P-DARTS
2019-ICCV-Progressive Differentiable Architecture Search Bridging the Depth Gap Between Search and Evaluation
- Tongji University && Huawei
- GitHub： 200+ stars
- Citation：49
Motivation
Question：
DARTS has to search the architecture in a shallow network while evaluate in a deeper one.
DARTS在浅层网络上搜索，在深层网络上评估（cifar search in 8-depth， eval in 20-depth）。
This brings an issue named the depth gap (see Figure 1(a)), which means that the search stage finds some operations that work well in a shallow architecture, but the evaluation stage actually prefers other operations that fit a deep architecture better.
Such gap hinders these approaches in their application to more complex visual recognition tasks.
Contribution
propose Progressive DARTS (P-DARTS), a novel and efficient algorithm to bridge the depth gap.
Bring two questions：
Q1： While a deeper architecture requires heavier computational overhead
we propose search space approximation which, as the depth increases, reduces the number of candidates (operations) according to their scores in the elapsed search process.
Q2：Another issue, lack of stability, emerges with searching over a deep architecture, in which the algorithm can be biased heavily towards skip-connect as it often leads to rapidest error decay during optimization, but, actually, a better option often resides in learnable operations such as convolution.
we propose search space regularization, which (i) introduces operation-level Dropout [25] to alleviate the dominance of skip-connect during training, and (ii) controls the appearance of skip-connect during evaluation.
Method
search space approximation
在初始阶段，搜索网络相对较浅，但是cell中每条边上的候选操作最多（所有操作）。在阶段 \(S_{k-1}\)中，根据学习到的网络结构参数（权值）来排序并筛选出权值（重要性）较高的 \(O_k\) 个操作，并由此搭建一个拥有 \(L_k\) 个cell的搜索网络用于下一阶段的搜索，其中， \(L_k > L_{k-1} , O_k < O_{k-1}\) .
这个过程可以渐进而持续地增加搜索网络的深度，直到足够接近测试网络深度。
search space regularization
we observe that information prefers to flow through skip-connect instead of convolution or pooling, which is arguably due to the reason that skip-connect often leads to rapid gradient descent.
实验结果表明在本文采用的框架下，信息往往倾向于通过skip-connect流动，而不是卷积。这是因为skip-connect通常处在梯度下降最速的路径上。
the search process tends to generate architectures with many skip-connect operations, which limits the number of learnable parameters and thus produces unsatisfying performance at the evaluation stage.
在这种情况下，最终搜索得到的结构往往包含大量的skip-connect操作，可训练参数较少，从而使得性能下降。
We address this problem by search space regularization, which consists of two parts.
First, we insert operation-level Dropout [25] after each skip-connect operation, so as to partially ‘cut off’ the straightforward path through skip-connect, and facilitate the algorithm to explore other operations.
作者采用搜索空间正则来解决这个问题。一方面，作者在skip-connect操作后添加Operations层面的随机Dropout来部分切断skip-connect的连接，迫使算法去探索其他的操作。
However, if we constantly block the path through skip-connect, the algorithm will drop them by assigning low weights to them, which is harmful to the final performance.
然而，持续地阻断这些路径的话会导致在最终生成结构的时候skip-connect操作仍然受到抑制，可能会影响最终性能。
we gradually decay the Dropout rate during the training process in each search stage, thus the straightforward path through skip-connect is blocked at the beginning and treated equally afterward when parameters of other operations are well learned, leaving the algorithm itself to make the decision.
因此，作者在训练的过程中逐渐地衰减Dropout的概率，在训练初期施加较强的Dropout，在训练后期将其衰减到很轻微的程度，使其不影响最终的网络结构参数的学习。
Despite the use of Dropout, we still observe that skip-connect, as a special kind of operation, has a significant impact on recognition accuracy at the evaluation stage.
另一方面，尽管使用了Operations层面的Dropout，作者依然观察到了skip-connect操作对实验性能的强烈影响。
This motivates us to design the second regularization rule, architecture refinement, which simply controls the number of preserved skip-connects, after the final search stage, to be a constant M.
因此，作者提出第二个搜索空间正则方法，即在最终生成的网络结构中，保留固定数量的skip-connect操作。具体的，作者根据最终阶段的结构参数，只保留权值最大的M个skip-connect操作，这一正则方法保证了搜索过程的稳定性。在本文中， M=2 .
We emphasize that the second regularization technique must be applied on top of the first one, otherwise, in the situations without operation-level Dropout, the search process is producing low-qualityarchitectureweights, basedon which we could not build up a powerful architecture even with a fixed number of skip-connects.
需要强调的是，第二种搜索空间正则是建立在第一种搜索空间正则的基础上的。在没有执行第一种正则的情况下，即使执行第二种正则，算法依旧会生成低质量的网络结构。
Experiments
Cell arch in different Search Stage
cifar10
ImageNet
·
Conclusion
we propose a progressive version of differentiable architecture search to bridge the depth gap between search and evaluation scenarios.
The core idea is to gradually increase the depth of candidate architectures during the search process.
- 2Q: computational overhead and instability
Search space approximate and Search space regularize
Our research defends the importance of depth in differentiable architecture search， depth is still the dominant factor in exploring the architecture space.
