1 Introduction
By dint of automatic feature engineering, deep neural networks (DNNs) have achieved remarkable success in various computer vision tasks, such as image classification
[37, 36, 41, 33, 32, 15, 39], visual generation [34, 35][40, 7, 8, 12] and semantic comprehension [18, 17]. In contrast, neural architecture search (NAS) aims at automatically learning the network architecture to further boost the performance for target tasks [10, 20, 43, 2, 19]. Nevertheless, previous NAS methods in general suffer from huge computation budget, such as 2000 GPU days of reinforcement learning
[43] and 3150 GPU days of evolution [26] with hundreds of GPUs.Current Oneshot NAS methods boost the search efficiency by modeling NAS as a oneshot training process of an overparameterized supernet. As a result, various architectures can be derived from the supernet, and share the same weights. For example, DARTS [21] and its variants [38, 1] parameterize the supernet with an additional categorical distribution for indicating what operations we want to keep. In contrast, recent single path methods adopt a nonparametric architecture modeling, and split the searching into two consecutive stages, i.e., supernet training and architecture sampling. For training supernet, only a single path consisting of a single operation choice is activated and gets optimized by regular gradientbased optimizers. After the supernet is trained well, it is regarded as a performance estimator for all architectures (i.e., paths). Then the optimal architecture can be searched using a holdout validation dataset via random search [16] or (reinforced) evolutionary [11, 4] algorithms under specified hardware constraint (e.g., FLOPs and latency). As only one path is activated for training, the memory cost coheres with that of traditional network training, and scales well on largescale datasets (e.g., ImageNet [27]).
Supernet matters for it serves as a fundamental performance estimator of different architectures (paths). Current methods [16, 11, 4, 3] hold the assumption that the supernet should estimate the (relative) performance accurately for all paths, and thus all paths are treated equally and trained simultaneously. However, the paths contained in the supernet are of fairly huge scale (e.g.,
). Hence it can be harsh for a single supernet to evaluate and give reasonable ranking on such a quantity of paths at the same time. In fact, the ultimate aim of supernet is only to identify a bunch of optimal paths. But the huge search space implies significant variance and variety of paths; there exist many architectures of inferior quality in terms of accuracy performance.
^{1}^{1}1For example, in a same supernet, MobileNetV2 [28] can achieve 72.0% Top1 accuracy on ImageNet dataset while an extreme case of almost all identity operations only has 24.1% [3]. Since the weights of all paths are highly shared, if a weak path is sampled and gets trained, it would disturb the weights of those potentiallygood paths. This disturbance will undermine their eventual performance estimation and affect the searched optimal architecture accordingly. The supernet is thus not supposed to care much on these weak paths and get updated for them. Besides, training on those weak paths actually involves unnecessary update of weights, and slows down the training efficiency more or less.In this paper, we ease the training burden by encouraging a greedy supernet. A greedy supernet is capable of shifting its focus on performance estimation of those potentiallygood paths instead of all paths. Concretely, during the supernet training, we propose a multipath sampling strategy with rejection to filter the weak paths, so the supernet will greedily train those potentiallygood paths. This path filtering can be efficiently implemented via evaluation using a surrogate portion of validation dataset, without harming the computation cost too much. Moreover, we also adopt an exploration and exploitation policy [14, 24] by introducing a candidate pool, which dynamically tracks those potentiallygood paths discovered during training. In this way, the supernet improves its training efficiency by switching its training space from all paths into those potentiallygood ones, and further into candidate pool by sampling from it, as shown in Figure 1.
We implement our proposed method GreedyNAS on the largescale benchmark ImageNet dataset [27], and extensive experimental results indicate our superiority in terms of accuracy performance and supernet training efficiency. For example, with the same search space, our method can achieve higher Top1 accuracy than that of other comparison methods under the same FLOPs or latency level, but reduces approximate 40% of supernet training cost. By searching on a larger space, we can also obtain new stateoftheart architectures.
2 Related Work
Oneshot NAS methods mainly aim to train an overparameterized network (a.k.a supernet) that comprises all architectures (paths), which share the same weights mutually. Then the optimal architecture can be derived or searched from the supernet. There are mainly two categories of oneshot NAS methods [9], which differ in how the architectures are modeled and elaborated as follows.
Parameterized architectures. To use the gradientbased optimizers for direct searching, an realvalued categorical distribution (architecture parameter) is usually introduced in the supernet, and can be thus jointly learned with the supernet weights, such as DARTS [21], FBNet [38] and MdeNAS [42]. When the supernet training is finished, the optimal architecture can be induced by sampling from the categorical distribution. However, it may suffer from the huge GPU memory consumption. ProxylessNAS [1] alleviates this issue by factorizing the searching into multiple binary selection tasks while SinglePathNAS [29] uses superkernels to encode all operation choices. Basically, they are difficult to integrate a hard hardware constraint (e.g., FLOPs and latency) during search but resort to relaxed regularization terms [38, 1].
