States
In Decthings, a model state is a piece of binary data which gets loaded into the model when it is executed. This binary data allows the model to store configuration information - for example, a neural network model would want to store the weights and biases of the network. When evaluating or training the model, the state is automatically loaded into the model, which allows the model code to access and use the state as intended.
Training is the process of providing some input data and letting our model "learn" from that data. For a neural network, training would mean to update the weights and biases of the neurons in the network. When a training session is completed, a new state which contains the new data (for example the weights and biases) is created.
State format
Because a state contains binary information, it is up to the model to serialize and deserialize the state to and from a binary format. The model can store any information it likes in the state - it does not have to be weights and biases.
For optimization purposes, the binary information of a state can be split into multiple chunks. It is up to the model code to select how many chunks it wants to output as well as the size of each chunk.
Stored states, and the active state
A model can have multiple states, one of which will be marked as the active state. When you create a new state, the state will be added to the state storage, but it will not immediately be marked as the active state. To use the new state you can mark it as the active state. This will cause all future evaluations and training sessions to use this particular state.
When evaluating or training you can always override the active state and instead use some other state. Usually, this is not recommmended. The benefit of using the active state when evaluating is that the device will automatically use the latest state without needing reconfiguration.
If you delete a state that is used by some running operation, the state will be removed once all operations using the state are finished.
To view the active state and all stored states, go to the model page, go to the "Train" tab and scroll down to the "States" panel.
Create a new, untrained state
Similarly to the train operation, the Create state operation also allows you to create a new state. However, the two operations have different use cases. When you run the create state operation, the model will be executed and the "createModelState" function will be executed within the model code. This function takes a set of input parameters, and outputs a new state. However, the function will not have access to the previous state. For this reason, the create state operation is intended to create a new, fresh state, that later can be trained.
For example, for a neural network, we usually want to start with randomized weights and biases. In that case, we can code the "createModelState" function so that it outputs a state that contains those randomized weights and biases. It does not output a state that has been actually been trained.
We can take this one step further, and customize some input parameters to our create state operation. For example, we could have one input parameter which specifies the number of layers that we want in our neural network. The "createModelState" function would then be able to read this input parameter, and create a state that contains both the weights and biases, as well as information required to later reconstruct the neural network accordingly. Once we have a new, randomized state, we can train it. The train function will then use this randomized state as a starting point, and adjust the weights and biases to improve the model.
You can upload a state by clicking the "Create state" button on the "States" panel of the "Train" tab of the models page. You can also call the Model.createState API function to upload a state using an API client.
Upload a state
You can upload a state by clicking the "Upload state" button on the "States" panel of the "Train" tab of the models page. You can also call the Model.uploadState API function to upload a state using an API client. Uploading a state should be an uncommon thing to do. However, it can be useful if you for example want to copy a state from one model to another. Note that when uploading a state, you are able to provide any binary data. If the model code is not able to interpret the binary data, evaluations or training sessions that you run using that state may fail.
