Setting up your image lab

Before you take a single photo, decide what each answer means. The program copies whatever pattern your labels contain, so a rule you apply one day and forget the next gives it two different targets to learn.

For this project:

• mug: a drinking container with a handle.
• glass: a drinking container without a handle. Plastic cups count as glass here.

These are rules for this project, not a claim about how everyone uses the words. A travel cup with a handle is a mug. A clear cup without a handle is a glass. Leave out three kinds of photo for now: photos with both objects in them, empty scenes, and containers whose handle you cannot see. The model can only ever answer mug or glass, so it has no way to say "both" or "neither".

Two words will keep coming up. A class is one of the allowed answers: mug or glass. A label is the answer you attach to one particular photo. If you label a glass photo as mug, training receives the wrong answer and has no way to ask what you meant.

Open Notes. The starting text already holds these two rules; reword them if you like. Then open Photos and check that Class 1 is mug and Class 2 is glass. Keep class names short and plain, with no percentages in them.

A model that has seen a photo during training has had the chance to learn that exact photo. Testing it on the same photo tells you very little. So before collecting, decide which objects will teach the model and which will be kept back to test it.

In Photos, the These photos are for menu gives each photo one of four jobs:

• Training: the only photos the model learns from.
• Development: photos kept out of training and used to check the model while you are still making changes. Many tools call this a validation set.
• Final test: photos you keep closed until you have picked your final version, then use once.
• Unlabelled targets: photos whose correct answer you do not know. The model can predict them, but nobody can score them.

The lab follows the menu exactly. If a photo of a reserved mug is added while Training is selected, the lab has no way to know it was meant to be kept back.

Split by physical object when you can: some mugs and glasses for training, different ones for development, and others again for the final test. All photos of one object stay in the same group. If you only own one mug and one glass, keep a separate session (a different room or background) for development, and write down that your test covers new views of familiar objects, not objects the model has never seen.

In Notes, list which objects (or which sessions) are for Training, which for Development, and which for Final test. Put the reserved objects to one side. Then check that These photos are for is set to Training.

With Training selected, press the mug button under the menu. Every photo you add now is stored as a training photo labelled mug.

Take photo opens the camera only when you ask for it. Check the frame, take one photo, and the camera closes. One photo at a time lets you look at each one before adding the next. Upload images does the same with JPG, PNG, or WebP files, such as photos copied from a phone. Either way, the lab cuts each picture to a centred square and shrinks it to 224 by 224 pixels, the size its image engine expects. Keep the object near the middle so the crop does not cut it off. The photos stay in this browser; they are not sent to a training server.

Variety matters more than quantity. Ten nearly identical photos teach almost the same thing as one. Between photos, change one of these:

A photo can be correctly labelled and still useless: a very dark or blurry frame shows no handle. Remove an accidental capture with the small x on its thumbnail.

Select Training and mug. Add photos with Take photo or Upload images, checking each thumbnail as it appears. For your own project, aim for about fifty varied photos per class. The lab lets you train with far fewer while you are learning the controls.

Press the glass button and collect glass photos the same way. The controls do not change; only the label attached to each new photo does. Check which button is highlighted every time you switch objects. The lab trusts the label you selected.

Now the important part. Suppose every mug sits on a wooden desk and every glass sits on a blue mat. Then "blue mat" predicts "glass" perfectly in your collection, and a model can score well by learning the mat instead of the object. This is called a shortcut: something that happens to line up with the label in your photos but is not the thing you care about. It fails the moment someone puts a mug on the mat. The fix is to give both classes the same range of surroundings: both on the desk, both on the mat, both held in a hand, both in dim light.

Keep the two counts roughly similar. Do not make up for having only a few different glasses by taking a hundred near-identical glass photos. Count different objects and conditions, not just files.

Select glass and collect a similar range of photos to your mugs. Open Train to see both counts side by side. If some background appears in only one class, add photos of the other class on that background.

You can learn the whole workflow without a camera. At the bottom of Photos, Use shape demonstration draws forty simple pictures in your browser: circles and squares in different sizes and positions, with the same colours and backgrounds in both classes. The lab gives them their jobs for you: 24 for Training, 8 for Development, and 8 for Final test.

The model you train on them is real, and so are its predictions. But the drawings are far simpler than photos, so a good score on them tells you nothing about mugs, glasses, or any real object.

Loading the demonstration replaces your current class names and photos. If you already have photos you want to keep, press Download project first (the next stage explains it). The lab asks before replacing anything: Replace with shapes or Keep my photos.

Only if you have no photos: press Use shape demonstration. Then switch These photos are for between Training, Development, and Final test to see how the drawings were shared out. If you are using your own photos, do not press the button; just read this stage.

The lab saves your photos and notes in this browser as you work, and the same project follows you through every Level 2 reading. Browser storage is limited, though, and clearing your browsing data wipes it. If the lab runs out of space, it shows a warning.

Download project, at the bottom of the lab, saves one file holding your photos, each photo's job, the class names, the rounds setting, and your notes. Open project loads such a file back. Think of it as the editable source of your experiment: with it, you can always add photos and train again.

A project file does not contain a trained model. You have not trained one yet; the next reading shows how to save the model separately.

In Notes, add a run name such as 'mug-glass run 1' and a line saying what you collected. Press Download project and keep the file somewhere you will find it again.