Bring your Level 2 model in

Before reusing a model, carry over the meaning of its labels. The example task distinguishes mugs with handles from glasses without handles. If an object could fit either class, write down how you will label it, so it is not labelled differently on different days.

The words printed here are a reminder for the person collecting examples. They do not train the model. Training still needs actual labelled images.

Create cups, then use two say statements to write the intended classes. Choose objects that fit your definitions.

The collection is the set of labelled photos. You can inspect it, correct a label, or add photos. A trained model is different: it holds learned numbers used for prediction, not photos.

This stack collects a small set of photos inside one run. Use several viewpoints and similar backgrounds for each class. Photos from another page or an earlier run are not available here.

Connect the mug and glass capture statements after one make-image-model block for cups. Inspect the groups before adding training.

Training learns from the collection as it is at that moment. If you correct labels or add examples, you must train again before the model includes that change. Editing the collection does not change a model that was already trained.

Give experiment versions descriptive notes: which objects, how many views, and what changed. A score is hard to interpret if you cannot identify the model and collection that produced it.

Put train image model cups after collection. Build the rest of the prediction stages before running if you want to avoid repeating collection.

The prediction photo arrives without a mug/glass answer attached. The model must choose using its learned parameters. Capturing this photo is not a training update, and the program does not automatically know whether you agree with its result.

Store the photo once and ask for both outputs from that stored input. Reusing the value keeps the label and score aligned.

Set photo from the camera after training, then set move and sure using cups and that photo variable. Print both.

The repeat block surrounds prediction only. Collection and training stay above it. Both test photos therefore use the same trained model, so their results can be compared.

Reuse is not continued learning. The second prediction is not automatically better because the model saw the first test photo. No add-photos or train operation occurred between them.

Wrap only capture, label, score, and output in repeat 2. Keep setup and train outside.

A classifier returns evidence; your application decides what to show. This gate asks for a clearer object when the score is low. It does not add the new view to training or promise the next score will improve.

A two-class model can still label unrelated objects confidently. If the real application needs an answer for neither class, it needs photos and testing for that case; a threshold does not create a new class.

Place the gate inside the prediction loop after sure is stored. Keep the ask-again message inside the low-score branch. In this version an unclear photo still uses up one of the attempts.

Numbered attempts help you connect observations to your experiment notes. Leave training fixed while collecting this small evaluation. Changing collection halfway through would mix results from different models.

After reviewing the three outcomes, you may start a new development cycle with corrected or more varied examples. Starting that new cycle is different from quietly replacing a poor score in the old report.

Set attempt to 0 before repeat. Add 1 to it at the start of each turn, then capture and print the model result.

This complete program reuses one trained model for three predictions within one run. Downloading its code keeps the instructions, but not the camera photos or the trained model. The next stages save the trained model itself as a file.

Keep a model file together with a note of its class names. Keep the photos separately if you want to improve the model later. Knowing which file you have avoids a common mistake: treating a program or a project file as if it were a ready-to-use trained model.

Build the complete sequence, run it, and identify which statements train versus predict. Use the final reference to compare structure only after making your own version.

A saved model file contains the learned numbers (the weights), the class names, and the settings needed for prediction. It is different from the program and from a collection of photos. This stage trains the model, then writes it to a file named image-model.codeguide.json. JSON is a common text format for storing data.

The editor offers the file for download. Keep the supplied filename if you want the next stage to load it without changing a path. You can also use a compatible model downloaded from the Level 2 image lab.

After collection and training, connect save image model cups to file image-model.codeguide.json. Run the complete stage and choose the resulting model download in Output.

Loading rebuilds cups from the numbers saved in the file. There are no collection or train blocks in this stack. The only camera capture is a new prediction input. This is reuse across runs, rather than only reuse inside a repeat loop.

A file must actually be present before its path can be loaded. The loader accepts a model file saved by this course, not a photograph or a program renamed to end in .json. It checks the file's format before using it.

First save a model with the previous stage, then upload that downloaded JSON through this editor's Files tab. Check its name is image-model.codeguide.json. Load this stage and Run; if you have no saved model yet, the previous stage creates one entirely here.

The application loop can use a loaded model exactly as it used the freshly trained one. Its job is to provide inputs and interpret outputs; it does not need to know whether the model was trained earlier in the run or loaded from a file.

Keep the model file and the class definitions together. If a different file contains different classes, the messages and assumptions around prediction may need to change. A compatible file format does not prove that a model performs your intended task.

Keep the saved model in Files. Put only prediction and the confidence gate inside repeat 3, with load above it.

The full reference includes collection, training, saving, loading, and prediction, so it can be rebuilt from an empty editor. Saving writes the trained model to the file, and loading replaces cups with the copy read back from that file. The predictions that follow use the reloaded copy.

For everyday reuse, keep the model file and run the shorter load-only stage. Keep an editable project or collection separately if you intend to improve the training data. This completes the four pieces: program, labelled collection, trained model file, and new input.

Build the round-trip stack, run it, and download the model file. Then use the preceding load-only stage to confirm that another prediction run needs only the model file and a new photo, rather than collecting the training groups again.