|

Dual Path Networks

This repository contains the code and trained models of:

Yunpeng Chen, Jianan Li, Huaxin Xiao, Xiaojie Jin, Shuicheng Yan, Jiashi Feng. "Dual Path Networks" (NIPS17).

• DPNs helped us won the 1st place in Object Localization Task in ILSVRC 2017, with all competition tasks within Top 3. (Team: NUS-Qihoo_DPNs)

Implementation

DPNs are implemented by MXNet @92053bd.

Augmentation

| Method | Settings | | :------------- | :--------: | | Random Mirror | True | | Random Crop | 8% - 100% | | Aspect Ratio | 3/4 - 4/3 | | Random HSL | [20,40,50] |

Note: We did not use PCA Lighting and any other advanced augmentation methods. Input images are resized by bicubic interpolation.

Normalization

The augmented input images are substrated by mean RGB = [ 124, 117, 104 ], and then multiplied by 0.0167.

Mean-Max Pooling

Here, we introduce a new testing technique by using Mean-Max Pooling which can further improve the performance of a well trained CNN in the testing phase without the need of any training/fine-tuining process. This testing technique is designed for the case when the testing images is larger than training crops. The idea is to first convert a trained CNN model into a convolutional network and then insert the following Mean-Max Pooling layer (a.k.a. Max-Avg Pooling), i.e. 0.5 * (global average pooling + global max pooling), just before the final softmax layer.

Based on our observations, Mean-Max Pooling consistently boost the testing accuracy. We adopted this testing strategy in both LSVRC16 and LSVRC17.

Results

ImageNet-1k

Single Model, Single Crop Validation Error:

Model	Size	GFLOPs	224x224		320x320		320x320 ( with mean-max pooling )
			Top 1	Top 5	Top 1	Top 5	Top 1	Top 5
DPN-68	49 MB	2.5	23.57	6.93	22.15	5.90	21.51	5.52
DPN-92	145 MB	6.5	20.73	5.37	19.34	4.66	19.04	4.53
DPN-98	236 MB	11.7	20.15	5.15	18.94	4.44	18.72	4.40
DPN-131	304 MB	16.0	19.93	5.12	18.62	4.23	18.55	4.16

ImageNet-1k (Pretrained on ImageNet-5k)

Single Model, Single Crop Validation Error:

Model	Size	GFLOPs	224x224		320x320		320x320 ( with mean-max pooling )
			Top 1	Top 5	Top 1	Top 5	Top 1	Top 5
DPN-68	49 MB	2.5	22.45	6.09	20.92	5.26	20.62	5.07
DPN-92	145 MB	6.5	19.98	5.06	19.00	4.37	18.79	4.19
DPN-107	333 MB	18.3	19.75	4.94	18.34	4.19	18.15	4.03

Note: DPN-107 is not well trained.

ImageNet-5k

Single Model, Single Crop Validation Accuracy:

Model	Size	GFLOPs	224x224		320x320		320x320 ( with mean-max pooling )
			Top 1	Top 5	Top 1	Top 5	Top 1	Top 5
DPN-68	61 MB	2.5	61.27	85.46	61.54	85.99	62.35	86.20
DPN-92	184 MB	6.5	67.31	89.49	66.84	89.38	67.42	89.76

Note: The higher model complexity comes from the final classifier. Models trained on ImageNet-5k learn much richer feature representation than models trained on ImageNet-1k.

Efficiency (Training)

The training speed is tested based on MXNet @92053bd.

Multiple Nodes (Without specific code optimization):

Model | CUDA
/cuDNN | #Node | GPU Card
(per node) | Batch Size
(per GPU) | kvstore | GPU Mem
(per GPU) | Training Speed*
(per node) :-------|:------------:|:----:|:---------------------:|:----------------------:|:---------:|:---------:|:-----------: DPN-68 | 8.0 / 5.1 | 10 | 4 x K80 (Tesla) | 64 |dist_sync| 9337 MiB | 284 img/sec DPN-92 | 8.0 / 5.1 | 10 | 4 x K80 (Tesla) | 32 |dist_sync| 8017 MiB | 133 img/sec DPN-98 | 8.0 / 5.1 | 10 | 4 x K80 (Tesla) | 32 |dist_sync| 11128 MiB | 85 img/sec DPN-131 | 8.0 / 5.1 | 10 | 4 x K80 (Tesla) | 24 |dist_sync| 11448 MiB | 60 img/sec DPN-107 | 8.0 / 5.1 | 10 | 4 x K80 (Tesla) | 24 |dist_sync| 12086 MiB | 55 img/sec

*This is the actual training speed, which includes data augmentation, forward, backward, parameter update, network communication, etc. MXNet is awesome, we observed a linear speedup as has been shown in link

Trained Models

Model | Size | Dataset | MXNet Model :--------|:------:|:---------:|:-----------------------------------: DPN-68 | 49 MB |ImageNet-1k|GoogleDrive DPN-68* | 49 MB |ImageNet-1k|GoogleDrive DPN-68 | 61 MB |ImageNet-5k|GoogleDrive DPN-92 | 145 MB |ImageNet-1k|GoogleDrive DPN-92 | 138 MB |Places365-Standard|GoogleDrive DPN-92* | 145 MB |ImageNet-1k|GoogleDrive DPN-92 | 184 MB |ImageNet-5k|GoogleDrive DPN-98 | 236 MB |ImageNet-1k|GoogleDrive DPN-131 | 304 MB |ImageNet-1k|GoogleDrive DPN-107*| 333 MB |ImageNet-1k|GoogleDrive

*Pretrained on ImageNet-5k and then fine-tuned on ImageNet-1k.

Third-party Implementations

• Caffe Implementation with trained models by soeaver

• Chainer Implementation by oyam

• Keras Implementation by titu1994

• MXNet Implementation by miraclewkf

• PyTorch Implementation by oyam

• PyTorch Implementation with trained models by rwightman

Other Resources

ImageNet-1k Trainig/Validation List: - Download link: GoogleDrive

ImageNet-1k category name mapping table: - Download link: GoogleDrive

ImageNet-5k Raw Images: - The ImageNet-5k is a subset of ImageNet10K provided by this paper. - Please download the ImageNet10K and then extract the ImageNet-5k by the list below.

ImageNet-5k Trainig/Validation List: - It contains about 5k leaf categories from ImageNet10K. There is no category overlapping between our provided ImageNet-5k and the official ImageNet-1k. - Download link: GoogleDrive

Places365-Standard Validation List & Matlab code for 10 crops testing: - Download link: GoogleDrive

Citation

If you use DPN in your research, please cite the paper:

@article{Chen2017,
  title={Dual Path Networks},
  author={Yunpeng Chen, Jianan Li, Huaxin Xiao, Xiaojie Jin, Shuicheng Yan, Jiashi Feng},
  journal={arXiv preprint arXiv:1707.01629},
  year={2017}
}