In this repo I will upload all problem statements I tackled using computer vison. I am exectuing these Notebooks on Google Colab, which is a fantastic platform offering free GPUs and TPUs. In these I am not using transfer learning as it would basically kill the idea of learning to design CNN architectures.
We use a Wide-ResNet and achieve 93.23% accuracy, beating the original ResNet56v1 by .2%. This is due to Dropout and the parallell convolutions in conv and identity modules.
My take on the classic MNIST dataset to classify handwritten digits, which achieves 99.6% accuracy.
There are a PyTorch and a Keras implementations in this repository. The Keras version hit 0.99614 on kaggle. It uses a slightly changed version of LeNet5 replacing 5x5 Filter with two 3x3 for non linerarity.
The PyTorch implementation, which uses a large number of epochs, checks for overfitting and will abort on its own. This version was programmed only on a CPU, hence the small network.