Class All/ Constraining Methods

[xleonplayz]

Method: crosstraining(Optimizer)

This method is used to train the data in a random order and interval towards the end of the optimization function. It ensures that training subsequent activation matrices does not overly disrupt the datasets already trained.

Process Details

1. Initialization:

   • Under the given optimizer's section in train.json, load:
     • The target_state for each activation matrix.
     • The last training matrix final_ombc for each activation matrix.
   • From data.json, retrieve the value iteration_crosstrain_b.
   • Initialize an internal attribute current_tp (Current Training Parameters) by loading final_ombc.

2. Cross-Training Loop (Loop A):

   The number of iterations is set by iteration_crosstrain_a from data.json.

   • For Each Iteration:
     • Random Selection:
     • Randomly determine two values:
       • intern_iter: An internal iteration count.
       • selected_ap: A selected activation matrix from the available activation matrices in data.json.
     • Cross-Training Helper:
     • Call the crosshelper() method with:
       • The matrix and target_state from selected_ap loaded from train.json.
       • current_tp.
       • The optimizer.
       • The value iteration_crosstrain_b.

3. Completion Check:

   • At the end, verify whether each activation matrix (AP) has been trained at least once.
     • If not, repeat the loop until the value top_cross_iterations from data.json is exceeded.
     • If the maximum is exceeded without training all APs, log an error message indicating that not every AP was cross-trained.

---

Method: crosshelper

Process Details

1. Determine Iterations:

   • Check if iteration_crosstrain_a is greater than intern_iter.
     • If yes, set b_iterations to iteration_crosstrain_b.
     • If not, use intern_iter as b_iterations.

2. Training Loop (Loop B):

   • For b_iterations Iterations:
     • Circuit Creation:
     • Create a circuit with:
       • current_tp as the training matrix.
       • selected_ap as the activation matrix.
     • Measurement:
     • Measure the circuit.
     • Optimization:
     • Call start(optimizer, target_state).
     • Execute the optimize method.

---

Documentation and Logging

Rules for the Training Matrices

• Saving Optimized Training Matrices:
  • Under the respective optimizer and optimization matrix in train.json, save:
  • Each optimized training matrix.
• Cross-Training Summary:
  • At the end, save a summary in train.json under crosstrain_selection.
  • This shows the order in which cross-training occurred.
• Random Numbers:
  • Write the randomly generated numbers (intern_iter, selected_ap) into train.json.

Logging Rules

• Logging Optimizer Completion:
  • Log the completion of each optimizer.
• Logging Activation Matrix Details:
  • Log each activation matrix with:
  • Their iteration counts.
  • The number of entries made in train.json.
• Logging Random Values:
  • Log the randomly generated numbers used during cross-training.

---

[xleonplayz]

needed files https://github.com/LILY-QML/LLY-DML/wiki/Optimizer https://github.com/LILY-QML/LLY-DML/issues/19 https://github.com/LILY-QML/LLY-DML/issues/34