jeandut
commented
7 years ago I get an even worse mAP by simply running with the default parameters every item is at AP=-1 , I made sure no overflow occured during training and I tried with two different learning rate and starting points any idea what could be the issue ? It seems the loss is not converging I seem to remember seeing an issue on unstable training in this repo.
For the time being I am writing a script to check RoI gradients as the pretrained model working well means that the forward is indeed ok.
Otherwise I do not know maybe trying to fiddle even more with the learning rate, momentum and inits...
ChiefGodMan
yelusaleng
Zealoe
Dose anyone try to train model not use pre-trained model? My mAP is very poor, where is the problem?TF init weights?
AP for aeroplane = 0.032 AP for bicycle = 0.0193 AP for bird = 0.0110 AP for boat = 0.0026 AP for bottle = -1.0000 AP for bus = 0.0164 AP for car = 0.1464 AP for cat = 0.0370 AP for chair = 0.0194 AP for cow = 0.0055 AP for diningtable = 0.0107 AP for dog = 0.0425 AP for horse = 0.1420 AP for motorbike = 0.1233 AP for person = 0.0693 AP for pottedplant = 0.0011 AP for sheep = -1.0000 AP for sofa = 0.0182 AP for train = 0.0188 AP for tvmonitor = 0.0083 Mean AP = -0.0638