1.perform object detection on MSCOCO

2.perform Dependency Parsing on MSCOCO

3.select the key word by 1 and 2

1.py-bottom-top  :obj detection finished.

2.try to synthesize features.

1.和泉ヤニナ received my paper(2480)

2.Reproduce bottom up and top dowm attention https://arxiv.org/pdf/1707.07998.pdf

1.Finish the ICISP paperhttps://mountain.elcs.kyutech.ac.jp/gao/icisp2021_gao/-/blob/master/GaoYuliang-ICISP.pdf

2.learn to use the feature of Faster-RCNN

3.Read paper https://arxiv.org/pdf/1707.07998.pdf

1.Contribute to the Journal of Yangda University and revise it according to the review opinions

2.Completes the data of ICISP papers, as well as drafts

1.draw attention heat map and add it to chinese paper

2.modify chinese paper again

1.Retraining the model by gradually adjusting the learning rate.

Compared with constant learning rate learning:Shock reduction，

Maximum accuracy is Slightly improved（sometimes）,sometimes droped,unstable.

2.Modify chinses paper

3.submit the abstract of ICISP

1.finish chinese paper

2.SA-Inception-ResnetV2 model finished(for ICISP)

1.SA-LIGHT-INCEPTION-RESNET model finish

2.ready to write papers

1.Yangzhou university mid-defend

2.try different improvements to inception(CABM,SE,SK)