Ootzk / SsENet.PyTorch

An experimental implementation to verify variation idea to Squeeze-and-Excitation Networks(SENet)

SsENet.PyTorch

An experimental implementation to verify variation idea to SENet(Squeeze-and-Excitation Networks).

Prerequisites

This code is under following environment:

• python=3.9.4
• pytorch=1.9.1
• torchvision=0.10.1

Explanation of Original Squeeze operation

Above figure shows the concept idea of SENet. They squeezed by taking representative value(scalar) per channel. (H * W * C -> 1 * 1 * C)
In original paper, to exploit input-specific descriptor z, global-average-pooling(nn.AdaptiveAvgPool2d on PyTorch) was used - it is just the average value of each 'image' whose size is H * W.

Variant Squeeze operations

We've conducted experiments under same training recipe but only different squeeze operation. The squeeze operation we used are:

1. baseline (no squeeze operation; naive architecture)
2. gap (global average pooling)
3. gmp (global max pooling)
4. std (standard deviation of pixels in H * W)
5. gapXstd (gap * std)
6. random (random value that have distribution on the interval [0, 1). not extracted from H * W)

Experimental Results

• ResNet18 + ImageNet

Squeeze operation	Top-1 Acc (%)	Top-5 Acc (%)	Improved Top-1 %p
baseline	65.71	86.29	+0.00
gap	66.91	87.35	+1.20
gmp	66.73	87.20	+1.02
std	66.47	86.94	+0.76
gapXstd	66.90	87.18	+1.19
random	65.36	86.22	-0.35

• ResNet50 + ImageNet

Squeeze operation	Top-1 Acc (%)	Top-5 Acc (%)	Improved Top-1 %p
baseline	70.94	89.84	+0.00
gap	70.12	89.34	-0.82
gmp	70.36	89.23	-0.58
std	70.34	89.33	-0.60
gapXstd	70.17	89.26	-0.77
random	68.99	88.71	-1.95