about the weight class training

[sunpeng1996]

hi, I use the ENET calculate_class_weighting.py generare the loss weight for my training. and I find some problems.

First, the weight is： 0.0819 0.4754 0.1324 1.5224 1.5190 2.4730 8.1865 5.2286 0.1870 1.4695 0.6893 1.9814 7.8091 0.4164 1.3809 1.1982 0.6273 5.3535 4.0939 ,the distrubution is too large.

Second, I use the weight for loss to train, And the Mean IOU drop 7%. Could you give me some tricks and help? thx a lot!

[Eromera]

Hi! What do you mean the distribution is too large? If you mean that the values are higher than 1, that is ok, the softmax will squash all values to probabilities-range anyway.

Are you training encoder or the full network (decoder mode)?

For this 7% drop did you do any modification to the network or is it only changing the weights?

You can try the ones that are uploaded in the my erfnet_pytorch code (here):

if (enc):  #encoder
    weight[0] = 2.3653597831726 
    weight[1] = 4.4237880706787 
    weight[2] = 2.9691488742828 
    weight[3] = 5.3442072868347 
    weight[4] = 5.2983593940735 
    weight[5] = 5.2275490760803 
    weight[6] = 5.4394111633301 
    weight[7] = 5.3659925460815 
    weight[8] = 3.4170460700989 
    weight[9] = 5.2414722442627 
    weight[10] = 4.7376127243042    
    weight[11] = 5.2286224365234    
    weight[12] = 5.455126285553 
    weight[13] = 4.3019247055054    
    weight[14] = 5.4264230728149    
    weight[15] = 5.4331531524658    
    weight[16] = 5.433765411377 
    weight[17] = 5.4631009101868    
    weight[18] = 5.3947434425354
else:
    weight[0] = 2.8149201869965     #road
    weight[1] = 6.9850029945374     #sidewalk
    weight[2] = 3.7890393733978     #building
    weight[3] = 9.9428062438965     #wall
    weight[4] = 9.7702074050903     #fence
    weight[5] = 9.5110931396484     #pole
    weight[6] = 10.311357498169     #traffic light
    weight[7] = 10.026463508606     #traffic sign
    weight[8] = 4.6323022842407     #vegetation
    weight[9] = 9.5608062744141     #terrain
    weight[10] = 7.8698215484619    #sky
    weight[11] = 9.5168733596802    #person
    weight[12] = 10.373730659485    #rider
    weight[13] = 6.6616044044495    #car
    weight[14] = 10.260489463806    #truck
    weight[15] = 10.287888526917    #bus
    weight[16] = 10.289801597595    #train
    weight[17] = 10.405355453491    #motorcycle
    weight[18] = 10.138095855713    #bicycle

[sunpeng1996]

Oh,thanks o lot! I will try yours. The distribution is too large means that the first class weight is 0.0819 ,and someone is 5.2286, the 5.2286/0.0819 is too big,please look at my weights.

[Eromera]

Yes, that distribution is a bit weird. What code did you use for calculation? Your weights are basically making the model "ignore" the classes with very small values (0.0819) and boost the ones with very high values. Did you try the weights from the pytorch code?

[Eromera]

I'm closing this but if you have more questions just reopen it. Thanks!

[zldodo]

Hi Eromera, thanks a lot for your wonderful works. Recently I have been training the ERFnet from scratch with my data. I noticed that you used different class weights for encoder trainning and decoder trainning. Could you please explain the reason behind it ? Is this just because different datasets are used for them?