Sampled singlepath architectures. By directly searching the discrete search space, the supernet is trained by sampling and optimizing a single path. The sampling can be uniform sampling [11, 16] or multipath sampling with fairness [4]
. After the supernet is trained, it is supposed to act as a performance estimator for different paths. And the optimal path can be searched by various searchers, such as random search and evolutionary algorithms
[6]. For example, ScarletNAS [3] employs a multiobjective searcher [22] to consider classification error, FLOPs and model size for better paths. Different to the previous parameterized methods, the hard hardware constraint can be easily integrated in the searchers. Our proposed method GreedyNAS is cast into this category.3 Rethinking path training of supernet
In Singlepath Oneshot NAS, we utilize an overparameterized supernet with parameter to substantialize a search space, which is formulated as a directed acyclic graph (DAG). In the DAG, feature maps act as the nodes, while the operations (or transformations) between feature maps are regarded as edges for connecting sequential nodes. Assume the supernet has layers, and each layer is allocated with operation choices , which can be basic convolution, pooling, identity or different types of building blocks, such as MobileNetV2 block [28] and ShuffleNetV2 block [23]. Then each architecture (i.e., path) denoted as can be represented by a tuple of size , i.e., where . As a result, the search space is discrete, and there will be (e.g., ) architectures in total, namely, .
Training supernet matters since it is expected to serve as a fundamental performance estimator. Due to the consideration of memory consumption, singlepath NAS methods implement training by sampling a single path from , then the sampled paths are all optimized on the training data . It can be formulated as minimizing an expected loss over the space , i.e.,
(1) 
where refers to the parameter of path , and is a discrete sampling distribution over .
After the supernet is trained well, we can evaluate the quality of each path by calculating its (Top1) accuracy (ACC) on the validation dataset , and the optimal path corresponds to the maximum ACC, i.e.,
(2) 
where w.r.t. path in the trained supernet .
3.1 Reshaping sampling distribution
Current methods assume that the supernet should provide a reasonable ranking over all architectures in . Thus all paths are treated equally, and optimized simultaneously [16, 11, 4, 3]. Then the sampling distribution
amounts to a uniform distribution
over , i.e.,(3) 
where is an indicator function. However, as previously discussed, it is a demanding requirement for the supernet to rank accurately for all paths at the same time. In the huge search space , there might be some paths of inferior quality. Since the weights are highly shared in the same supernet, training on these weak paths does have negative influence on the evaluation of those potentiallygood paths. To alleviate this disturbance, an intuitive idea is to block the training of these weak paths.
For simplifying the analysis, we assume the search space can be partitioned into two subsets and by an Oracle good but unknown supernet , where
(4) 
and indicates the potentiallygood paths while is for weak paths, i.e.,
(5) 
holds for all on validation dataset . Then to screen the weak paths and ease the burden of the supernet training, we can just sample from the potentiallygood paths instead of all paths . The sampling distribution is equivalently reshaped by truncation on , i.e., and
(6) 
In this way, the supernet is expected to thoroughly get trained on the potentiallygood paths and thus give decent performance ranking. Besides, since the valid search space has been shrunken from into , the training efficiency of supernet is improved accordingly.
3.2 Greedy path filtering
Nevertheless, in the supernet training the Oracle supernet is unknown, thus we can not sample paths according to Eq.(6) since it relies on . In this paper, we propose to use greedy strategy and during training, current supernet is progressively regarded as a proxy of the Oracle . Thus during the supernet training, we greedily sample paths according to the reshaped sampling distribution given by current , namely, . The sampled paths will get optimized, then the supernet is get updated and evolves to a decent performance estimator over .
However, a natural question arises: even given a supernet , how can we sample from the shaped distribution ? In other words, how can we accurately identify whether a path is from or ? Note that the partition of is determined by traversing all paths in as Eq.(5), which is not affordable in computation. Since we can not accurately know whether a single path is good or weak, to solve this issue, we propose a multipath sampling strategy with rejection.
Suppose we uniformly sample a path from , then it amounts to be sampled from with probability , and sampled from with probability
. In this way, if we sample multiple paths independently at a time, we have the following results based on binomial distribution.
Theorem 1.
From Theorem 1, we can see by sampling paths, the probability that at least paths are from is very high when the proportion of potentiallygood paths is medially large or is medially small (see Figure 2). For example, if we conservatively assume 60% paths have the potential to be good (i.e., ), we will have 83.38% confidence to say at least 5 out of 10 paths are sampled from . In this way, based on the definition of Eq.(4) and Eq.(5), we just rank the sampled paths using validation data , keep the Top paths and reject the remaining paths.
