reproduce results for wresnet28x10 @ cifar10/100.

update usage with setup.py.

Author: ildoonet

README.md

@@ -4,13 +4,54 @@ Unofficial PyTorch Reimplementation of RandAugment. Most of codes are from [Fast
 
 ## Introduction
 
+TODO
+
 ## Install
 
+```bash
+$ pip install git+https://github.com/ildoonet/pytorch-randaugment
+```
+
 ## Usage
 
+```python
+from torchvision.transforms import transforms
+from RandAugment import RandAugment
+
+transform_train = transforms.Compose([
+    transforms.RandomCrop(32, padding=4),
+    transforms.RandomHorizontalFlip(),
+    transforms.ToTensor(),
+    transforms.Normalize(_CIFAR_MEAN, _CIFAR_STD),
+])
+
+# Add RandAugment with N, M(hyperparameter)
+transform_train.transforms.insert(0, RandAugment(N, M))
+```
+
 ## Experiment
 
+We use same hyperparameters as the paper mentioned. 
+
+### CIFAR-10
+
+| Model             | Paper's Result | Ours         |
+|-------------------|---------------:|-------------:|
+| Wide-ResNet 28x10 | 97.3           | 97.4         |
+| Shake26 2x96d     | 
+| Pyramid272        | TODO           |
+
+### CIFAR-100
+
+| Model             | Paper's Result | Ours         |
+|-------------------|---------------:|-------------:|
+| Wide-ResNet 28x10 | 83.3           | 83.3         |
+
+### ImageNet
+
+TODO
+
 ## References
 
-- RandAugment : https://arxiv.org/abs/1909.13719
+- RandAugment : [Paper](https://arxiv.org/abs/1909.13719)
 - Fast AutoAugment : [Code](https://github.com/kakaobrain/fast-autoaugment) [Paper](https://arxiv.org/abs/1905.00397)

RandAugment/__init__.py

@@ -0,0 +1,2 @@
+
+from RandAugment.augmentations import RandAugment

RandAugment/networks/wideresnet.py

@@ -4,6 +4,9 @@
 import numpy as np
 
 
+_bn_momentum = 0.1
+
+
 def conv3x3(in_planes, out_planes, stride=1):
     return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=True)
 
@@ -21,10 +24,10 @@ def conv_init(m):
 class WideBasic(nn.Module):
     def __init__(self, in_planes, planes, dropout_rate, stride=1):
         super(WideBasic, self).__init__()
-        self.bn1 = nn.BatchNorm2d(in_planes, momentum=0.9)
+        self.bn1 = nn.BatchNorm2d(in_planes, momentum=_bn_momentum)
         self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, padding=1, bias=True)
         self.dropout = nn.Dropout(p=dropout_rate)
-        self.bn2 = nn.BatchNorm2d(planes, momentum=0.9)
+        self.bn2 = nn.BatchNorm2d(planes, momentum=_bn_momentum)
         self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=True)
 
         self.shortcut = nn.Sequential()
@@ -56,7 +59,7 @@ def __init__(self, depth, widen_factor, dropout_rate, num_classes):
         self.layer1 = self._wide_layer(WideBasic, nStages[1], n, dropout_rate, stride=1)
         self.layer2 = self._wide_layer(WideBasic, nStages[2], n, dropout_rate, stride=2)
         self.layer3 = self._wide_layer(WideBasic, nStages[3], n, dropout_rate, stride=2)
-        self.bn1 = nn.BatchNorm2d(nStages[3], momentum=0.9)
+        self.bn1 = nn.BatchNorm2d(nStages[3], momentum=_bn_momentum)
         self.linear = nn.Linear(nStages[3], num_classes)
 
         # self.apply(conv_init)

confs/shake26_2x96d_cifar10_b512.yaml

@@ -0,0 +1,17 @@
+model:
+  type: shakeshake26_2x96d
+dataset: cifar10
+aug: fa_reduced_cifar10
+cutout: 16
+batch: 512
+epoch: 1800
+lr: 0.01
+lr_schedule:
+  type: 'cosine'
+  warmup:
+    multiplier: 4
+    epoch: 5
+optimizer:
+  type: sgd
+  nesterov: True
+  decay: 0.001

confs/wresnet28x10_cifar100_b512.yaml renamed to confs/wresnet28x10_cifar100_b256.yaml

@@ -6,13 +6,13 @@ randaug:
   N: 2
   M: 14
 cutout: 16
-batch: 512
+batch: 256
 epoch: 200
 lr: 0.1
 lr_schedule:
   type: 'cosine'
   warmup:
-    multiplier: 4
+    multiplier: 2
     epoch: 5
 optimizer:
   type: sgd

confs/wresnet28x10_cifar10_b256.yaml

@@ -0,0 +1,20 @@
+model:
+  type: wresnet28_10
+dataset: cifar10
+aug: randaugment
+randaug:
+  N: 3
+  M: 5    # from appendix
+cutout: 16
+batch: 256
+epoch: 200
+lr: 0.1
+lr_schedule:
+  type: 'cosine'
+  warmup:
+    multiplier: 2
+    epoch: 5
+optimizer:
+  type: sgd
+  nesterov: True
+  decay: 0.0005

requirements.txt

@@ -10,4 +10,4 @@ sklearn
 ray
 matplotlib
 psutil
-requests
+requests

setup.py

@@ -0,0 +1,26 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import setuptools
+
+_VERSION = '0.1'
+
+# 'opencv-python >= 3.3.1'
+REQUIRED_PACKAGES = [
+]
+
+DEPENDENCY_LINKS = [
+]
+
+setuptools.setup(
+    name='RandAugment',
+    version=_VERSION,
+    description='Unofficial PyTorch Reimplementation of RandAugment',
+    install_requires=REQUIRED_PACKAGES,
+    dependency_links=DEPENDENCY_LINKS,
+    url='https://github.com/ildoonet/pytorch-randaugment',
+    license='MIT License',
+    package_dir={},
+    packages=setuptools.find_packages(exclude=['tests']),
+)