About model size

[shudct]

hi all, in the paper, the trained model is 5.4M on PASCAL VOC 2007 test. However, the link to the trained model from the author is 20M~. Does the network has some differenence? Anyone can explain that for me? THX!

[Robert-JunWang]

The 5.4M is the number of parameters. The actual model size on hard disk depends on the file format. To reduce mode size, you can store weights with float16 precision so that the actual model size can be halved and the mAP is almost the same. You can further reduce the model size by model pruning.

[shudct]

@Robert-JunWang Thank for your explanation. Caffe seems not support float16 for training and inference. Can you give me some advice for store the weights? THX!

[ujsyehao]

@shudct NVcaffe supports for float16.