Implementation of ResNext by chainer
git clone https://github.com/nutszebra/resnext.git
cd resnext
git submodule init
git submodule update
python main.py -g 0
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.
| 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 |
| 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 |
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
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
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)
Aggregated Residual Transformations for Deep Neural Networks [1]