Model¶

The third component of Emmental’s pipeline is to build deep learning model with your tasks.

Emmental Model¶

The following describes elements of used for model creation.

Emmental model.

class emmental.model.EmmentalModel(name=None, tasks=None)[source]¶

Bases: torch.nn.modules.module.Module

A class to build multi-task model.

Parameters

name (Optional[str]) – Name of the model, defaults to None.
tasks (Union[EmmentalTask, List[EmmentalTask], None]) – A task or a list of tasks.

add_task(task)[source]¶

Add a single task into MTL network.

Parameters: task (EmmentalTask) – A task to add.
Return type: None

add_tasks(tasks)[source]¶

Build the MTL network using all tasks.

Parameters: tasks (Union[EmmentalTask, List[EmmentalTask]]) – A task or a list of tasks.
Return type: None

collect_state_dict()[source]¶

Collect the state dict.

Return type: Dict[str, Any]

flow(X_dict, task_names)[source]¶

Forward based on input and task flow.

Note

We assume that all shared modules from all tasks are based on the same input.

Parameters

X_dict (Dict[str, Any]) – The input data
task_names (List[str]) – The task names that needs to forward.

Return type

Dict[str, Any]

Returns

The output of all forwarded modules

forward(uids, X_dict, Y_dict, task_to_label_dict, return_loss=True, return_probs=True, return_action_outputs=False)[source]¶

Forward function.

Parameters

uids (List[str]) – The uids of input data.
X_dict (Dict[str, Any]) – The input data.
Y_dict (Dict[str, Tensor]) – The output data.
task_to_label_dict (Dict[str, str]) – The task to label mapping.
return_loss – Whether return loss or not, defaults to True.
return_probs – Whether return probs or not, defaults to True.
return_action_outputs – Whether return action_outputs or not,
False. (defaults to) –

Return type

Union[Tuple[Dict[str, List[str]], Dict[str, Tensor], Dict[str, Union[ndarray, List[ndarray]]], Dict[str, Union[ndarray, List[ndarray]]], Dict[str, Dict[str, Union[ndarray, List]]]], Tuple[Dict[str, List[str]], Dict[str, Tensor], Dict[str, Union[ndarray, List[ndarray]]], Dict[str, Union[ndarray, List[ndarray]]]]]

Returns

The uids, loss, prob, gold, action_output (optional) in the batch of all tasks.

load(model_path, verbose=True)[source]¶

Load model state_dict from file and reinitialize the model weights.

Parameters

model_path (str) – Saved model path.
verbose (bool) – Whether log the info, defaults to True.

Return type

None

load_state_dict(state_dict)[source]¶

Load the state dict.

Parameters: state_dict (Dict[str, Any]) – The state dict to load.
Return type: None

predict(dataloader, return_loss=True, return_probs=True, return_preds=False, return_action_outputs=False)[source]¶

Predict from dataloader.

Parameters

dataloader (EmmentalDataLoader) – The dataloader to predict.
return_loss (bool) – Whether return loss or not, defaults to True.
return_probs (bool) – Whether return probs or not, defaults to True.
return_preds (bool) – Whether return predictions or not, defaults to False.
return_action_outputs (bool) – Whether return action_outputs or not,
False. (defaults to) –

Return type

Dict[str, Any]

Returns

The result dict.

remove_task(task_name)[source]¶

Remove a existing task from MTL network.

Parameters: task_name (str) – The task name to remove.
Return type: None

save(model_path, iteration=None, metric_dict=None, verbose=True)[source]¶

Save model.

Parameters

model_path (str) – Saved model path.
iteration (Union[float, int, None]) – The iteration of the model, defaults to None.
metric_dict (Optional[Dict[str, float]]) – The metric dict, defaults to None.
verbose (bool) – Whether log the info, defaults to True.

Return type

None

score(dataloaders, return_average=True)[source]¶

Score the data from dataloader.

Parameters

dataloaders (Union[EmmentalDataLoader, List[EmmentalDataLoader]]) – The dataloaders to score.
return_average (bool) – Whether to return average score.

Return type

Dict[str, float]

Returns

Score dict.

update_task(task)[source]¶

Update a existing task in MTL network.

Parameters: task (EmmentalTask) – A task to update.
Return type: None

Configuration Settings¶

Visit the Configuring Emmental page to see how to provide configuration parameters to Emmental via .emmental-config.yaml.

The model parameters are described below:

# Model configuration
model_config:
    model_path: # path to pretrained model
    device: 0 # -1 for cpu or gpu id (e.g., 0 for cuda:0)
    dataparallel: True # whether to use dataparallel or not
    distributed_backend: nccl # what distributed backend to use for DDP [nccl, gloo]