ResNet-example

A Residual Neural Network trained to classify objects from sound samples We loaded the data using the provided gcommend_loader, with a batch size of 20 and shuffle on the train. We chose to implement the network based on a Resnet network. In Order to do so, we crated a network with a convolution layer using an input size 1 and an output size 20, a kernel size 3, a stride of 1 and padding of 1. We performed a batch normalization on the convolutional layer, applied the relu activation function, and created 3 layers so that each layer will have 2 residual blocks. On two of the 3 layers we performed a convolution layer using filter of 3, and a stride of 2. Each Residual layer was performed with the following twice: convolutional layer, a batch normalization and a relu activation function. At the end, before the second relu was applied we combined the input of the layer with the output of the batch normalization. Also, we added avg pooling of size 8. At the end we added one fully connected layer with an input size of 900, that was calculated by the output of the avg pooling and an output of 30 as the number of class we have, and applied the softmax function on the output. We used 10 epochs with a learning rate of 0.01, we tried multiple options of values for both, but these gave us the best results. Also, we used nll loss and Adam optimizer since they gave us the best results. We managed to reach an accuracy of 91.1