Note-Liu
Note-Liu
> 是可以的,需要你改一下那个model.py里面加载weights的函数,把遇到你新加模块那边的隔开就可以了,你可以参照我用80分类预训练模型给不同分类模型做预训练的写法,就是-pt那部分 > > 我在yolov3-mobilenet中颈部加入了cbam模块,但是训练的时候通道数出现问题,请问一下为什么会这样,加载或不加载预训练模型都一样。提示runtimeerror:Given Groups=1, weight ofsize 1024 160 1 1,except input [16, 1, 1, 1] to have 1024 channels, but got 1 channels instead 加载不加载预训练模型都报错,说明和预训练模型没关系。先能跑通再加载吧。
This problem has been solved!
TRAIN: LR: 0.001 MOMENTUM: 0.9 DECAY: 0.0005 BURN_IN: 5 MAXEPOCH: 300 COS: True SYBN: False#True MIX: True NO_MIXUP_EPOCHS: 30 LABAL_SMOOTH: True BATCHSIZE: 1 IMGSIZE: 608 IGNORETHRE: 0.7 train script: python...
If I use only one GPU , Is my train script right? thanks
> @Note-Liu have you solverd it?? no.[cry][cry]
I have the same issue,how did you solve it?
> 可以,更换网络结构就行; > nin_gc.py里模型就有用分组卷积结构,效果不错; > 大一些的数据集还需要测试。 对于shufflenet包含的channel shuffle结构,也能处理吗?
I set batch_size = 64 ,and im2col_step=64,but still meet the error
Why the parameter 'epoch' in scheduler.step(epoch) will affect the lr ?
> rovide the version of n Have you solved this problem yet?