Mmdetection: finetune CascadeRcnn generate too large model weights

[BeBeauty]

hello, I fine tune 'cascade_rcnn_x101_64x4d_fpn_1x' on my own datasets to detect 4 classes(img_scale:(1280x1024)). After 13 training epochs, the size of generated weights is 1G, which is quite larger than the size of pre_trained model(rcnn_x101_64x4d_fpn_1x:500M). Then I change the img_scale from (1280x1024) to (320x256),the size of generated model weights is still nearly 900M. I wonder if there any methods to minimize the model weights and decrease the inference time in prediction?
Thank you!

[karlind]

It is irrelated to img_scale. mmdetection will save optimizer's status by default, which is almost the same size as model's parameters. So if you don't want optimizer's status, just setcheckpoint_config = dict(interval=1, save_optimizer=False). It should halve the saved pth's size.

[BeBeauty]

It is irrelated to img_scale. mmdetection will save optimizer's status by default, which is almost the same size as model's parameters. So if you don't want optimizer's status, just setcheckpoint_config = dict(interval=1, save_optimizer=False). It should halve the saved pth's size.

It does works! Thank you very much!

[karlind]

If the problem has been solved, please close the issue.
Feel free to reopen.