Skip to content

Workflow

Training

When your dataset is ready, train a model without leaving the app. Pick a task, a model, and hyperparameters, then run it on your own GPU for full privacy — or push the job to Google Colab when you need more compute.

The training dialog

Click Train to open the training dialog. Everything for a run lives on one screen: the dataset and its split on the left, and the task, model, and hyperparameters on the right.

VisoLabel 'Train a new model' dialog. Left: dataset summary with image and annotation counts and train/validation/test split sliders. Right: task type selector, RF-DETR model cards, and hyperparameter fields. Bottom: Train locally and Train in Colab buttons.
The training dialog — dataset and split on the left, task, model, and hyperparameters on the right.

Dataset & split

Choose which dataset to train on — the Current Project or any augmented dataset you've created. VisoLabel shows the image and annotation counts so you know exactly what's going in.

The dataset split sliders divide your data into:

  • Train — the images the model learns from (e.g. 80%).
  • Validation — held back to measure progress during training (e.g. 10%).
  • Test — held back for a final, unbiased evaluation.

Task type

Match the task to what you annotated:

  • Object Detection — bounding boxes.
  • Instance Segmentation — pixel-level masks.
  • Classification — whole-image labels.

Choosing a model

For detection, VisoLabel trains RF-DETR models — a modern, real-time detection transformer. For segmentation projects, the RF-DETR-Seg variants are used, and classification projects train a ConvNeXt model. For RF-DETR, pick a size based on the speed/accuracy trade-off you want:

Model Speed Accuracy Best for
RF-DETR-N (Nano) Fast Good Edge devices, real-time, quick experiments
RF-DETR-S (Small) Medium Better A solid general-purpose default
RF-DETR-M (Medium) Medium Best Higher accuracy when you have the GPU for it
RF-DETR-L (Large) Slow Best Maximum accuracy, offline batch inference

Start with RF-DETR-N for your first run — it trains fastest, so you get feedback quickly.

Hyperparameters

The defaults are sensible; tune these as you iterate:

Setting What it does Starting point
Epochs Full passes over the dataset. 50
Batch size Images per step — higher uses more VRAM. 8
Image size Pixels per side. Must be a multiple of 24. 560
Learning rate How fast the weights update. 0.0001

If training runs out of GPU memory, lower the batch size or the image size (keeping it a multiple of 24) and try again.

Train locally or on Colab

When everything's set, choose where the run happens:

  • Train locally — runs on your own GPU (or CPU if you have no GPU — slower, but it works). Your data never leaves the machine, which is ideal for sensitive or proprietary datasets.
  • Train in Colab — VisoLabel packages your dataset and a ready-to-run notebook, connects to your Google account, and trains on Colab's GPU. The trained weights are pulled back into your project automatically. See the step-by-step walkthrough below.

An NVIDIA GPU with 8 GB+ VRAM is recommended for local training. No GPU? Use Train in Colab — see the system requirements.

Training in Colab, step by step

No GPU, or want to keep your machine free? Train in Colab runs the whole job on Google Colab's GPU. Your dataset is encrypted before it leaves your machine and decrypted only inside the notebook, and checkpoints are uploaded back into VisoLabel as they're produced.

1. Click "Train in Colab". In the training dialog, set up your run as usual, then click Train in Colab instead of Train locally.

The training dialog footer with the blue 'Train in Colab' button next to 'Train locally'.
Choose Train in Colab in the training dialog.

2. Wait for the dataset to upload. VisoLabel encrypts and packages your dataset and pushes it to Colab. Keep the window open until it finishes and the Colab token is ready.

'Upload dataset to Colab' dialog showing an 'Uploading dataset… 6%' progress bar and a Cancel button.
The dataset is encrypted and uploaded to Colab.

3. Open the Colab notebook. Once the upload completes, click Open Colab notebook to launch the ready-to-run trainer in your browser.

The 'Open Colab notebook' button shown after the dataset upload completes.
Launch the trainer with Open Colab notebook.

4. Open the VisoLabel Colab Trainer cell. In the Colab tab, expand the VisoLabel Colab Trainer section. The notebook downloads the encrypted bundle, decrypts it locally in Colab, validates the dataset, trains the selected model, and uploads checkpoints back to VisoLabel.

Google Colab notebook showing the 'VisoLabel Colab Trainer' section and a cell instructing to paste the token and click Run full pipeline.
The VisoLabel Colab Trainer notebook, ready to run.

5. Paste the token and run the pipeline. Back in VisoLabel, click Copy token, paste it into the Colab token field, then click ▶ Run full pipeline. This resolves the bundle, starts training, and uploads results in one click.

Colab notebook form with the 'COLAB TOKEN' field filled and a row of buttons: Run full pipeline, Resolve bundle, Start training, Upload results, Clear logs.
Paste the token, then click Run full pipeline.

6. Watch the generated files. Training now runs on Colab. As checkpoints and logs are produced, they appear in the Generated files list inside VisoLabel — it refreshes every 15 seconds. Select what you need and click Download selected to pull it back into your project.

VisoLabel panel with the Colab token, a 'Copy token' button, and a 'Generated files' list that refreshes every 15 seconds with a 'Download selected' button.
Files produced by Colab show up under Generated files.

The token is what links the notebook to your run — copy it from VisoLabel and keep it private. Leave the Colab tab open while training runs so it can keep uploading checkpoints.

After training

When a run finishes you get a trained model attached to your project. From here you can run inference on new images or move on to export and deploy.


Next: Export & deploy — take your model live →