6. Training on Small Dataset

Github Link : https://github.com/satyajitghana/ProjektDepth/blob/master/notebooks/13_DepthModel_TrainOnSmallDataset.ipynb Colab Link : https://colab.research.google.com/github/satyajitghana/ProjektDepth/blob/master/notebooks/13_DepthModel_TrainOnSmallDataset.ipynb

Now the model was trained on 96x96 images, so then later we can perform transfer learning to train on 192x192 images

The entire training took about 5.5 hrs, but its actually took more, since there were bugs that i had to patch up and then restart the run.

Albeit the program does checkpoint every epoch and also stores the best accuracy model, every model metric is logged using Tensorboard.

Quite a tiring experience overall 😪😪

LR Range Test

GitHub Link : https://github.com/satyajitghana/ProjektDepth/blob/master/notebooks/12_DepthModel_LRRangeTest.ipynb Colab Link : https://colab.research.google.com/github/satyajitghana/ProjektDepth/blob/master/notebooks/12_DepthModel_LRRangeTest.ipynb

So the max_lr should be about 0.2-0.4

Colab was kept alive by using my chrome extension https://github.com/satyajitghana/colab-keepalive

Train for 15 Epochs

LR Value

Train Loss

Train - Segmentation Loss

Train Depth Loss

Train Accuracy

Train mIOU

Train mRMSE

Test Loss

Test - Segmentation Loss

Test Depth Loss

Test Accuracy

Test mIOU

Test mRMSE

Results

After 1st epoch

After 2nd epoch

After 4th epoch

After 9th epoch

After 11th epoch

After 15th epoch