Week #9
26 Dec 2018Done
-
Solve the image visualization problem existed last week.
-
Try to Understanding Categorical Cross-Entropy Loss, Binary Cross-Entropy Loss, Softmax Loss, Logistic Loss, Focal Loss. link
-
Try to use 3 concatenated random cropped image datasets as a whole training set, the total number of training number is 480, the source code is here. The experiments are as follows:
| Number | Data source | Training Sample | epoch of training | Batch size | learning rate | cropped size | momentum | Mimimum loss attained |
|---|---|---|---|---|---|---|---|---|
| 01 | KITTI | 480 | 20 | 1 | 0.001 | 128 | 0.9 | 0.651 |
| 02 | KITTI | 480 | 20 | 1 | 0.0001 | 128 | 0.9 | 0.672 |
| 03 | KITTI | 480 | 20 | 1 | 0.25 | 128 | 0.9 | 0.686 |
| 04 | KITTI | 480 | 20 | 1 | 0.25, every 5 epoch divide by 5 | 256 | 0.9 | 0.620 |
| 05 | KITTI | 480 | 20 | 1 | 0.5, every 5 epoch divide by 5 | 256 | 0.9 | 0.634 |
- Try to use other datasets to study the image segmentation(still need to fix the loss function problem):
- Review of cs231n blog, especially the loss function
To do
- Hyperparameter still need to be tuned
- Figure out why the loss doesn’t decrease
- Try to train the network using non-cropped image for the KITTI datasets
- More Experiments