eSigning CSV Made Easy

Esigning csv

[Music] [Applause] hi and welcome to coding tensorflow a show where we focus on coding machine learning and AI applications I'm Laurence Moroney a developer advocate for tensorflow and in this episode were going to look at using JavaScript for machine learning in the browser in the previous episode we looked at creating a very basic machine learning scenario in the browser you took data that had a linear relationship and built a basic model that could predict future values based on that relationship it might have looked a little strange in how you fed the data into the model when you were training it with one tensor of the x-values and another for the y-values that's because one of the core concepts that you need to learn as a tensorflow developer is all about that how to shape your data and how to get it ready for training this process is a major part of data science and today we'll look at a more complex program and how you can get your data ready for training a machine learn model with it so instead of a simple linear arrangement which you don't really need machine learning for let's consider a classification problem and this is when there are multiple items of data about a thing and then there's something about how they're related that determines the classification of that thing so for example an email might be a spam if it's from a particular sender or it contains particular keywords or pictures an animal might be a doggie if it has four paws and a cute wet nose now it's very difficult to program traditional if-then type code for these scenarios and this is why machine learning can be such a powerful tool so let's take a look at some public data and we'll use that to build a classification system we're going to use the well-known iris dataset which collects data points from a hundred and fifty different samples of flower taking petal length and width as well as sepal length and width these measurements are then associated with one of three types of iris by training a neural network with these measurements telling it the values and then the classes of flower the chorus to those values you could then build a neural network that can infer from a new measurement what type of flower they represent so let's take a look at the data and here it is and you might typically get data like this as a bunch of comma separated values as you can see each entry has five values the four measurements that I mentioned earlier and then a value zero one or two in the final column indicating the category of flower that the data represents consider the first four values to be your X's and the last one is your Y thus given a set of four X's you'd want to predict or classify the Y so now that you have the data you can use it to train a model to do that you'll use tensors for the training X's and tensors for the training wise in addition to that you can use some of your data to test your model so what you should do is you take your percentage of your data for training the model and then with the remainder compare the predicted value with their actual value and from there you can determine how well your model is behaving so let's take a look at the code that we use to prepare this data for training first of all we'll split the data into different arrays for each of the classes this for loop iterates through the iris classes length and it creates two arrays one for the data of that class and one for the values for that class if you then look at the data you'll see that there are three classes so we'll have three of each array the data by class will contain the four measurements and the targets by class will contain zero one or two based on the flower type once we've created these arrays we can now iterate through the data and sort the values into the array based on the target so the data for class 0 will get loaded into data by class 0 and the targets for class 0 likewise etc etc if I now log these arrays to the console I can view them in my developer tools the next step will be to convert these values into tensors with four sets of tensors an X for training annex for test a Y for training and a y4 test we do this according to the test split which is a parameter that we pass into the function in this demo I said it's a point two so that 80% of my data is used for training and 20% for testing the workhorse here is the convert to tensors function this takes the data the targets and the split and loads all this value into tensors splitting them into training and test sets respectively let's take a look at that next here is the convert to tensors function it calculates the number of test examples by rounding the sample size by the split and the number of training examples will just be the remainder it then creates a two dimensional tensor of the data as you can see here and a one heart encoding of the label data now one heart encoding is a way of helping a machine understand how your data is being classified so instead of the flowers being 0 1 or 2 what happens is you get an encoded array where instead of a flower for 0 you would get 1 0 0 in that array instead of 1 you get a 0 1 0 etc etc the idea is this that that this array will just map to your output neurons once you've done that the data will be sliced into the four arrays based on the size determined by the test split the last step is just having a nice clean linear set of tensors to feed into the training instead of the 2d one that you have right now this is achieved using TF concat along axis 0 let's take a look at the code for this and here's the code you can see I set the concat access to be 0 and then I'll return my set of 4 tensors where i'm concatenating them into a one dimensional tensor as an example if i log the x trains against the concatenated x trains you'll see the difference this has the reducing the overall complexity of the data being fed into the model it doesn't have to try and figure out multiple dimensions and this makes training quicker and more accurate congratulations you've now taken raw data and you've learned how to pre-process it into tensors that make for efficient training including how to one heart encode the output data this is a massive part of designing any machine learning system getting your data right in the next video we'll train a neural network with this data and we'll see how you can design that network and then how you can do classification given the terrain model you can find that right here on the tensorflow youtube channel so don't forget to hit that subscribe button right now [Music] [Applause] you