frame

1. Inherit dataset class from task class.

2. (Optional) Define class attribute data_path, if not defined, default to data root + name of dataset class.

3. Define list_data:

   • Input:

     • self.path - the dataset path

   • Output:

     • return X1, X2, ..., Xn, Y1, Y2, ..., Ym, lists of inputs and outputs

       Typically return X, Y

   • Inout:

     • self.dict - universal recording (across experiments)

       Typically used for mapping between class name and class index

   • Effects:

     List all the data in self.path. Do any preprocessing on raw data. This function will be called in __init__ iff do split. The returned data type should be reasonable to store in csv files.

     Typically X, Y are paths to the actual data

4. Define read_data:

   • Input:

     • x1, x2, ..., xn, y1, y2, ..., ym - a tuple of input and output corresponding to the output of list_data.

   • Output:

     • x1, x2, ..., xn, y1, y2, ..., ym - data to pass to model and loss function

   • Effects:

     Defines how to read actual data from listed data. Do any online preprocessing. This function will be called in __getitem__. The returned data type should be torch.Tensor.

5. (Optional) Define n_inputs if it is not 1.

   n_inputs and n_outputs (inferred) must be consistent between data and model

6. Define any parameter to model as a class attribute.

7. Define model

   • __init__(self, params from data, **kwargs)
   • forward(self, x1, x2, ..., xn)

8. (Optional) Define loss function