search

lzx1413 / CAFFE_SSD

Other
106 stars 55 forks source link

change mbox_source_layers #3

Closed liangxi627 closed 6 years ago

liangxi627 commented 6 years ago

Hello, when I use the FSSD, I changed the mbox_source_layers and have a error when training my own dataset, and the training loss decline fast at first, finally maintain about 5%, how can I solve it?

F0512 14:40:32.193830 15163 solver.cpp:464] Check failed: result[j]->width() == 5 (1 vs. 5)

lzx1413 commented 6 years ago

can you run the training process on the VOC dataset successfully?

liangxi627 commented 6 years ago

Yes, it runs successfully on VOC, I meet the same error when I use the original SSD codes.

lzx1413 commented 6 years ago

It seems like that this error occurs while testing the model, the test result should have 5 elements rather than 1. So maybe there is something wrong with the dataset for testing?

liangxi627 commented 6 years ago

Ok, thank you, maybe I will check the dataset. And I found SSD is hard to training on small objects, is there any suggestions for the hyperparameters setting?

lzx1413 commented 6 years ago
  1. Set the min_size and max_size for priorbox generation with appropriate values for your objects.
  2. Use feature map with feature stride 4 may be helpful to small objects.
liangxi627 commented 6 years ago

Many thanks for your prompt help!