Repository files navigation

2019.07.15 add OHEM and Mixup methods

2019.06.06 add warm up method of lr

2019.05.28 add reduced-resnet and different optimizers for cifar and fashionminist dtatset

2019.05.11 add support for cifar and fashionmnist dataset

Training and Testing on Python3.5

pytorch = 0.4.0torchvision>=0.2.0matplotlibnumpyscipyopencvpyyamlpackagingPILtqdmtime

• MINC-2500 is a patch classification dataset with 2500 samples per category. This is a subset of MINC where samples have been sized to 362 x 362 and each category is sampled evenly. Error rate and five fold cross validation are employed for evaluating. Based on resnet50, we can achieve a comparable result with state-od-the-arts.

• CIFAR100. In this experiment, we choose thereduced-resnet as our backbone network(you can choose yours).

• More dataset coming soon ......

1. Basic data augmentation methods
   • horizontal/vertical flip
   • random rot (90)
   • color jitter
   • random erasing
   • test augmentation
   • lighting noise
   • mixup
2. Multiple backbones
   • Resnet
   • Desnsenet
   • Reduced-resnet
3. Other methods
   • Focal loss
   • Label smooth
   • Combining global max pooling and global average pooling
   • Orthgonal center loss based on subspace masking
   • Learning rate warmup
   • OHEM(online hard example mining)

• MINC-2500. The data structure is following the Materials in Context Database (MINC)

• data/minc-2500
  • images
  • labels

• CIFAR100. The data would be automaticly downloaded to the folder： "./data"

• MINC-2500

python experiments/recognition/main.py - -dataset minc - -loss CrossEntropyLoss - -nclass 23 - -backbone resnet50 - -checkname test - -ocsm

• CIFAR100

python experiments/recognition/main.py - -backbone resnet_reduce - -res_reduce_depth 20 - -solver_type SGD - -lr-step 200,300 - -dataset cifar100 - -lr 0.1 - -epochs 375 - -batch-size 384 - -mixup

Note: (- -lr-step 200,300) indicates that leanrning rate is decayed by 10 at 200-th and 300-th epoch; (- -lr-step 200,) indicates that learning rate is decayed by 10 evary 200 epochs. (- - batch-size 384 - -ohem 192) indicates choosing 192 hard examples from 384 instances.

• MINC-2500. For example:

python experiments/recognition/main.py - -dataset minc - -nclass 23 - -backbone resnet18 - -test-batch-size 128 - -eval --resume experiments/recognition/runs/minc/deepten/09-3/*.pth

• CIFAR100. For example:

python experiments/recognition/main.py - -backbone resnet_reduce - -res_reduce_depth 20 - -dataset cifar100 - -test-batch-size 128 - -eval --resume experiments/recognition/runs/cifar100/deepten/0/*.pth

PyTorch Encoding

Random Erasing