Rocket Launching

PyTorch code for "Rocket Launching: A universal and efficient framework for training well-performing light net" https://arxiv.org/abs/1708.04106

About this code

This code is based on the attention-transfer code, the code uses PyTorch.

What's in this repo so far:

Rocket-Interval code for CIFAR-10,CIFAR-100 experiments
Code for Rocket-Bottom (ResNet-16-ResNet-40)
gradient block
parameter sharing

bibtex:

@article{zhou2017Rocket,
  title={Rocket Launching: A unified and effecient framework for training well-behaved light net},
    author={Zhou, Guorui and Fan, Ying and Cui, Runpeng and Bian, Weijie and Zhu, Xiaoqiang and Kun, Gai},
      journal={arXiv preprint arXiv:1708.04106},
        year={2017}
}

Requirements

First install PyTorch, then install torchnet:

pip install git+https://github.com/pytorch/tnt.git@master

Then install OpenCV with Python bindings (e.g. conda install -c menpo opencv3), and other Python packages:

pip install -r requirements.txt

Experiments

Table 1

This section describes how to get the results in the table 1 of the paper.

Third column, train light basic:

python rocket_interval.py --save logs/resnet_16_1_basic --depth 16 --width 1
python rocket_interval.py --save logs/resnet_16_2_basic --depth 16 --width 2
python rocket_bottom.py --save logs/resnet_bottom_16_1_basic --depth 16 --width 1

Ninth column, train booster only:

python rocket_interval.py --save logs/resnet_40_1_booster --depth 40 --width 1
python rocket_interval.py --save logs/resnet_40_2_booster --depth 40 --width 2

Fouth column, train attention transfer:

python rocket_interval.py --save logs/at_16_1_40_1 --width 1 --teacher_id resnet_40_1_booster --beta 1e+3
python rocket_interval.py --save logs/at_16_2_40_2 --width 2 --teacher_id resnet_40_2_booster --beta 1e+3

Fifth column, train with KD:

python rocket_interval.py --save logs/kd_16_1_16_2 --width 1 --teacher_id resnet_40_1_booster --alpha 0.9
python rocket_interval.py --save logs/kd_16_2_16_2 --width 2 --teacher_id resnet_40_2_booster --alpha 0.9
python rocket_bottom.py --save logs/kd_bottom_16_1_40_1 --teacher_id resnet_40_1_booster --alpha 0.9

Sixth column and eighth column, train rocket launching:

python rocket_interval.py --save logs/rocket_interval_16_1_40_1 --width 1 --student_depth 16  --depth 40 --gamma 0.03 
python rocket_interval.py --save logs/rocket_interval_16_2_40_2 --width 2 --student_depth 16  --depth 40 --gamma 0.03 
python rocket_bottom.py --save logs/rocket_bottom_16_1_40_1 --width 1 --student_depth 16  --depth 40 --gamma 0.03

Seventh column, train rocket launching with KD:

python rocket_interval.py --save logs/rocket_interval_16_1_40_1_R_KD --width 1 --student_depth 16  --depth 40 --teacher_id resnet_40_1_booster --gamma 0.03 --alpha 0.9
python rocket_interval.py --save logs/rocket_interval_16_2_40_2_R_KD --width 2 --student_depth 16  --depth 40 --teacher_id resnet_40_2_booster --gamma 0.03 --alpha 0.9
python rocket_bottom.py --save logs/rocket_bottom_16_1_40_1_R_KD --width 1 --student_depth 16  --depth 40 --teacher_id resnet_40_1_booster --gamma 0.03 --alpha 0.9

Table 2

rocket launching without gradient block

python rocket_interval.py --save logs/rocket_interval_16_1_40_1_no_gb --width 1 --student_depth 16  --depth 40 --gamma 0.03 --grad_block False

rocket launching without parameter sharing

python rocket_interval.py --save logs/rocket_interval_16_1_40_1_no_ps --width 1 --student_depth 16  --depth 40 --gamma 0.03 --param_share False

Table 4 (only Cifar 100)

The same running command with Table 1 just add parameter --dataset CIFAR100

For example: rocket-launching

python rocket_interval.py --save logs/rocket_interval_16_1_40_1_cifar100 --width 1 --student_depth 16  --depth 40 --gamma 0.03 --dataset CIFAR100

lilujunai / rocket-launching Goto Github PK