search

bubbliiiing / ssd-keras

这是一个ssd-keras的源码，可以用于训练自己的模型。
MIT License
167 stars 68 forks source link

ssd 訓練loss 降不下去 #25

Open j2233ack opened 2 years ago

j2233ack commented 2 years ago

你好，請問一下我用自己標註的資料，訓練很多次得到的Train loss大概都是 24.361875534057617 23.856924057006836 23.58797836303711 20.466842651367188 17.641437530517578 16.859132766723633 16.480562210083008 16.4239501953125 16.142189025878906 15.949203491210938 16.07608985900879 15.916677474975586 16.07317543029785 15.945028305053711 15.92853832244873 16.18751335144043 15.703150749206543 15.671076774597168 15.861603736877441 15.706280708312988 15.933178901672363 15.665262222290039 16.22136116027832 16.081159591674805 16.072917938232422 15.817296981811523 15.466055870056152 15.689082145690918 15.644898414611816 15.513778686523438 15.390292167663574 15.63227367401123 15.55762767791748 15.397815704345703 15.544672012329102 15.496471405029297 15.702784538269043 15.536251068115234 15.315775871276855 15.344732284545898 15.502330780029297 15.26860523223877 15.21063232421875 15.171586990356445 15.373653411865234 15.013505935668945 15.055233001708984 14.86623764038086 14.83223819732666 14.92117691040039 14.696207046508789 14.838573455810547 14.655868530273438 14.613218307495117 14.572738647460938 14.700054168701172 14.290406227111816 14.5927734375 14.125246047973633 14.32200813293457 14.265861511230469 14.036972999572754 14.076470375061035 13.975605010986328 13.952387809753418 14.032809257507324 13.8345308303833 14.009840965270996 13.703845024108887 13.896482467651367 13.783814430236816 13.866425514221191 14.005176544189453 13.810897827148438 13.803141593933105 13.929206848144531 這樣就降不下去... 有什麼建議能讓模型順利收斂嗎？ 感謝

我影像尺寸是960x540，我修改一些參數 ` #------------------------------------------------------#

input_shape 输入的shape大小

#------------------------------------------------------#
input_shape     = [960, 540]
#------------------------------------------------------#
#   可用于设定先验框的大小，默认的anchors_size
#   是根据voc数据集设定的，大多数情况下都是通用的！
#   如果想要检测小物体，可以修改anchors_size
#   一般调小浅层先验框的大小就行了！因为浅层负责小物体检测！
#   比如anchors_size = [21, 45, 99, 153, 207, 261, 315]
#------------------------------------------------------#
# anchors_size    = [30, 60, 111, 162, 213, 264, 315]
anchors_size = [21, 45, 99, 153, 207,315]`
bubbliiiing commented 2 years ago

emm……你先按默认的来一遍吧