []
        for j in range(2+i):
          node_str = '{:}<-{:}'.format(i, j)
          ws = weights[ self.edge2index[node_str] ]
          for k, op_name in enumerate(self.op_names):
            if op_name == 'none': continue
            edges.append( (op_name, j, ws[k]) )
        edges = sorted(edges, key=lambda x: -x[-1])
        selected_edges = edges[:2]
        gene.append( tuple(selected_edges) )
      return gene
    with torch.no_grad():
      gene_normal = _parse(torch.softmax(self.arch_normal_parameters, dim=-1).cpu().numpy())
      gene_reduce = _parse(torch.softmax(self.arch_reduce_parameters, dim=-1).cpu().numpy())
    return {'normal': gene_normal, 'normal_concat': list(range(2+self._steps-self._multiplier, self._steps+2)),
            'reduce': gene_reduce, 'reduce_concat': list(range(2+self._steps-self._multiplier, self._steps+2))}

  def forward(self, inputs):
    def get_gumbel_prob(xins):
      while True:
        gumbels = -torch.empty_like(xins).exponential_().log()
        logits  = (xins.log_softmax(dim=1) + gumbels) / self.tau
        probs   = nn.functional.softmax(logits, dim=1)
        index   = probs.max(-1, keepdim=True)[1]
        one_h   = torch.zeros_like(logits).scatter_(-1, index, 1.0)
        hardwts = one_h - probs.detach() + probs
        if (torch.isinf(gumbels).any()) or (torch.isinf(probs).any()) or (torch.isnan(probs).any()):
          continue
        else: break
      return hardwts, index

    normal_hardwts, normal_index = get_gumbel_prob(self.arch_normal_parameters)
    reduce_hardwts, reduce_index = get_gumbel_prob(self.arch_reduce_parameters)

    s0 = s1 = self.stem(inputs)
    for i, cell in enumerate(self.cells):
      if cell.reduction: hardwts, index = reduce_hardwts, reduce_index
      else             : hardwts, index = normal_hardwts, normal_index
      s0, s1 = s1, cell.forward_gdas(s0, s1, hardwts, index)
    out = self.lastact(s1)
    out = self.global_pooling( out )
    out = out.view(out.size(0), -1)
    logits = self.classifier(out)

    return out, logits