fix depth of resnet/preresnet on cifar10/cifar100

[luzai]

Hi~ thanks for your nice code!
I made some modification, according to the discussion in #6 and #8:

• The default building block for resnet on cifar10/cifar100 is BasicBlock
• Depth=6n+2 when using basicblock and Depth=9n+2 when using bottleneck
• Add an command line option to choose building block.

And I do some experiments using resnet and cifar10. The training receipts and results are:

python cifar.py -a resnet --depth 164 --block-name Bottleneck --epochs 164 --schedule 81 122 --gamma 0.1 --wd 1e-4 --checkpoint checkpoints/cifar10/resnet-164-bottleneck 

python cifar.py -a resnet --depth 110 --block-name BasicBlock --epochs 164 --schedule 81 122 --gamma 0.1 --wd 1e-4 --checkpoint checkpoints/cifar10/resnet-110-basicblock

python cifar.py -a resnet --depth 110 --block-name Bottleneck --epochs 164 --schedule 81 122 --gamma 0.1 --wd 1e-4 --checkpoint checkpoints/cifar10/resnet-110-bottleneck 

block name	depth	params(M)	best top-1 error(%)
Bottleneck	164	1.70	6.26
Bottleneck	110	1.14	6.71
BasicBlock	110	1.73	6.60

[BIGBALLON]

hello, @bearpaw ,
I think this merge pull request is important since the preresnet-110 in original code is not 110 layers, so the results is wrong. hope to fixed the bug and update readme.md , please!

[mountains-high]

@luzai , thank you for your update about depth. I am getting an assertion error when I wanted to use resnet34.
Could you help me to solve this issue, please?
Thank you

[luzai]

Hi mountains-high, I guess the assert error comes from 6n+2. It means if n=5, then the number of layers is 32.

@luzai , thank you for your update about depth. I am getting an assertion error when I wanted to use resnet34. Could you help me to solve this issue, please? Thank you