However, ranking paths involves calculation of ACC over all validation dataset as Eq.(5), which is also computationally intensive during the supernet training.^{2}^{2}2For example, the size of on ImageNet dataset is 50k. In fact, in our multipath sampling, what we care about is the obtained ranking; we empirically find that it suffices to rank based on the loss (e.g., cross entropy loss for classification) over a surrogate subset of (e.g., 1k images on ImageNet dataset), denoted as . The consistency between this rank and that given by ACC on all is fairly significant. More details and analysis refer to the ablation studies in Section 5.3.1. Then the sampling works as Algorithm 1.
As a result, path filtering can be efficiently implemented for it can run in a simple feedforward mode (e.g., eval()
mode in Pytorch) on a small portion of validation data. In this sense, we block the weak paths greedily during the supernet training. And the validation data
acts as a rough filter to prevent the training of those lowquality or even harmful paths, so that the supernet can get sufficient training on those potentiallygood ones.4 Proposed Approach: GreedyNAS
In this section, we formally illustrated our proposed NAS method (a.k.a. GreedyNAS) based on a greedy supernet. Our GreedyNAS is composed with three procedures, i.e., supernet training, searching paths and retraining the searched optimal path. The last retraining corresponds to conventional training a given network. We mainly elaborate the first two as follows.
4.1 Greedy training of supernet
As previously discussed, we propose to maintain a greedy supernet during its training. By doing this, we gradually approximate the sampling by keeping the Top paths and filtering the bottom paths by evaluating using . Then those weak paths are prevented from getting trained, which allows the supernet to focus more on those potentiallygood paths and switch its training space from into .
4.1.1 Training with exploration and exploitation
After the greedy path filtering, we have actually identified some potentiallygood paths, which amount to some empiricallygood ones given by current supernet. Then to further improve the training efficiency, inspired by the Monte Carlo tree search [14] and deep Qlearning (DQN) [24], we propose to train the supernet with an exploration and exploitation (EE) strategy by reusing these paths.
Concretely, we introduce a candidate pool to store the potentiallygood paths discovered during training. Each path is represented as a tuple of operation choices. Besides, each is also allocated with an evaluation loss . The candidate pool is thus formulated as a fixedsize ordered queue with priority . With more potentiallygood paths involved, the candidate pool can be maintained by a minheap structure in real time.
As a result, we can conservatively implement local search by sampling from the candidate pool since it consists of a smaller number (but promising) of paths. However, this greedy exploitation brings in the risks of losing path diversity for the training. In this way, we also favor a global search with the hope of probing other promising paths that are yet to be sampled and get trained, which can be easily fulfilled by uniform sampling from . For achieve a balanced tradeoff of exploration and exploitation, we adopt a typical sampling policy, i.e., implementing uniform sampling both from and pool (line 4 of Algorithm 1),
(8) 
where indicates the probability of sampling from the pool . Note that candidate pool runs through the training process of supernet; however, it might be not reliable at first since the priority is calculated based on a much lesstrained supernet. In this case, we propose to actively anneal the pool sampling probability from 0 to a predefined level. In our experiment, we find will be a good option.
Training with exploration and exploitation encourages the supernet to refine the alreadyfound good paths as well as probing new territory for more better paths. Besides, it actually also contributes to our greedy path filtering by improving our filtering confidence. Basically, the collected potentiallygood paths can be regarded as a subset of , then sampling from amounts to increasing the probability of Theorem 1 into
(9) 
which refers to the proportion of potentiallygood paths. For example, assume we evenly sample from or (), then the probability of sampling at least 5 good paths out of 10 paths will rise from to according to Theorem 1. Comparing reducing to increase the sampling confidence, sampling with is almost costneglectable since we only need to maintain a minheap. The supernet thus gradually shifts its training from more to , and the training efficiency will be further improved accordingly.
4.1.2 Stopping principle via candidate pool
Different to conventional networks, a supernet serves as a performance estimator and it is difficult to judge when it is trained well. Current singlepath NAS methods control its training by manually specifying an maximum epoch number. In our GreedyNAS, however, we propose an adaptive stopping principle based on the candidate pool.
Candidate pool indicates a bunch of best empirical paths, and it is updated dynamically during the supernet training. In consequence, if a supernet is trained well, the pool should tend to be steady. This steadiness can be measured by the update frequency of candidate pool , i.e.,
(10) 
where refers to the old in previous iterations. Thus smaller implies that fewer new paths are involved in the pool within iterations, and is more steady. Given a certain tolerance level ^{3}^{3}3 suffices in our experiment., when the update frequency is less than , we believe the supernet has been trained enough, and its training can be stopped accordingly.
