AMC: AutoML for Model Compression and Acceleration on Mobile Devices
Citations
6,222 citations
Cites background from "AMC: AutoML for Model Compression a..."
...Model compression (Han et al., 2016; He et al., 2018; Yang et al., 2018) is a common way to reduce model size by trading accuracy for efficiency....
[...]
1,841 citations
1,201 citations
Cites background from "AMC: AutoML for Model Compression a..."
...Besides searching for new architectures, works such as [27, 5] focus on adapting existing models to improve efficiency....
[...]
977 citations
892 citations
Cites background or methods from "AMC: AutoML for Model Compression a..."
...271 Recently, [25] uses a similar pruning technique to Network Slimming [6] to automate the design of net272 work architectures; [26] prune channels using reinforcement learning and automatically compresses 273 the architecture....
[...]
...Recently, Gordon et al. (2018) uses a similar pruning technique to Network Slimming (Liu et al., 2017) to automate the design of network architectures; He et al. (2018) prune channels using reinforcement learning and automatically compresses the architecture....
[...]
...[26] Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, and Song Han....
[...]
References
123,388 citations
55,235 citations
"AMC: AutoML for Model Compression a..." refers methods in this paper
...We also compare our method to the original channel pruning paper (CP in Table 3), in which the sparsity ratios of pruned VGG-16 [45] are carefully tuned by human experts (conv5 are skipped, sparsity ratio for conv4 and remaining layers is 1 : 1.5)....
[...]
...Compelling results have been demonstrated for MobileNet, MobileNet-V2, ResNet and VGG on Cifar and ImageNet....
[...]
...To demonstrate the wide and general applicability, we evaluate our AMC engine on multiple neural networks, including VGG [45], ResNet [21], and MobileNetV1 [23], and we also test the generalization ability of the compressed model from classification to object detection....
[...]
...The results of pruned VGG-16 are presented in Table 3....
[...]
...Apart from VGG-16, we also test AMC on modern efficient neural networks MobileNet [23] and MobileNet-V2 [44]....
[...]
40,257 citations
17,184 citations
15,005 citations