search

yhenon / pytorch-retinanet

Pytorch implementation of RetinaNet object detection.
Apache License 2.0
2.14k stars 665 forks source link

Training on coco2014 using retinanet-101 just reached AP(@0.5:0.95) of 0.302 after 16 Epochs #142

Open LiRWZ opened 4 years ago

LiRWZ commented 4 years ago

Is that normal? or I need more Epochs?

SISTMrL commented 4 years ago

hello, I'm sorry to disturbing you due to this is my homework, I havae a question. I encounter this question: No such file: '/home/mrl/Documents/retinanet/RetinaNet/datasets/COCO/images/train2017/000000171857.jpg, when run the python code: for iter_num, data in enumerate(dataloader_train), did you face the same problem, thanks!

LiRWZ commented 4 years ago

sorry i have never met this erro.but i think this problem may cause by the picture (000000171857.jpg) is not in the folder.check it.if it's not there,download it.

---Original--- From: "SISTMrL"<notifications@github.com> Date: 2020/4/24 18:40:23 To: "yhenon/pytorch-retinanet"<pytorch-retinanet@noreply.github.com>; Cc: "Author"<author@noreply.github.com>;"LiRWZ"<942464482@qq.com>; Subject: Re: [yhenon/pytorch-retinanet] Training on coco2014 using retinanet-101 just reached AP(@0.5:0.95) of 0.302 after 16 Epochs (#142)

hello, I'm sorry to disturbing you due to this is my homework, I havae a question. I encounter this question: No such file: '/home/mrl/Documents/retinanet/RetinaNet/datasets/COCO/images/train2017/000000171857.jpg, when run the python code: for iter_num, data in enumerate(dataloader_train), did you face the same problem, thanks!

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

SISTMrL commented 4 years ago

I'm sorry to disturb you, when i see the code, i saw the retinanet.train() followed by retinanet.module.freeze_bn() in line 109 and 110 in train.py, i don't understood. I check the pytorch documentation, eval is used to model evaluation, why author use freeze_bn in train.py? thanks!

feiyuhuahuo commented 4 years ago

@SISTMrL I heard that when training batch size is small (e.g. <=4), then it's not suitable to use batch normalization.