Methods  performance  supernet training efficiency  
Top1 (%)  FLOPs  latency  Params  #optimization  #evaluation  corrected #optimization  
ProxylessR (mobile) [1]  74.60  320M  79 ms  4.0M       
Random Search  74.07  321M  69 ms  3.6M  1.23M120    147.6M 
Uniform Sampling [11]  74.50  326M  72 ms  3.8M  1.23M120    147.6M 
FairNASC [4]  74.69  321M  75 ms  4.4M  1.23M150    184.5M 
Random SearchE  73.88  320M  91 ms  3.7M  1.23M73    89.8M 
Uniform Sampling [11]E  74.17  320M  94 ms  3.6M  1.23M73    89.8M 
GreedyNAS (FLOPsM)  74.85  320M  89 ms  3.8M  1.23M46  2.40M46  89.7M 
GreedyNAS (latencyms)  74.93  324M  78 ms  4.1M  1.23M46  2.40M46  89.7M 
4.2 Searching with candidate pool
After the supernet is trained, we can use supernet to evaluate the quality (ACC) of each path on validation dataset , and search the optimal path as Eq.(2). However, enumerating all paths in is prohibitively computationintensive. One remedy is by dint of evolutionary algorithms [11] or reinforced version (e.g., MoreMNAS [5]), which takes the supernet as an offtheshelf evaluator. In our paper, we adopt the multiobjective NSGAII [6] algorithm for searching, where the hardware constraint can be easily integrated in the evolution process. If a path violates the predefined hardware constraint (e.g., under 330 FLOPs), we just ditch it for good.
Besides, evolutionary algorithms need to initialize population with size before implementing iterative mutation and crossover. Current methods usually random sample paths under the constraint as initial population. In contrast, our method makes the initialization with the help of candidate pool , and select its Top paths instead. As Figure 3 shows, searching with candidate pool can boost the evolutionary performance for supplying a good initial population. The ACC of searched paths using candidate pool is on average higher than that using random initialization. More details of our searching algorithm refer to the supplementary materials.
5 Experimental Results
5.1 Configuration and settings
Dataset. We conduct the architecture search on the challenging ImageNet dataset [27]. As [1], we randomly sample 50,000 images (50 images per class) from training dateset as the validation dataset (K), and the rest of training images are used for training. Moreover, we use the original validation dataset as the test dataset to report the accuracy performance.
Search space. Following [1, 4], we adopt the same macrostructure of supernet for fair comparison as shown in Table 5 (see supplementary materials). Moreover, we use MobileNetV2 inverted bottleneck [28] as the basic building block. For each building block, the convolutional kernel size is within and expansion ratio is selected in . An identity block is also attached for flexible depth search. As a result, with 21 building blocks, the search space is of size . In addition, we also implement searching on a larger space by augmenting each building block with an squeezeandexcitation (SE) option. The size of the larger search space is thus .
Supernet training. For training the supernet, Algorithm 1 is adopted to sample paths and filter paths. We randomly sample images (1 image per class) from the validation dataset for evaluating paths in Algorithm 1
. For training each path, we use a stochastic gradient descent (SGD) optimizer with momentum 0.9 and Nesterov acceleration. The learning rate is decayed with cosine annealing strategy from initial value
. The batch size is . As for candidate pool, we empirically find 1000 is a good option for pool size , which approximates the amount of paths involved in one epoch. The candidate sampling probability is linearly increased from to . Instead of specifying an epoch number [11, 4], we use the proposed principle to stop the supernet training with tolerance .Methods 









SCARLETC [4]  75.6  92.6  280  67  6.0  single path  10  12  
MobileNetV2 1.0 [28]  72.0  91.0  300  38  3.4        
MnasNetA1 [30]  75.2  92.5  312  55  3.9  single path + RL    
GreedyNASC  76.2  92.5  284  70  4.7  single path  7  
ProxylessR (mobile) [1]  74.6  92.2  320  79  4.0  two paths    
FairNASC [4]  74.7  92.1  321  75  4.4  single path  10  2  
Uniform Sampling [11]  74.7    328      single path  12  
SCARLETB [4]  76.3  93.0  329  104  6.5  single path  10  12  
GreedyNASB  76.8  93.0  324  110  5.2  single path  7  
SCARLETA [4]  76.9  93.4  365  118  6.7  single path  10  12  
EfficientNetB0 [31]  76.3  93.2  390  82  5.3  single path      
DARTS [21]  73.3  91.3  574    4.7  a whole supernet    
GreedyNASA  77.1  93.3  366  77  6.5  single path  7 
Evolutionary searching. For searching with NSGAII [6] algorithm, we set the population size as and the number of generations as . The population is initialized by the candidate pool while other comparison methods use random initialization. During searching, we use constraint of FLOPs or latency. All our experiments use Qualcomm® Snapdragon™ 855 mobile hardware development kit (HDK) to measure the latency.
