hi is the resnet part for extract features use pretrained model?(im too noob to make sure that i understood the code right..) )

yuxiangsun / RTFNet

RGB-Thermal Fusion Network for Semantic Segmentation of Urban Scenes

MIT License

166 stars 39 forks source link

hi is the resnet part for extract features use pretrained model?(im too noob to make sure that i understood the code right..) ) #3

Closed lvcat closed 5 years ago

yuxiangsun commented 5 years ago

Yes, it loads the pre-trained ResNet weight using: pretrained=True

lvcat commented 5 years ago

Yes, it loads the pre-trained ResNet weight using: pretrained=True

thanks, it matters to me. : )

lvcat commented 5 years ago

Yes, it loads the pre-trained ResNet weight using: pretrained=True

hi me again. Would u mind provide the training log that record the loss and acc. i want to reference it for my training= = . Thank u very much.QAQ

lvcat commented 5 years ago

您好我按照您论文里的参数重新训练试了一下，跟您论文里的以及给出的权重有一些差距，有什么我应该注意的小细节吗

yuxiangsun commented 5 years ago

You may want to set drop_last as false for the train dataloader. But I think that the initialization scheme mainly introduces the variance, you may try training it again.