Your first training run

Training is the step where the program adjusts the model to fit your labelled photos. People also call it fitting the model. It is the guess, check, adjust loop from Level 1: the model guesses a class for a training photo, a loss number measures how wrong the guess was, and the learned numbers inside the model are nudged to make the loss smaller.

The Training rounds menu sets how many times that happens across your whole collection. One round is one pass through every training photo. The usual technical name for a round is an epoch. Leave it at 20 for your first run, so later runs have the same setting to compare against.

Press Train model once. The first time, the lab downloads its image engine and the engine's starting numbers, so it needs an internet connection and may take a little while. Then it reads each training photo and runs the rounds. The status line at the bottom shows the round and the loss. The loss should mostly shrink as the rounds go by. It measures mistakes on the training photos only, so a small loss is not a score on new photos.

Everything runs in your browser; your photos are not uploaded anywhere. Pressing Train model again does not make it faster. To cancel, press Stop and wait a moment for the current step to finish.

In Train, check both counts, keep 20 rounds, and press Train model. Wait until the status says training is complete; the lab then switches to Test & export by itself. If an error appears instead, read it. The usual causes are a class with fewer than two training photos or a download that failed. Download your project before reloading the page, and never fix a failed run by moving reserved photos into Training.

A trained model is ready to answer questions. The fair question is one about photos it has never seen: your reserved development objects.

In Photos, switch These photos are for to Development. Press the correct label button for each reserved object before adding it. You decide the right answer with your rule, before the model has a say. Add two or three photos: a mug, a glass, and one unusual angle. If you loaded the shape demonstration, its development drawings are already there.

In Test & export, choose Development under Predict this collection and press Predict collection. Each development photo goes through the trained model, which gives it a score for each class. The class with the bigger score is the model's answer, its prediction. Because you supplied the right answer for each development photo, the lab can mark each prediction correct or incorrect. Predicting never changes the model: nothing is learned from these photos. Above the results the lab also shows an Accuracy line and a small table. With only two or three photos they mean very little, so leave them until Test it honestly.

Each result starts with an ID. It is the uploaded file's name without its ending (dev01.jpg becomes dev01), or a generated name for a camera photo. If two files share a name, the lab adds a number to the second so every ID stays different.

Add two or three development photos, predict the Development collection, and match one result to its photo. For that result, read the ID, the predicted class, the actual class, and both scores.

Suppose a result shows mug at 82% and glass at 18%. The two scores always add up to 100%. The model answered mug because mug got the bigger share. That is all the scores say: how the model split its preference between the two answers it knows. People call them confidence scores.

Confidence is not accuracy. 82% does not mean the model is right 82% of the time, and it does not even mean this answer is right: a model can give 95% to a wrong answer. How often a model is right can only be measured by checking many predictions against known answers, which the next module does.

Remember too that this model knows only two answers. Show it an empty table or a shoe and it still has to share 100% between mug and glass, so it may say mug at 90%. A photo outside the task like this is called out of scope. A high score on it shows a limit of the model; it is not evidence that the shoe is a mug.

Pick one correct result, and one wrong result if you have one. In Notes, copy each ID, the prediction, the two scores, and the actual answer. Label them as observations from a few photos, not a measurement.

Your trained model stays in this page while you move between Level 2 readings, but a refresh or a closed tab loses it. Your photos and notes survive, because the browser saves them; the training would have to be run again.

To keep the model itself, press Download model in Test & export. The file holds the learned numbers inside the model and its two class names. It does not hold your photos or notes. Open model loads it back, so you can predict again without retraining. When you open one, check that the class names it reports match your project.

So there are three kinds of file, each with its own job. The project is what you edit and retrain from. The model is what you predict with. A predictions file, which you will make in a later reading, is a record of answers for particular photos. None of them can stand in for another.

Press Download model in Test & export. Press Download project again too, since you have added development photos. Rename both files so they start with the same run name, such as run1, so you can tell later which model came from which photos.