pt.darts源码分析

解释博客：

Intuitive Explanation of Differentiable Architecture Search (DARTS)

SearchCNN 构建

SearchCNNController

models.search_cnn.SearchCNNController

n_ops = 8, n_nodes = 4

之所有有个i + 2是因为include 2 input nodes， 一个是上一个输入，一个是跨层连接的上上个输入。

for i in range(n_nodes):
    self.alpha_normal.append(nn.Parameter(1e-3*torch.randn
    (i+2, n_ops)))
    self.alpha_reduce.append(nn.Parameter(1e-3*torch.randn
    (i+2, n_ops)))

Variable	Value
C_in	3
C	16
n_classes	10
n_layers	8
n_nodes	4
stem_multiplier	3

SearchCNNController构造函数的最后：

self.net = SearchCNN(C_in, C, n_classes, n_layers, n_nodes, stem_multiplier)

SearchCNN

每个cell有两个输入，一个是上一个输入，一个是跨层连接的上上个输入。

刚开始c_pp = c_p =48，c_cur = 16

通过跨层连接将cells组织起来：

SearchCell中nodes之间DAG的构建

models.search_cells.SearchCell#__init__

op = ops.MixedOp(C, stride)
self.dag[i].append(op)

看到MixedOp内部

self._ops = nn.ModuleList()
for primitive in gt.PRIMITIVES:
    op = OPS[primitive](C, stride, affine=False)
    self._ops.append(op)

# genotypes.py
PRIMITIVES = [
    'max_pool_3x3',
    'avg_pool_3x3',
    'skip_connect', # identity
    'sep_conv_3x3',
    'sep_conv_5x5',
    'dil_conv_3x3',
    'dil_conv_5x5',
    'none'
]

models.search_cells.SearchCell#forward

models/search_cells.py:53

s.shape
Out[2]: torch.Size([