VGG type Network
- ConvBlock (32, 360, 360)
- ConvBlock (64, 180, 180)
- ConvBlock (128, 90, 90)
- ConvBlock (256, 45, 45)
- ConvBlock (512, 22, 22)
- Global Average Pooling
- Dense(2)
ConvBlock
- Conv3x3 + BN + ELU
- Conv3x3 + BN + ELU
- MaxPool (2, 2)
- Batch: 32
- Optimizer: Adam
- LR: 1e-4
- 5 fold
- Loss: L1
- Augmentation: flips
Training
A
B
C
D
E
F
BCE = - \sum_i \left( y_i \ln (p_i) + (1 - y_i) \ln(1 - p_i)\right)
BCE=−∑i(yiln(pi)+(1−yi)ln(1−pi))
DICE = 2 \frac {\sum_i y_i p_i} {\sum_u y_i + \sum p_i}
DICE=2∑uyi+∑pi∑iyipi
LOSS = BCE - \ln \left( DICE \right)
LOSS=BCE−ln(DICE)
BCE = - \sum_i \left( y_i \ln (\hat{y_i}) + (1 - y_i) \ln(1 - \hat{y_i})\right)
BCE=−∑i(yiln(yi^)+(1−yi)ln(1−yi^))
SoftJaccard = \frac{1}{n}\sum\limits_{i=1}^n\left(\frac{y_i\hat{y}_i}{y_{i}+\hat{y}_i-y_i\hat{y}_i}\right)
SoftJaccard=n1i=1∑n(yi+y^i−yiy^iyiy^i)
LOSS = \alpha \times BCE + (1 - \alpha) \times SoftJaccard
LOSS=α×BCE+(1−α)×SoftJaccard
CCE = - \frac {1} {n} \sum_{c=1}^7 \sum_{i=1}^{n} y_i^c \ln (\hat{y_i}^c)
CCE=−n1∑c=17∑i=1nyicln(yi^c)
SoftJaccard =\frac{1}{n}\sum_{c=1}^7w_c\sum\limits_{i=1}^n\left(\frac{y_i^c\hat{y}^c_i}{y_{i}^c+\hat{y}^c_i-y_i^c\hat{y}_i^c}\right)
SoftJaccard=n1∑c=17wci=1∑n(yic+y^ic−yicy^icyicy^ic)
LOSS = \alpha \times CCE + (1 - \alpha) \times SoftJaccard
LOSS=α×CCE+(1−α)×SoftJaccard
Jaccard | |
---|---|
Public Test | 0.636 |
Private Test | 0.613 |
Jaccard | |
---|---|
Public Test | 0.493 |
Private Test | 0.523 |
deck
By Vladimir Iglovikov
deck
- 1,362