CNN-backbone
This is a summary repository of the sota backbone of the CNNs
Quickly start
1. Download CIFAR-10 Dataset
Download the cifar-10 dataset and extract it to the ./Data directory.
You can change the download parameter download=True in the train file
or download it from the link below (recommended method):
baiduURL：https://pan.baidu.com/s/16U9FhTlv3BVuB3ixipayXg pw:ifdu
2. training model
The ./train.py can be used to train models contained in the ./models module.
Many comments are written in the main function .train.py to facilitate the interpretation of the training process
Code architecture
│ train.py # The training process and all parameters
│
├─Data # Downloaded dataset's location
│ │
│ └─cifar-10-batches-py
│ batches.meta
│ data_batch_1
│ data_batch_2
│ ...
├─lib
│ │ __init__.py # No implement
│
├─models # cnn models
│ │ DeformLeNet.py
│ │ densenet.py
│ │ GhostNet.py
│ │ LeNet.py
│ │ OctNet.py
│ │ OctResnet.py
│ │ ResNet18.py
│ │ __init__.py
│ │
│ ├─layer # some common layer
│ │ conv_layer.py
│ │ deform_conv_v2.py
│ │ OctConv.py
│ │ __init__.py
│
│─utils
│ │ common.py # some tools,such as Logger func
│
├─output # trained model will be saved here
│ └─test # a sample
│ net_best.pth # Best model for validation
│ net_latest.pth # Newly trained models
│ train_log.txt # training log