train error!!!

[zhepherd]

[xingyizhou]

Hi, can you check if the number of GPUs in your machine matches the length of batchsize list https://github.com/xingyizhou/ExtremeNet/blob/master/config/ExtremeNet.json#L22 ?

[zhepherd]

@xingyizhou my GPUs is 4

[xingyizhou]

Then you should change this line into "chunk_sizes": [4, 5, 5, 5], and change this line into "batch_size": 19, and expect a different result from the paper.

[zhepherd]

thx @xingyizhou

[zhepherd]

one more question， how to train my own data， where i need to change？

[xingyizhou]

Basically, you will need to generate the annotation file https://github.com/xingyizhou/ExtremeNet/blob/master/tools/gen_coco_extreme_points.py, create the data preprocessing https://github.com/xingyizhou/ExtremeNet/blob/master/db/coco_extreme.py, and change the dataset in your config file.

[zhepherd]

i need to generate the annotation file like [(Xt,Yt, Xb,Yb,Xl,Yl, Xr, Yr),……]?

[xingyizhou]

Yes.

[zhepherd]

i also need to change the https://github.com/xingyizhou/ExtremeNet/blob/master/sample/coco_extreme.py and train.py ?

[zhepherd]

if i change my own data like coco data format and have 3 class , i think this way is more easy, so i want to ask where i need to change? @xingyizhou thx for help ~