Retraining. To train the obtained architecture, we use the same strategy as [1]
for search space without SE. As for the augmented search space, we adopt a RMSProp optimizer with
momentum as Mnasnet [30]. Learning rate is increased from to in the first epochs with batch size , and then decays every epochs. Besides, exponential moving average is also adopted with decay .5.2 Performance comparison with stateoftheart methods
Searching on same search space. For fair comparison, we first benchmark our GreedyNAS to the same search space as [1] to evaluate our superiority to other Singlepath Oneshot NAS methods. We also cover a baseline method Random Search, which shares the same supernet training strategy with Uniform Sampling [11]; but during search, instead of using evolutionary algorithms it randomly samples 1000 paths, and retrains the rank1 path according to Top1 ACC on the supernet. As Table 1 shows, when searching with similar 320 FLOPs, our GreedyNAS achieves the highest Top1 ACC. We further align our searched constraint to latency of 80 ms. Table 1 indicates that with similar latency, GreedyNAS is still consistently superior to other comparison methods. For example, GreedyNAS can search an architecture with 74.93% Top1 ACC, enjoying a 0.43% improvement over uniform sampling, which in a way illustrates the superiority of our greedy supernet to a uniform supernet.
Besides advantages on the classification performance of searched models, we also evaluate our superiority in terms of supernet training efficiency. Since the main differences of our GreedyNAS and other Singlepath Oneshot NAS methods lie in the supernet training, we report in Table 1 the supernet training cost. To eliminate the efficiency gap due to different implementation tools (e.g., GPU types, dataloader wrappers), we calculated the accumulated number of images involved in a whole gradientbased optimization step, i.e., #optimization in Table 1. Our GreedyNAS has an additional evaluation process during training, thus we also report the accumulated number of images for forward evaluation, i.e., #evaluation. For overall efficiency comparison, we empirically find the cost of a whole optimization step is approximately 3.33 times larger than that of a forward evaluation. The corresponding corrected #optimization is covered accordingly.
From Table 1, we can see that the training cost of our GreedyNAS is much smaller than that of other comparison methods, which indicates GreedyNAS enjoys significant efficiency in supernet training since it greedily shrinks its training space into those potentiallygood paths. Besides, we also implement Random Search and Uniform Sampling using same training cost of GreedyNAS, denoted as Random SearchE and Uniform SamplingE, respectively. The results show that with decreased iterations of supernet training, the searched architectures are inferior to those of larger iterations. In contrast, our method can achieve higher accuracy by a large margin (almost 1%). This implies that GreedyNAS is capable of learning a decent supernet with much less iterations.
Searching on augmented search space. To comprehensively illustrate our superiority to various stateoftheart NAS methods, we implement searching by augmenting the current space with an SE option. Moreover, we search the architectures under different FLOPs constraint. But we also report the corresponding latency and parameter capacity to comprehensively analyze the statistics of searched models. As Table 2 shows, our GreedyNAS achieves new stateoftheart performance with respect to different FLOPs and latency levels. For example, with similar FLOPs and latency, GreedyNASC has higher Top1 ACC than the competing SCARLET method by a margin of 0.6%. Our searched models are visualized in Figure 6 (see supplementary materials). It shows smaller network (GreedyNASC) tends to select more identity blocks to reduce the FLOPs while larger networks will exploit more SE modules (GreedyNASA&B) to further improve the accuracy performance. Moreover, GreedyNASA adopts more 33 kernels to have smaller latency since 33 kernels are optimized more maturely in mobile inference framework. We also report our real training cost in Table 2 based on Tesla V100 GPU. It shows that GreedyNAS can significantly reduce the training time compared to other NAS methods, which empirically validates the efficiency of GreedyNAS.
5.3 Ablation studies
5.3.1 Effect of evaluation in path filtering
To filter the weak paths, GreedyNAS evaluates each path by a small portion (1000) of validation images as a surrogate for the whole validation dataset (50K images). We first investigate whether this approximation suffices in our experiment. By random sampling 1000 paths from supernet, we examine the correlation of two path orderings, which are generated by ranking the evaluation results using 1000 and 50K validation images, respectively. In Table 3, we report the widelyused Spearman rho [25] and Kendall tau [13] rank correlation coefficient, which are in the range and larger values mean stronger correlation. We also cover three types of supernets, i.e., randomly initialized, trained by uniform sampling and our greedy sampling.
From Table 3, we can see that our greedy supernet achieves fairly high rank correlation coefficient (0.997 and 0.961), which indicates that the ranking of greedy supernet using 1000 validation images is significantly consistent with that of all validation dataset. Moreover, supernet trained with uniform sampling has smaller correlation coefficient, even with different evaluation images (see left Figure 4). This implies in a sense that our greedy supernet is more discriminative since it can use less images to identify whether a path is good or weak. Nevertheless, as left Figure 4 shows, too few evaluation images might have weak correlation while too many evaluation images mean greater evaluation cost. But 1000 evaluation images enjoy a balanced tradeoff between the rank correlation and evaluation cost. Note that we also report the results w.r.t. ranking using the ACC of 1000 images, which is smaller than that using loss. This results from that during training the value of loss might be more informative than that of ACC.
