Add note about randomization

docs/source/tutorials/crafting_standard_datasets.md

@@ -45,6 +45,6 @@ def default_light_curve_observation_post_injection_transform(x: LightCurveObserv
     return x
 ```
 
-It's a function that takes in a `LightCurveObservation` and spits out two `Tensor`s, one for the fluxes and one for the label to predict. Most of the data transform functions within have names that are largely descriptive, but we'll walk through them anyway. `remove_nan_flux_data_points_from_light_curve_observation` removes time steps from a `LightCurveObservation` where the flux is NaN. `randomly_roll_light_curve_observation` randomly rolls the light curve (a random cut is made and the two segments' order is swapped). `from_light_curve_observation_to_fluxes_array_and_label_array` extracts two NumPy arrays from a `LightCurveObservation`, one for the fluxes and one from the label (which in this case will be an array with a single value). `make_uniform_length` performs the uniform lengthening on the fluxes as we discussed in the previous section. `pair_array_to_tensor` converts the pair of NumPy arrays to a pair of PyTorch tensors (PyTorch's equivalent of an array). `normalize_tensor_by_modified_z_score` normalizes a tensor via based on the median absolute deviation. Notice, this is only applied to the flux tensor, not the label tensor. The `randomize` parameter enables or disables randomization of the functions which may include randomization. During training, randomization should be on to make sure we get variation in training observations. During evaluation and inference, it should be off to get repeatable results.
+It's a function that takes in a `LightCurveObservation` and spits out two `Tensor`s, one for the fluxes and one for the label to predict. Most of the data transform functions within have names that are largely descriptive, but we'll walk through them anyway. `remove_nan_flux_data_points_from_light_curve_observation` removes time steps from a `LightCurveObservation` where the flux is NaN. `randomly_roll_light_curve_observation` randomly rolls the light curve (a random cut is made and the two segments' order is swapped). `from_light_curve_observation_to_fluxes_array_and_label_array` extracts two NumPy arrays from a `LightCurveObservation`, one for the fluxes and one from the label (which in this case will be an array with a single value). `make_uniform_length` performs the uniform lengthening on the fluxes as we discussed in the previous section. `pair_array_to_tensor` converts the pair of NumPy arrays to a pair of PyTorch tensors (PyTorch's equivalent of an array). `normalize_tensor_by_modified_z_score` normalizes a tensor via based on the median absolute deviation. Notice, this is only applied to the flux tensor, not the label tensor. The `randomize` parameter enables or disables randomization of the functions which may include randomization. During training, randomization should be on to make sure we get variation in training observations. During evaluation and inference, it should be off to get repeatable results. In our previous example, to keep the code simple, we did not disable randomization for the validation dataset. Although in most cases it will not make a major difference, randomization should be disabled on the validation dataset. It should only be enabled for the training dataset.
 
 It's worth noting, `default_light_curve_observation_post_injection_transform` is just a function that can be replaced as desired. To remove one of the preprocessing steps or add in an addition one, we can simply make a modified version of this function. Additionally, `qusi` does not require the transform function to output only the fluxes and a binary label. The `Hadryss` NN model expects these two types of values for training, but other models may take advantage of the times of the light curve, or they may predict multi-class or regression labels.