squeezezenet_ssd_lite correct rate is very low about pytorch-ssd

[AlanStark]

What kind of training strategy you've been used? SGD usually requires more training iterations to converge.

[qfgaohao]

Hi @wangnamu @AlanStark a pretrained ImageNet model can help. Also use a smaller learning rate, as squeezenet doesn't have batchnorm layers. Increasing the number of anchors may help as well.

Be aware, the squeeze net here is modified, as the original impl is not onnx-compatible due to the ceiling_mode in polling layers.

If you don't have to use Squeezenet, Mobilenet v2 with width-multi 0.6 has the similar speed as Squeezenet, but has higher accuracy. I don't have the exact benchmark numbers. but it's worth to try.

[wangnamu]

hi @qfgaohao I will try your suggestion, thank you very much.