When using Azure Form Recognizer together with its connected containers, The training of custom models occurs in the cloud using Azure Form Recognizer. You would send your training data to Azure, and Azure Form Recognizer would create a custom model. After training, you'll receive a model ID.
The container does not inherently have the model within it. Instead, when you run the container, you'll will provide configuration details, one of which is the model ID you received from the training step. When the container starts, it fetches the model from Azure and uses it locally. The model is stored in a local cache directory that you specify, so it doesn't have to fetch it every time the container starts.
If your container gets corrupted or if you need to restart it, the model will still exist in this cache directory, and you won't lose it. However, if this cache directory is deleted or corrupted, then you would lose the locally cached model. It's good practice to have a reliable storage solution for this cache directory. You provide a local cache directory using the eula=accept and model-cache= parameters when starting the container.
Training data does get sent to Azure during the training process, when you're using connected containers, the inference (or prediction) happens entirely on-premises. The container doesn't send the data you're analyzing back to Azure. It processes the data locally using the cached model.
