allrank/config_template.json

{
  "model":
  {
    "fc_model":
    {
      "sizes": <array of ints, dimmensions of hidden layers of a fully connected network applied to the input>,
      "input_norm": <boolean, whether to to apply LayerNorm to input>
      "activation": <Optional[str], name of the activation funtion to be used between hidden layers, as in torch.nn.modules.activation>
      "dropout": <float, dropout probability>
    },
    "transformer":
    {
      "N": <int, number of transformer blocks>
      "d_ff": <int, hidden dimmension>
      "h": <int, number of transformer heads>
      "positional_encoding": <Optional[dict] with two keys, strategy (either "fixed" or "learned") and max_indices, int>
      "dropout": <float, dropout probability
    },
    "post_model":
    {
      "output_activation": <Optional[str], name of the activation funtion to be used on the output, as in torch.nn.modules.activation>
      "d_output": <int, number of output neurons>
    }
  },
  "data":
  {
    "path": <str, path to dataset dir>,
    "validation_ds_role": <str, name of the validation dataset to be used, excluding .txt extension>,
    "num_workers": <int, num workers in the data loaders>
    "batch_size": <int, batch size>
    "slate_length": <int, slate length used in training, slates will be subsampled or padded to match that length>
  },
  "optimizer":
  {
    "name": <str, name of the optimizer as in torch.optim>
    "args": <dict, arguments of the optimizer>
  },
  "lr_scheduler":
  {
    "name": <str, name of the LR scheduler as in torch.optim.lr_scheduler>
    "args": <dict, arguments of the LR scheduler>
  },
  "training":
  {
    "epochs": <int, num epochs>
    "early_stopping_patience": <int, number of epochs for early stopping>
    "gradient_clipping_norm": <Optional[float]
  },
  "metrics": <list of metrics for evaluation of the form name_k where name is the name of the metric
              as defined in models/metrics.py and k is the rank used for evaluation>,
"loss":
  {
    "name": <str, name of the loss to be used, as defined in models/losses
    "args": <dict, arguments of the loss function, empty if no arguments are needed>
  },
"val_metric": <str, name of the metric to be monitored in early stopping>,
"detect_anomaly": <boolean, whether to use pytorch's detect anomaly feature>
"expected_metrics" :
  {
  "val": <dict, keys are metric names are values are their expected values,
          useful for seeing if changes to your code degrade model performance of previously trained model>
  },
 "click_model":  <this is an optional section, used when running 'rank_and_click' script.
                 it defines the implementation of the click model to use.
                 For the list of click_models see allrank.click_models package
    "name": <str, fully qualified name of a class to instantiate, e.g. `allrank.click_models.cascade_models.BaseCascadeModel`
    "args": <dict, arguments of the click_model, empty if no arguments are needed, can contain nested {"name":..., "args":...}
                   for nested object construction>
  }
}