bootstrap.Engine

class bootstrap.engines.engine.Engine

Contains training and evaluation procedures

eval()

Launch evaluation procedures

eval_epoch(model, dataset, epoch, mode='eval', logs_json=True)

Launch evaluation procedures for one epoch

List of the hooks (mode='eval' by default):

• mode_on_start_epoch: before the evaluation procedure for an epoch
• mode_on_start_batch: before the evaluation precedure for a batch
• mode_on_forward: after the forward of the model
• mode_on_print: after the print to the terminal
• mode_on_end_batch: end of the evaluation procedure for a batch
• mode_on_end_epoch: before saving the logs in logs.json
• mode_on_flush: end of the evaluation procedure for an epoch

Returns:mean of all the scalar outputs of the model, indexed by output name, for this epoch Return type:out(dict)

generate_view()

Generate a view.html via an asynchronous call to self.view.generate()

hook(name)

Run all the callback functions that have been registered for a hook.

Parameters:name – the name of the hook

is_best(out, saving_criteria)

Verify if the last model is the best for a specific saving criteria

Parameters:

• out (dict) – mean of all the scalar outputs of model indexed by output name
• saving_criteria (str) –

Returns:

is_best(bool)

Example usage:

out = {
    'loss': 0.2,
    'acctop1': 87.02
}

engine.is_best(out, 'loss:min')

load(dir_logs, name, model, optimizer)

Load a checkpoint

Parameters:

• dir_logs – directory of the checkpoint
• name – name of the checkpoint
• model – model associated to the checkpoint
• optimizer – optimizer associated to the checkpoint

load_state_dict(state)

register_hook(name, func)

Register a callback function to be triggered when the hook is called.

Parameters:

• name – the name of the hook
• func – the callback function (no argument)

Example usage:

def func():
    print('hooked!')

engine.register_hook('train_on_start_batch', func)

resume()

Resume a checkpoint using the bootstrap.lib.options.Options

save(dir_logs, name, model, optimizer)

Save a checkpoint

Parameters:

• dir_logs – directory of the checkpoint
• name – name of the checkpoint
• model – model associated to the checkpoint
• optimizer – optimizer associated to the checkpoint

state_dict()

train()

Launch training procedures

List of the hooks:

• train_on_start: before the full training procedure

train_epoch(model, dataset, optimizer, epoch, mode='train')

Launch training procedures for one epoch

List of the hooks:

• train_on_start_epoch: before the training procedure for an epoch
• train_on_start_batch: before the training precedure for a batch
• train_on_forward: after the forward of the model
• train_on_bachward: after the backward of the loss
• train_on_update: after the optimization step
• train_on_print: after the print to the terminal
• train_on_end_batch: end of the training procedure for a batch
• train_on_end_epoch: before saving the logs in logs.json
• train_on_flush: end of the training procedure for an epoch

class bootstrap.engines.logger.LoggerEngine

LoggerEngine is similar to Engine. The only difference is a more powerful is_best method. It is able to look into the logger dictionary that contains the list of all the logged variables indexed by name.

Example usage:

out = {
    'loss': 0.2,
    'acctop1': 87.02
}
engine.is_best(out, 'loss:min')

# Logger().values['eval_epoch.recall_at_1'] contains a list
# of all the recall at 1 values for each evaluation epoch
engine.is_best(out, 'eval_epoch.recall_at_1')