yehao
yehao
I follow your advice > remove confidence prediction layers, I use lr = 0.0005. The result is 3 percentage points lower than author reports. I am trying to adjustment.
I do ablation experiments, the results verify if set param {1,1,2,0}, it will have higher accuracy.
about 0.2% - 0.5% higher.
Hi, @miaow1988 Can you share your tricks about training squeeze net v1.1 model? Thank you in advance!
@forresti Hi，I follow @miaow1988 tutorial and train the model, It performs better when removing relu after squeeze net layer, Do you have some ideas about it?
@goodwood123 Who tell you need squeezenet_iter_74000.caffemodel? You should train squeezenet-ssd model with a pretrain model trained on imagenet. you can refer #6 to know about the pretrain model.
what is training parameter with random initial value? You can use squeezenet-ssd hyper-parameters regardless of the dataset and object categories
I think @chuanqi305 use the pretrained weights on imagenet from [https://github.com/DeepScale/SqueezeNet](url)
@chuanqi305 Hi, Can you give me some advice? Thank you in advance.