使用tensorflow進行FCN網絡訓練時出現loss值是負值情況1

簡單的FCN網絡出現結果如下：

epoch=0,i=54747 of 78989, loss=-624.140625

epoch=0,i=54748 of 78989, loss=-739.443359

epoch=0,i=54749 of 78989, loss=-603.046875

epoch=0,i=54750 of 78989, loss=-594.843750

epoch=0,i=54751 of 78989, loss=-509.031250

epoch=0,i=54752 of 78989, loss=-656.093750

epoch=0,i=54753 of 78989, loss=-725.562500

epoch=0,i=54754 of 78989, loss=-589.484375

epoch=0,i=54755 of 78989, loss=-691.789062

epoch=0,i=54756 of 78989, loss=-123.398438

epoch=0,i=54757 of 78989, loss=-561.562500

epoch=0,i=54758 of 78989, loss=-554.531250

epoch=0,i=54759 of 78989, loss=-557.578125

epoch=0,i=54760 of 78989, loss=-543.281250

epoch=0,i=54761 of 78989, loss=-592.968750

經過驗證，是softmax與sigmoid函數選擇不恰當，本人做的是兩分類，換成sigmoid函數計算loss之後，發現所有的loss值固定為一個數，在此之前deconv是沒有加入bias的，當最後加入bias訓練之後，得到的結果如下所示：

epoch=7,i=57416 of 78989, loss=-44566832.000000

epoch=7,i=57417 of 78989, loss=-27127590.000000

epoch=7,i=57418 of 78989, loss=-27127606.000000

epoch=7,i=57419 of 78989, loss=-33217624.000000

epoch=7,i=57420 of 78989, loss=-35709012.000000

epoch=7,i=57421 of 78989, loss=-45951332.000000

epoch=7,i=57422 of 78989, loss=-29065516.000000

去掉bias之後，結果如下

epoch=0,i=22875 of 78989, loss=798.504578

epoch=0,i=22876 of 78989, loss=798.504578

epoch=0,i=22877 of 78989, loss=798.504578

epoch=0,i=22878 of 78989, loss=798.504578

epoch=0,i=22879 of 78989, loss=798.504578

epoch=0,i=22880 of 78989, loss=798.504578

epoch=0,i=22881 of 78989, loss=798.504578

epoch=0,i=22882 of 78989, loss=798.504578

固定為一個值不動，經過重新試驗，我将fcn網絡的最後一層的softmax改為了sigmoid函數之後，結果如下:

epoch=6,i=693 of 78989, loss=798.504578

epoch=6,i=694 of 78989, loss=798.504578

epoch=6,i=695 of 78989, loss=798.504578

epoch=6,i=696 of 78989, loss=798.504578

epoch=6,i=697 of 78989, loss=798.504578

epoch=6,i=698 of 78989, loss=798.504578

還是固定為同一個值不變，可能陷入局部最優解，将學習率從0.0001調整為0.001之後，結果如下所示：

epoch=6,i=696 of 78989, loss=798.504578

epoch=6,i=697 of 78989, loss=798.504578

epoch=6,i=698 of 78989, loss=798.504578

繼續調整學習率到0.01，結果不變。即跟學習率無關。

繼續探索原因，後續會補上結果。