As for the random supernet, the correlation coefficient is fairly small (0.155 and 0.113). This makes sense since the ranking is based on the classification performance; however, a random supernet fails to learn sufficient domainrelated information but gives disordered ranking of paths. This smaller correlation coefficient implies that it might be not sensible to implement greedy sampling from a random supernet since the ranking evaluated by 1000 validation images will be rather noisy. In this way, we record the trend of rank correlation coefficients with uniform sampling in right Figure 4. It shows that with more iterations, the correlation coefficients increase and at 10K iteration, they tend to be steady at a high level (e.g., 0.81 for Kendall Tau). As a result, in our GreedyNAS we propose to have a warmup stage by uniform sampling for 10K iterations, so that we can safely use 1000 validation images to evaluate paths.
Spearman rho  Kendall tau  

random  uniform(ACC)  greedy  random  uniform(ACC)  greedy 
0.155  0.968(0.869)  0.997  0.113  0.851(0.699)  0.961 
5.3.2 Effect of path filtering and candidate pool
To study the effect of our proposed path filtering and the candidate pool, we implement experiments on the search space without SE. In our GreedyNAS, path filtering is to block the training of weak paths. In contrast, the use of candidate pool is mainly threefold as shown in Table 4. First, we can sample from it as the exploitation process; second, we can initialize the evolutionary searching with the pool for better paths; third, we can use it to adaptively stop the supernet training. Then we control each factor and obtain 6 variants of GreedyNAS as well as 6 corresponding searched architectures Net1Net6. For fair comparison, we search all nets under 330M FLOPs. Besides, if the candidate pool is not used for stopping training, we specify a maximum epoch 60 as [3].
As Table 4 shows, comparing with the baseline Net1 (Net3), Net2 (Net6) achieves 0.28% (0.41%) better Top1 ACC, which indicates that path filtering does contribute to the supernet training, and thus improves the searching results. By involving the candidate pool, Net6 can increase its accuracy from 74.59% (Net2) to 74.89%. In specific, initialization with candidate pool in evolutionary algorithms enables to have a 0.18% gain on Top1 ACC since it helps to search paths with higher ACC on supernet (also see Figure 3). Note that stopping by candidate pool usually saves training cost; however, full training with candidate pool (Net5) seems to drop the accuracy a bit (0.05% w.r.t. Net6). It might result from that extreme greedy exploitation on the candidate pool harms the supernet training instead. Then the stopping in a sense brings benefits for a more balanced tradeoff between exploration and exploitation.
path filtering  candidate pool  Top1 (%)  
sampling  evolutionary  training  
(exploitation)  initialization  stopping  
Net1          74.31 
Net2  ✓        74.59 
Net3    ✓  ✓  ✓  74.48 
Net4  ✓  ✓    ✓  74.71 
Net5  ✓  ✓  ✓    74.84 
Net6  ✓  ✓  ✓  ✓  74.89 
6 Conclusion
Training a supernet is a key issue for Singlepath Oneshot NAS methods. In stead of treating all paths equally, we propose to greedily focus on training those potentiallygood ones. This greedy path filtering can be efficiently implemented by our proposed multipath sampling strategy with rejection. Besides, we also adopt an exploration and exploitation policy and introduce a candidate pool to further boost the supernet training efficiency. Our proposed method GreedyNAS shows significant superiority in terms of both accuracy performance and training efficiency.
References
 [1] (2018) Proxylessnas: direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332. Cited by: §1, §2, Table 1, §5.1, §5.1, §5.1, §5.2, Table 2.
 [2] (2019) Detnas: neural architecture search on object detection. arXiv preprint arXiv:1903.10979. Cited by: §1.
 [3] (2019) Scarletnas: bridging the gap between scalability and fairness in neural architecture search. arXiv preprint arXiv:1908.06022. Cited by: §1, §2, §3.1, §5.3.2, footnote 1.
 [4] (2019) Fairnas: rethinking evaluation fairness of weight sharing neural architecture search. arXiv preprint arXiv:1907.01845. Cited by: §1, §1, §2, §3.1, Table 1, §5.1, §5.1, Table 2.
 [5] (2019) Multiobjective reinforced evolution in mobile neural architecture search. arXiv preprint arXiv:1901.01074. Cited by: §4.2.

[6]
(2002)
A fast and elitist multiobjective genetic algorithm: nsgaii
.IEEE transactions on evolutionary computation
6 (2), pp. 182–197. Cited by: Appendix A, §2, §4.2, §5.1.  [7] (2019) Unsupervised semanticpreserving adversarial hashing for image search. IEEE Transactions on Image Processing 28 (8), pp. 4032–4044. Cited by: §1.
