zt1112/pytorch-priv

Pytorch implementation for Classification, Semantic Segmentation and Object Detection

pytorch-priv

Pytorch implementation for Classification, Semantic Segmentation, Pose Estimation and Object Detection

• Image Classification
• Semantic Segmentation (progressing...)
• Object Detection (progressing...)
• Pose Estimation (progressing...)

Install

• Install PyTorch>=0.3.0
• Install torchvision>=0.2.0
• Clone recursively

  git clone --recursive https://github.com/soeaver/pytorch-priv
  

• pip install easydict

Training and Evaluating

For training:

1. Modify the .yml file in ./cfg/imagenet/air50-1x64d:

   • the ckpt is used to save the checkpoints
   • if you want use cosine learning rate, please set cosine_lr: True, then lr_schedule and gamma will not be used
   • for resuming training, add the model.pth.tar to resume: and modify start_epoch
   • rotation, pixel_jitter and grayscale are extra data augmentation, recommended for training complex networks only

2. Train a network:

   python cls_train.py --cfg ./cfg/imagenet/air50_1x64d.yml 
   

   2.1 Training with mixup (optional):

   python tools/cls_mixup_train.py --cfg ./cfg/imagenet/air50_1x64d_mixup.yml 
   

   for better performace:

   • double the epochs for training with mixup
   • a few extra epochs with no mixup after the process above

   2.2 Ttraining cifar dataset (optional):

   python tools/cls_cifar.py --cfg ./cfg/cifar10/resnext29_8x64d.yml
   

   or with mixup (usually weight_decay: 0.0001):

   python tools/cls_mixup_cifar.py --cfg ./cfg/cifar10/resnext29_8x64d_mixup.yml
   

For evaluating:

1. Modify the .yml file in ./cfg/cls/air50-1x64d:

   • add the model.pth.tar to pretrained:
   • set the evaluate: True

2. Evaluate a network:

   python train_cls.py --cfg ./cfg/imagenet/  air50_1x64d.yml 
   

For evaluating image by image:

1. Modify the tools/cls_eval.py file

2. Evaluate a network:

   python tools/cls_eval.py
   

Features

• Aligned Inception ResNet (air)
• Cosine Learning Rate
• Mixup
• Random-Erasing

Results

ImageNet1k

Single-crop (224x224) validation error rate is reported.

Network	Flops (M)	Params (M)	Top-1 Error (%)	Top-5 Error (%)	Speed (im/sec)
resnet50-1x64d	4109.4	25.5	22.96	6.54	160.1
air50-1x32d	1543.9	9.5	24.99	7.62	91.7
air50-1x64d	6148.2	35.9	21.02	5.55	86.9
air50-1x80d	9597.9	55.4	20.50	5.41	81.3
air101-1x64d	11722.9	64.4	20.13	5.02	48.1

• Speed test on single Titan xp GPU with batch_size: 1.

Cifar10 & Cifar100

Single-crop (32x32) validation error rate is reported.

Network	Flops (M)	Params (M)	Cifar10 Top-1 Error (%)	Cifar100 Top-1 Error (%)
resnext29-8x64d	5387.2	34.4	3.73	18.55
resnext29-8x64d-mixup	5387.2	34.4	2.90	--
resnext29-8x64d-ea	5387.2	34.4	3.55	--

License

pytorch-priv is released under the MIT License (refer to the LICENSE file for details).

Contribute

Feel free to create a pull request if you find any bugs or you want to contribute (e.g., more datasets and more network structures).