A custom ResNet architecture that uses One Cycle Policy to train the model

The model is trained on CIFAR-10 dataset using a custom Resnet architecture as shown below

Figure 1 : Resnet Architecture

The training is done using one cycle policy with following parameters

• Total Epochs = 24
• Max at Epoch = 5
• LRMIN = FIND
• LRMAX = FIND
• NO Annihilation

The following image augmentations are done using Albumentation Library

• RandomCrop 32, 32 (after padding of 4)
• FlipLR
• CutOut(8, 8)

The summary of the model is shown below

Figure 2 : Model summary

Result

The training logs, loss and accuracy can be found here

Figure 3 : Plot for loss and accuracy