Test accuracy is very low in multitask net.

[baole...@gmail.com]

Hello everyone,

I am trying to train a multitask CNN nets which is used to both classification and regression.

When I trained it with 100000 iteration,the classification accuracy is still very low,and its loss is still high:

Test net output #0: accuracy = 0.001875

I1202 06:35:36.006667  7062 solver.cpp:408]     Test net output #1: class_loss = 3.09879 (* 4 = 12.3952 loss)

I1202 06:35:36.006677  7062 solver.cpp:408]     Test net output #2: reg_loss = 0.0493825 (* 1 = 0.0493825 loss)

I finetune my net with Caffenet. It takes about 10 hours for 100000 iters.

My solver is like this:

test_iter: 100

test_interval: 1000

base_lr: 0.0001

lr_policy: "step"

gamma: 0.1

stepsize: 20000

display: 20

max_iter: 200000

momentum: 0.9

weight_decay: 0.0005

Any advice for this problem?

Thanks,

Bill.

[Jobs Bill]

I just train it longer, now the training loss is about 5, and test loss is about 10 while test accuracy is about 0.05.
在 2015年12月2日星期三 UTC+8上午9:55:01，Jobs Bill写道：

[helxsz]

can you specify your dataset and possibly the training prototxt as well

[Jobs Bill]

Hi, helxsz:

     I have said I just finetune from caffenet, so my prototxt file is same as caffenet's.But I also add some layers, so that the whole net has two branches. 

The net's datalayer has four top: data1, data2, label1, label2(1 for classification and 2 for regression), the rest layer is in attach. Hope it will help.

Thanks,

Bill.

在 2015年12月5日星期六 UTC+8下午11:24:09，helxsz写道：