ResNet-50/101/152

There are two types of ResNet in Deep Residual Learning for Image Recognition, by Kaiming He et al. One for ImageNet and another for CIFAR-10.

I had implemented the ResNet-50/101/152 (ImageNet one) by Python with Tensorflow in this repo. You can train my ResNet-50/101/152 without pretrain weights or load the pretrain weights of ImageNet. I had trained and tested my ResNet-50/101/152 on Kaggle Plant Seedings Classification.

My Environment

Environment 1

• Operating System:
  • Arch Linux 4.20.7-1
• Memory
  • 64GB
• CUDA:
  • CUDA V10.0.130
• CUDNN:
  • CUDNN 7.0.5-2
• GPU:
  • GTX 1070 8G
• Nvidia driver:
  • 390.25
• Python:
  • python 3.6.4
• Python package:
  • tqdm, bs4, opencv-python, pydensecrf, cython...
• Tensorflow:
  • tensorflow-gpu 1.5.0

Downloading the plant seeding dataset from Kaggle

• Plant seeding dataset

Setup Dataset

My directory structure

./ResNet/
├── dataset
├── models
└── pretrain_models

Plant seeding dataset directory structure

./train/
├── Black-grass
├── Charlock
├── Cleavers
├── Common Chickweed
├── Common wheat
├── Fat Hen
├── Loose Silky-bent
├── Maize
├── Scentless Mayweed
├── Shepherds Purse
├── Small-flowered Cranesbill
└── Sugar beet

• Put train.zip and test.zip (Plant seeding dataset) into {PATH}/ResNet-PreAct/ResNet/dataset/

mv {PATH}/train.zip {PATH}/ResNet-PreAct/ResNet/dataset/
mv {PATH}/test.zip {PATH}/ResNet-PreAct/ResNet/dataset/

• Unzip train.zip and test.zip to the {PATH}/ResNet-PreAct/ResNet/dataset

unzip train.zip
unzip test.zip

• Resize and split the dataset into train, val and test

python ./dataset/resize.py

Demo (See Usage for more details)

Download pretrain model training on plant seeding dataset

• Pretrain model
  • Move files from resnet_v2_50_val to '{PATH}/models/resnet_v2_50'
• Run test

  python test.py --layers 50 --batch 1 --recover 272  
  

• Run train (See Training for more details)

  python train.py --layers 50 --batch 16 --val 0 --recover 272 --pretrain 0  
  

• Performance

Key	Value
Method	ResNet50-Preact
Language	Python with Tensorflow
Pretrain weight	ImageNet
Dataset	20% to validation
Best epoch	272
Validation set accuracy	96.72%
Submission time	2019/1/18
Result	96.72%

• Other performance

Training (See Usage for more details)

Download pretrain weights of ImageNet

• tensorflow/models
  • Put resnet_v2_*.ckpt in 'pretrain_models'

Train network

• Set parameters
  • set number of layers for ResNet
  • set loading pretrain weights of ImageNet or not
  • set adding validation set into training set or not
  • Check Usage for other parameters

python train.py --layers [50|101|152] --pretrain [0|1] --val [0|1]    

• For example

python train.py --layers 50 --pretrain 1 --val 0    

Check performance (See Usage for more details)

Test validation set

python val.py --layer [50|101|152] --recover [0+]    

• For example

python val.py --layer 50 --recover 299    

Testing (See Usage for more details)

Test network

• Result will be stored as '.csv' file in {PATH}/ResNet/dataset/

python test.py --layers [50|101|152] --batch 1 --recover [0+]

• For example

python test.py --layers 50 --batch 1 --recover 299

Training on your own dataset

• Put the images of training set into {PATH}/ResNet/train/{class_name}/
• If you have testing set, put it into {PATH}/ResNet/test/
• Modify the class name in resize.py

Usage

dataset/resize.py

• Resize and split the dataset into train, val, test

eval_.py

• Compute the accuracy of the result
• May be called by train.py and val.py

load_data.py

• Load and preprocess dataset
• May be called by train.py, val.py and test.py

parser.py

• Parse the command line argument
• May be called by train.py, val.py and test.py

resnet.py

• The implementation of the network

test.py

• Test the testing set and store the result in the 'dataset' as .csv file

usage: test.py [-h] [--layers LAYERS] [--classes CLASSES] [--batch BATCH]
               [--recover RECOVER]

optional arguments:
  -h, --help         show this help message and exit
  --layers LAYERS    select number of layers from [50, 101, 152] (default:
                     101)
  --classes CLASSES  number of classes (default: 12)
  --batch BATCH      batch size (default: 16)
  --recover RECOVER  recover weights (default: 0)

train.py

• Train the network

usage: train.py [-h] [--layers LAYERS] [--classes CLASSES] [--batch BATCH]
                [--epoch EPOCH] [--lr LR] [--pretrain PRETRAIN] [--save SAVE]
                [--recover RECOVER] [--val VAL]

optional arguments:
  -h, --help           show this help message and exit
  --layers LAYERS      select number of layers from [50, 101, 152] (default:
                       101)
  --classes CLASSES    number of classes (default: 12)
  --batch BATCH        batch size (default: 16)
  --epoch EPOCH        max epoch (default: 300)
  --lr LR              learning rate (default: 0.001)
  --pretrain PRETRAIN  load pretrain weights or not. [0, 1]=[False, True]
                       (default: 1)
  --save SAVE          save step for saver (default: 1)
  --recover RECOVER    recover weights and continue training ot not. [-1,
                       0+]=[False, True] (default: -1)
  --val VAL            add validation set into training set or not. [0,
                       1]=[False, True] (default: 0)

val.py

• Test the validation set

usage: val.py [-h] [--layers LAYERS] [--classes CLASSES] [--batch BATCH]
              [--recover RECOVER]

optional arguments:
  -h, --help         show this help message and exit
  --layers LAYERS    select number of layers from [50, 101, 152] (default:
                     101)
  --classes CLASSES  number of classes (default: 12)
  --batch BATCH      batch size (default: 16)
  --recover RECOVER  recover weights (default: 0)

Reference

• [1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. Deep Residual Learning for Image Recognition.
• [2] Tensorflow Slim