search

BCV-Uniandes / DMS

Dynamic Multimodal Instance Segmentation Guided by Natural Language Queries, ECCV 2018
https://biomedicalcomputervision.uniandes.edu.co
MIT License
75 stars 8 forks source link

Why net.train() in evaluate function? #30

Closed RyanHTR closed 5 years ago

RyanHTR commented 5 years ago

Hi, i notice that the network is set to train mode in evaluate function. Is that a special setting? Does this line affect the results of evaluation?

https://github.com/BCV-Uniandes/DMS/blob/9fa3a3a2ef5980dd17e21b73234a4cd0b3d00e16/dmn_pytorch/train.py#L342

andfoy commented 5 years ago

Hi, we evaluate on train mode due to the unstability of BN blocks of the backbone network, i.e., DPN-92. As we train with batch size = 1, the model is unable to capture proper statistics per "batch" (Instance, actually)

andfoy commented 5 years ago

Due to a lack of responses, I'm closing this one