Skip to content

[4 29, Ikjae] Experiments

Justin Kim edited this page Apr 30, 2021 · 1 revision

Models used so far

FCN8s, DeepLabv3 ResNet50, UNet ResNet34, DeepLabv3+ ResNeXt101 32x16d

Augmentation (Brightness)

Fixed parameters

  • contrast = hue = saturation = 0.0
  • model = DeepLabv3+ ResNeXt101 32x16d with instagram pretrain-weights
  • batch = 8, lr = 1e-5
  • input_size = 256x256
  • loss function = cross entropy
  • optimizer = adam
  • seed = 21
brightness epoch loss val_loss val_mIoU LB
0.0 7 0.2008 0.3413 0.4963 0.6149
0.2 7 0.1981 0.3422 0.4897 N/A
0.4 9 0.1757 0.3173 0.5133 0.6158
0.6 7 0.2161 0.3362 0.5218 0.6082
  • observations There are minor LB(less than 0.0010) score difference in between different values of brightness. However, it's not worth finding optimal parameter.
  • thought setting brightness = x, randomly applies brightness transformation to images what if I set up a parameter to have the same effect throughout all images?

failures

model with loss function, -log(DiceLoss), is hard to train

  • thought
    consider mixture of cross entropy and dice, with non-equal weights something like 0.7ce - 0.3log(dice)

Overall thoughts

  • so far all the parameter tuning barely affects the performance
  • the larger model appears to perform better, thus try efficientnet
  • unclear which validation measure to evaluate the performance of the model on unseen data