[DEV] ssds.pytorch v1.5 roadmap

[foreverYoungGitHub]

This is the roadmap for ssds.pytorch v1.5. The development for ssds.pytorch v1.5 is fully reconstruct and almost done. The main features are listed at here:

Documentations:

• install;
• basic usage;
• basic parameters;
• sample tutorials;
• basic python api.

Framework:

• add training, validation, and inference support for pytorch v1.5 based model;
• add support to convert the model from pytorch to onnx, and allows the user to convert the onnx model the tensorrt 7 through the code in retinanet-examples.

Dataset:

• remove the current the voc dataset;
• add the dalicoco and dalitfrecord dataset for fast data loading.

Anchor box matching:

• remove the current anchor mathcing strategy;
• add the anchor box matching for each level to make user understand ssd-like training and inference easier.

Loss:

• FocalLoss
• SmoothL1
• IOU, GIOU, DIOU, CIOU Loss
• MultiBoxLoss (Not recommend, not fully tested)

Pipeline:

• add DataParallel for basic multiple gpu or single gpu training (slow)
• add apex for multiple gpu training (fast)

Visualization:

• add visualization for anchor strategy in each feature map (the distribution of scale and ratio in the dataset);
• add visualization for defualt anchor boxes in each feature map;
• prepare the images for readme.

Support SSDs head:

• ssd;
• fpn in retinanet;
• bifpn in efficientdet;
• yolov3 and yolov4
• shelf in shelfnet

Support backbone (feature extractor):

• resnet
• regnetx
• mobilenet v1 and v2
• shufflenet v2
• darknet
• densenet
• efficientNet (memory cost)

Others:

• Provide the dockerfile to allow user directly build the ssds.pytorch docker quickly;
• Provide the setup.py to allow user directly install the ssds.pytorch by pip;
• Prepare the pretrained models for different backbone and detection heads.

Bug Fix:

Please let me know if you have any problem when you use the ssds.pytorch or any suggestion to make the ssds.pytorch better!