Moeinh77 / resnext

Implementation of ResNext by chainer (Aggregated Residual Transformations for Deep Neural Networks: https://arxiv.org/abs/1611.05431)

What's this

Implementation of ResNext by chainer

Dependencies

git clone https://github.com/nutszebra/resnext.git
cd resnext
git submodule init
git submodule update

How to run

python main.py -g 0

Details about my implementation

All hyperparameters and network architecture are the same as in [1] except for data-augmentation.

• Data augmentation
  Train: Pictures are randomly resized in the range of [32, 36], then 32x32 patches are extracted randomly and are normalized locally. Horizontal flipping is applied with 0.5 probability.
  Test: Pictures are resized to 32x32, then they are normalized locally. Single image test is used to calculate total accuracy.

Cifar10 result

network	model detail	total accuracy (%)
ResNext [1]	16x64d	96.42
ResNext [1]	8x64d	96.35
ResNext [1]	2x64d	95.98 (by my eyes on Figure 7)
my implementation	2x64d	95.72

Converted ILSVRC pretrained model

C	d	total layers	url	original model: ILSVRC top-1 error (%)	converted model: ILSVRC top-1 error (%)	converted model: ILSCRC top-5 error (%)
64	4	101	https://1drv.ms/u/s!AtHe5bQGa25xiIswQGB9cdcHWDUhNA	20.4	21.4	5.86
32	4	101	https://1drv.ms/u/s!AtHe5bQGa25xiIsyHnhhdNNcugAqLA	21.2	22.3	6.24
32	4	50	https://1drv.ms/u/s!AtHe5bQGa25xiIsxun5XuoIpd_bFjg	22.2	23.4	6.96

How to convert ILSVRC pretrained model by yourself

To run ilsvrc_converter.py, you need to install pytorch and download t7 file from here, then type like this on terminal:

ipython
run ilsvrc_converter.py -t /path/to/t7/file
model.save_model('path/to/save/model')

Note: Please do not rename t7 file

How to test converted ILSVRC pretrained model

run test_ilsvrc.py -g 0 -b 16 -c 32 -d 4 -l 50 -m /path/to/converted/chainer/model -ld ./ILSVRC

g: gpu number
b: batch number
c: cardinality
d: d of Cxd
l: total layers
m: converted model
ld: path to root directory of ilsvrc

How to load converted ILSVRC pretrained model

import sys
sys.path.append('./trainer')
import resnext_ilsvrc
import data_augmentation

layers = 101 # this parameter is up to model
C = 64 # this parameter is up to model

if layers == 101 and C == 64:
    model = resnext.ResNext(1000, block_num=(3, 4, 23, 3), C=C, d=d, multiplier=1)
elif layers == 101 and C == 32:
    model = resnext.ResNext(1000, block_num=(3, 4, 23, 3), C=C, d=d, multiplier=2)
elif layers == 50 and C == 32:
    model = resnext.ResNext(1000, block_num=(3, 4, 6, 3), C=C, d=d, multiplier=2)
else:
    model = resnext.ResNext(1000, C=C, d=d)

model.load_model('path/to/converted/chainer/model')
model.check_gpu(-1) # -1 means cpu. If you'd like to use gpu, give gpu id here

preprocess = data_augmentation.DataAugmentationNormalizeBigger

img = preprocess.test('path/to/image').x
x = model.prepare_input([x], volatile=True)
y = model(x, train=False)

References

Aggregated Residual Transformations for Deep Neural Networks [1]