gaojiah / octconv-tfkeras

Unofficial implementation of Octave Convolutions (OctConv) in TensorFlow / Keras.

OctConv-TFKeras

Unofficial implementation of Octave Convolutions (OctConv) in TensorFlow / Keras.

Y. Chen, H. Fang, B. Xu, Z. Yan, Y. Kalantidis, M. Rohrbach, S. Yan, J. Feng. Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution. (2019). https://arxiv.org/abs/1904.05049

(Update 2019-04-26) Official implementation by MXNet is available : https://github.com/facebookresearch/OctConv

Usage

from oct_conv2d import OctConv2D
# high, low = some tensors or inputs
high, low = OctConv2D(filters=ch, alpha=alpha)([high, low])

Colab Notebook

Train OctConv ResNet (TPU)
https://colab.research.google.com/drive/1MXN46mhCk6s-G_nfJrH1B6_8GXh-a_QH

Measuring prediction time (CPU)
https://colab.research.google.com/drive/12MdVXyB9K3FnpzYNmyc3qu5s59-53WNE

CIFAR-10

Experimented with Wide ResNet (N = 4, k = 10). Train with colab TPUs.

alpha	Test Accuracy
0	88.68%
0.25	94.25%
0.5	94.06%
0.75	93.66%

Prediction Time

CPU and GPU are colab environment. On CPU, use 256 samples for prediction, and on GPU, use 50000 samples for prediction. Both CPU and GPU are 32x32 resolution each.

CPU

alpha/s	Mean	Median	S.D.	Median/sample[ms]	Relative measured value	Theoretical FLOPs cost
0	39.18	38.96	0.6807	152.19	100	100
0.25	29.79	29.55	0.7705	115.43	76	67
0.5	20.61	20.46	0.5052	79.92	53	44
0.75	14.38	14.17	0.7874	55.35	36	30

Theoretical FLOPs cost are from the paper.

GPU

alpha/s	Mean	Median	S.D.	Median/sample[ms]	Relative measured value
0	60.43	59.94	1.693	1.20	100
0.25	62.24	62.09	0.6996	1.24	104
0.5	47.87	47.73	0.6224	0.95	80
0.75	34.15	34	0.6747	0.68	57

You can see that the time spent on prediction decreases with theoretical FLOPs costs, both CPU and GPU.

Details (Japanese)

[最新論文]Octave Convolution（OctConv）を試してみる
https://qiita.com/koshian2/items/0e40a5930f1aa63a66b9