search

liuzhuang13 / DenseNet

Densely Connected Convolutional Networks, In CVPR 2017 (Best Paper Award).
BSD 3-Clause "New" or "Revised" License
4.69k stars 1.06k forks source link

What is architecture of (L=40,k=12) DenseNet on Cifar10 #45

Closed Johnson-yue closed 5 years ago

Johnson-yue commented 6 years ago

I will implement DenseNet on CIFAR but， I don·t know how many dense block 。

DanielTea commented 6 years ago

Do a model.summary() to print the architecture.

DanielTea commented 6 years ago

ups sorry wrong language

Johnson-yue commented 6 years ago

@DanielTea Sorry，It is my fault. In the original paper: Densely Connected Convolutional Networks, it provided ImageNet architecture that shown in Table 1。 But It is not provide CIFAR model，for example, how many DenseNet-B block? I only know, total layers is 100 or 250 or 190

Ouwen commented 5 years ago

@Johnson-yue For cifar10, there are 3 dense blocks. There will be 1 conv2d layer initially which is a 3x3 conv, stride 1, and 3 transition layers (4 other layers)

So for depth = 40, we take 40-4 (other layers) = 36, then we take 36/3 = 12,

So there are 12 layers in each dense block. If bottlenecks are used, then there will be 6 layers and 6 bottlenecks in each dense block.

Hope this helps!

yuffon commented 5 years ago

@Johnson-yue For cifar10, there are 3 dense blocks. There will be 1 conv2d layer initially which is a 3x3 conv, stride 1, and 3 transition layers (4 other layers)

So for depth = 40, we take 40-4 (other layers) = 36, then we take 36/3 = 12,

So there are 12 layers in each dense block. If bottlenecks are used, then there will be 6 layers and 6 bottlenecks in each dense block.

Hope this helps!

Can you post out your training curve of densenet-40-12 on cifar10? On my computer the acc is bad except for very deep densenet.

Johnson-yue commented 5 years ago

@yuffon Sorry, It is long time before,so I don't have any information about this.