 [8] (2019) Twostream deep hashing with classspecific centers for supervised image search. IEEE Transactions on Neural Networks and Learning Systems. Cited by: §1.

[9]
(2019)
Neural architecture search: a survey..
Journal of Machine Learning Research
20 (55), pp. 1–21. Cited by: §2. 
[10]
(2019)
Irlas: inverse reinforcement learning for architecture search.
In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
, pp. 9021–9029. Cited by: §1.  [11] (2019) Single path oneshot neural architecture search with uniform sampling. arXiv preprint arXiv:1904.00420. Cited by: §1, §1, §2, §3.1, §4.2, Table 1, §5.1, §5.2, Table 2.

[12]
(2018)
Attributeaware attention model for finegrained representation learning
. In Proceedings of the 26th ACM international conference on Multimedia, pp. 2040–2048. Cited by: §1.  [13] (1938) A new measure of rank correlation. Biometrika 30 (1/2), pp. 81–93. Cited by: §B.3, §5.3.1.
 [14] (2006) Bandit based montecarlo planning. In European conference on machine learning, pp. 282–293. Cited by: §1, §4.1.1.

[15]
(2020)
Learning student networks with few data.
In
Proceedings of the AAAI Conference on Artificial Intelligence
, Cited by: §1.  [16] (2019) Random search and reproducibility for neural architecture search. arXiv preprint arXiv:1902.07638. Cited by: §1, §1, §2, §3.1.
 [17] (2019) A realtime crossmodality correlation filtering method for referring expression comprehension. arXiv preprint arXiv:1909.07072. Cited by: §1.
 [18] (2019) PPDM: parallel point detection and matching for realtime humanobject interaction detection. arXiv preprint arXiv:1912.12898. Cited by: §1.
 [19] (2019) Graphguided architecture search for realtime semantic segmentation. arXiv preprint arXiv:1909.06793. Cited by: §1.
 [20] (2018) Progressive neural architecture search. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 19–34. Cited by: §1.
 [21] (2019) DARTS: differentiable architecture search. In ICLR (Poster), Cited by: §1, §2, Table 2.
 [22] (2018) NSGAnet: a multiobjective genetic algorithm for neural architecture search. arXiv preprint arXiv:1810.03522. Cited by: §2.
 [23] (2018) Shufflenet v2: practical guidelines for efficient cnn architecture design. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 116–131. Cited by: §3.
 [24] (2013) Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602. Cited by: §1, §4.1.1.
 [25] (2004) S pearman rank correlation coefficient. Encyclopedia of statistical sciences. Cited by: §B.3, §5.3.1.

[26]
(2019)
Regularized evolution for image classifier architecture search
. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33, pp. 4780–4789. Cited by: §1.  [27] (2015) ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) 115 (3), pp. 211–252. External Links: Document Cited by: §1, §1, §5.1.
 [28] (2018) Mobilenetv2: inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 4510–4520. Cited by: §3, §5.1, Table 2, footnote 1.
 [29] (2019) Singlepath nas: designing hardwareefficient convnets in less than 4 hours. arXiv preprint arXiv:1904.02877. Cited by: §2.
 [30] (2019) Mnasnet: platformaware neural architecture search for mobile. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2820–2828. Cited by: §5.1, Table 2.

[31]
(2019)
EfficientNet: rethinking model scaling for convolutional neural networks
. In International Conference on Machine Learning, pp. 6105–6114. Cited by: Table 2.  [32] (2020) Reborn filters: pruning convolutional neural networks with limited data. In Proceedings of the AAAI Conference on Artificial Intelligence, Cited by: §1.
 [33] (2019) Bringing giant neural networks down to earth with unlabeled data. arXiv preprint arXiv:1907.06065. Cited by: §1.
 [34] (2018) Perceptual adversarial networks for imagetoimage transformation. IEEE Transactions on Image Processing 27 (8), pp. 4066–4079. Cited by: §1.
 [35] (2019) Evolutionary generative adversarial networks. IEEE Transactions on Evolutionary Computation 23 (6), pp. 921–934. Cited by: §1.

[36]
(2018)
The devil of face recognition is in the noise
. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 765–780. Cited by: §1.  [37] (2017) Residual attention network for image classification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3156–3164. Cited by: §1.
 [38] (2019) Fbnet: hardwareaware efficient convnet design via differentiable neural architecture search. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 10734–10742. Cited by: §1, §2.
 [39] (2019) Deep comprehensive correlation mining for image clustering. In Proceedings of the IEEE International Conference on Computer Vision, pp. 8150–8159. Cited by: §1.
 [40] (2018) Shared predictive crossmodal deep quantization. IEEE transactions on neural networks and learning systems 29 (11), pp. 5292–5303. Cited by: §1.
