Finally!! I created this blog. I’ll be posting articles and codes about deep learning especially focused on vision criteria. Thank U~

Reference: I referred my blog format from ratsgo’s blog

Sick of VGGNet and ResNet??

We have been seeing VGGNet and ResNet for years!! I think it’s time to move on!!

Here are the fastest and most accurate convolution neural networks for year 2018~

.

Top accuracy

1. AmoebaNet

2. NASNet

3. SENet

4. DPN

5. PolyNet

.

Info. Density (Less calculation)

1. SqueezeNext

2. TinyDarkNet

3. SqueezeNet

4. MobileNet

5. ZynqNet

And here are the ratings by overall NetScore proposed by this paper which rated 60 different deep CNNs for the ILSVRC 2012 dataset.

NetScore

1. SqueezeNext

2. CondenseNet

3. MobileNetv2

4. TinyDarkNet

5. MobileNetv1 (1.0-128)

6. MobileNetv1 (1.0-160)

7. ShuffleNet (x2)

8. ShuffleNet (1.5)

9. MobileNetv1(1.0-192)

10. ZynqNet

26. DenseNet-121 (k=32)

33. ResNet-50

.

We should note that less computation does not mean faster testing time. But if accuracy is similar, models with less computation can be more beneficial. Also implementation on mobile phones would be much better.

Also we should note that this post doesn’t mean that SqueezeNext is way better than ResNet. There should be further testing with other datasets.

Here is the link to SqueezeNext which I will try to write a new post covering this network someday.