PyTorch Tutorial : Forward propagation

Key Takeaways

welcome back here we are going to explain forward propagation also known as forward pass an important step on training and evaluation of all classifiers let's take a classifier maybe a neural network no worries if you don't know what it is I promise that you are going to soon learn about it and the principle is general for most classifiers the first step is to evaluate what classifier is doing given your data given some data the yellow layer the model does all the operation in the following blue layers until it gives some result in the output brown layer this step is called the forward step we are going to start with a simple example let's assume that we have a model it could be a neural network some other type of classifier or in more abstract terms a computational graph intuitively a computational graph is a network of nodes that represent numbers scalars or tensors and are connected via added that represent functions of operations for simplicity in visualizations instead of using large multi dimensional tensors we are going to use colors or graph has notes a B C and D and some operations between the nodes like adding not a with not B or multiplying no see with no D our job is to do all the operations until we get the result G first we add a and B getting 2 plus minus 4 equals 2 minus 2 putting the result in node E similarly we multiply C with D getting minus 2 times 2 equals 2 minus 4 and we put the result in node F now our nodes in the graph or nodes ENF containing values minus 2 and minus 4 the last step is to multiply these two values and to put the result in G we do that getting minus 2 times minus 4 equals 8 and so finishing our exercise let's implement it on fighters we first initialize dancers a b c and d to the corresponding values - 4 - 2 & 2 we then add a and B - 10 sorry and we multiply CD putting the result in tensor F finally we multiply e with F putting the result in tensor G and then we print the values of these three tenses getting the same result as in the slice now there is hardly anything fancy in this exercise however understanding it is of you to importance neural networks and most of the other classifiers can be understood as computational graphs in fact your code gets converted to a computational graph just that the graphs will be much larger and the tensors within them will contain millions of entries an additional benefit of computational graph is that they make the automatic computation of derivatives or gradients much easier as we will see in the next lesson let's do a quick exercise to solidify the knowledge about computational graphs and forward