 [41] (2017) Learning from multiple teacher networks. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1285–1294. Cited by: §1.
 [42] (2019) Multinomial distribution learning for effective neural architecture search. In Proceedings of the IEEE International Conference on Computer Vision, Cited by: §2.
 [43] (2017) Neural architecture search with reinforcement learning. In ICLR, Cited by: §1.
Appendix A Details of evolutionary searching in Section 4.2
Appendix B More Experimental Results
b.1 Details of (augmented) search space
The macrostructure of supernet is presented in Table 5, where each operation choice for Choice Block is attached in Table 6.
input  block  channels  repeat  stride 

conv  32  1  2  
MB1_K3  16  1  1  
Choice Block  32  4  2  
Choice Block  40  4  2  
Choice Block  80  4  2  
Choice Block  96  4  1  
Choice Block  192  4  2  
Choice Block  320  1  1  
conv  1280  1  1  
global avgpool    1    
FC  1000  1   
block type  expansion ratio  kernel  SE 

MB1_K3  1  3  no 
ID       
MB3_K3  3  3  no 
MB3_K5  3  5  no 
MB3_K7  3  7  no 
MB6_K3  6  3  no 
MB6_K5  6  5  no 
MB6_K7  6  7  no 
MB3_K3_SE  3  3  yes 
MB3_K5_SE  3  5  yes 
MB3_K7_SE  3  7  yes 
MB6_K3_SE  6  3  yes 
MB6_K5_SE  6  5  yes 
MB6_K7_SE  6  7  yes 
b.2 Calculating corrected #optimization in Table 1
In our GreedyNAS, when equipped with the stopping principle of candidate pool, the supernet training is stopped at approximately 46th epoch. Thus the accumulated number of examples calculated for a whole optimization step is equal to
where 1.23M refers to the quantity of training dataset. As for the path filtering, we evaluate 10 paths based on 1000 validation images, and select 5 paths for training, whose batch size is 1024. In this way, the number of images for evaluation amounts to
Given our empirical findings that the cost of a whole optimization step is approximately 3.33 times larger than that of a forward evaluation, the corrected #optimization is thus
b.3 Details of rank correlation coefficient
In ablation study 5.3.1, we use two Spearman rho [25] and Kendall tau [13] rank correlation coefficient to evaluate the correlation of two path orderings, which are generated by ranking the evaluation results using 1000 and 50K validation images, denoted as and , respectively. Basically, we aim to calculate the correlation of and .
For Spearman rho rank correlation coefficient, it is simply the Pearson correlation coefficient between random variable
and , if we regard andas two observation vectors of random variable
and , i.e.,where is the covariance of two variables, and
is the standard deviations of
. Based on observation vectors, it can be more efficiently calculated aswhere in our experiment.
For Kendall tau rank correlation coefficient, it focuses on the pairwise ranking performance. For any pair and , it is said to be concordant if and hold simultaneously, or also for and . Otherwise, it will be called disconcordant. Then the coefficient is calculated as
where refers to the total number of pairs. In this way, if two rankings and are exactly the same, will be 1 while if the two are completely different (i.e., one ranking is the reverse of the other), will be 1. According to the definition, it can also be calculated in a closedform as
where is the sign function.
b.4 More ablation studies
b.4.1 Performance of trained supernet
To further investigate the performance of the trained supernet, we implement two different searching methods (random search and evolutionary search) on various trained supernet, i.e., greedy supernet, uniform supernet (full training) and uniform supernetE (same training cost with GreedyNAS). The results can be regarded as supplementary for Table 1.
supernet  searcher  Top1 (%)  FLOPs 

uniform  random  74.07  321M 
uniformE  random  73.88  320M 
greedy  random  74.22  321M 
uniform  evolutionary  74.50  326M 
uniformE  evolutionary  74.17  320M 
greedy  evolutionary  74.85  320M 
From Table 7, we can see that a greedy supernet improves consistently the classification accuracy in terms of different searchers. This validates the superiority of our greedy supernet since it helps searchers to probe better architectures. Moreover, to comprehensively investigate the effect of supernets, we implement systematic sampling ^{4}^{4}4https://en.wikipedia.org/wiki/Systematic_sampling to sample paths from paths, which are discovered by the evolutionary algorithms and ranked according to the accuracy on supernet. Then we retrain these paths from scratch, and report their distribution histogram in Figure 5.
As shown in Figure 5, we can see that on average, paths searched with our greedy supernet have higher retraining Top1 accuracy than that with uniform supernet. This implies that our greedy supernet serves as a better performance estimator, so that those good paths can be eventually identified and searched.
b.5 Visualization of searched architectures
We visualize the searched architectures by our GreedyNAS method in Figure 6.
Comments
There are no comments yet.