pth-file size much larger than pre-trained coco file

[alexander-sony]

Have trained 5 classes on a custom dataset starting from config https://github.com/RangiLyu/nanodet/blob/main/config/nanodet-plus-m-1.5x_416.yml (adjusted num_classes) and https://github.com/RangiLyu/nanodet/blob/main/tools/train.py. Training went just fine and reached mAP 0.8 after 300 epochs. Resulting model_best.ckpt is 125MB and nanodet_model_best.pth is 31MB.

Here is my question:

• Why is my pth file 3x larger compared with the coco pretrained file from model zoo, NanoDet-Plus-m-1.5x-416, https://drive.google.com/file/d/16FJJJgUt5VrSKG7RM_ImdKKzhJ-Mu45I/view?usp=sharing

It seems that my pth file includes stuff only required for training. Anyone knows if this is the case and if it can be removed? Have tried torch.save(model.state_dict(), PATH) but the resulting file is still 30MB

model was created as in main() in https://github.com/RangiLyu/nanodet/blob/main/tools/export_onnx.py

Interestingly, the onnx model size is 10MB.

[alexander-sony]

iterate over the state_dict and remove items starting with "aux" - those are only requires during traing as remarked in the config file