def predict(self, pipeline_config, network, predict_loader):
if pipeline_config["torch_num_threads"] > 0:
torch.set_num_threads(pipeline_config["torch_num_threads"])
device = Trainer.get_device(pipeline_config)
Y = predict(network, predict_loader, device)
return {'Y': Y.detach().cpu().numpy()}
def predict(self, pipeline_config, network, predict_loader):
"""Predict using trained neural network
Arguments:
pipeline_config {dict} -- The user specified configuration of the pipeline
network {BaseNet} -- The trained neural network.
predict_loader {DataLoader} -- The data to predict the labels for.
Returns:
dict -- The predicted labels in a dict.
"""
if pipeline_config["torch_num_threads"] > 0:
torch.set_num_threads(pipeline_config["torch_num_threads"])
device = Trainer.get_device(pipeline_config)
Y = predict(network, predict_loader, device)
return {'Y': Y.detach().cpu().numpy()}
def fit(self, hyperparameter_config, pipeline_config, train_loader,
valid_loader, network, optimizer, optimize_metric,
additional_metrics, log_functions, budget, loss_function,
training_techniques, fit_start_time, refit,
hyperparameter_config_id):
"""Train the network.
Arguments:
hyperparameter_config {dict} -- The sampled hyperparameter config.
pipeline_config {dict} -- The user specified configuration of the pipeline
train_loader {DataLoader} -- Data for training.
valid_loader {DataLoader} -- Data for validation.
network {BaseNet} -- The neural network to be trained.
optimizer {AutoNetOptimizerBase} -- The selected optimizer.
optimize_metric {AutoNetMetric} -- The selected metric to optimize
additional_metrics {list} -- List of metrics, that should be logged
log_functions {list} -- List of AutoNetLofFunctions that can log additional stuff like test performance
budget {float} -- The budget for training
loss_function {_Loss} -- The selected PyTorch loss module
training_techniques {list} -- List of objects inheriting from BaseTrainingTechnique.
fit_start_time {float} -- Start time of fit
refit {bool} -- Whether training for refit or not.
Returns:
dict -- loss and info reported to bohb
"""
self.hyperparameter_config_id = hyperparameter_config_id
self.pipeline_config = pipeline_config
self.budget = budget
hyperparameter_config = ConfigWrapper(self.get_name(),
hyperparameter_config)
logger = logging.getLogger('autonet')
logger.debug("Start train. Budget: " + str(budget))
if pipeline_config["torch_num_threads"] > 0:
torch.set_num_threads(pipeline_config["torch_num_threads"])
trainer = Trainer(
model=network,
loss_computation=self.batch_loss_computation_techniques[
hyperparameter_config["batch_loss_computation_technique"]](),
metrics=[optimize_metric] + additional_metrics,
log_functions=log_functions,
criterion=loss_function,
budget=budget,
optimizer=optimizer,
training_techniques=training_techniques,
device=Trainer.get_device(pipeline_config),
logger=logger,
full_eval_each_epoch=pipeline_config["full_eval_each_epoch"])
trainer.prepare(pipeline_config, hyperparameter_config, fit_start_time)
model_params = self.count_parameters(network)
logs = trainer.model.logs
epoch = trainer.model.epochs_trained
training_start_time = time.time()
print("RESULT: ", pipeline_config['result_logger_dir'])
#Instantiate the SummaryWriter class if tensrobaord is activated
if 'use_tensorboard_logger' in pipeline_config and pipeline_config[
'use_tensorboard_logger']:
writer = SummaryWriter()
while True:
# prepare epoch
log = dict()
trainer.on_epoch_start(log=log, epoch=epoch)
# training
optimize_metric_results, train_loss, stop_training = trainer.train(
epoch + 1, train_loader)
if valid_loader is not None and trainer.eval_valid_each_epoch:
valid_metric_results = trainer.evaluate(valid_loader)
# evaluate
log['loss'] = train_loss
log['model_parameters'] = model_params
for i, metric in enumerate(trainer.metrics):
log['train_' + metric.name] = optimize_metric_results[i]
if valid_loader is not None and trainer.eval_valid_each_epoch:
log['val_' + metric.name] = valid_metric_results[i]
if trainer.eval_additional_logs_each_epoch:
for additional_log in trainer.log_functions:
log[additional_log.name] = additional_log(
trainer.model, epoch)
# wrap up epoch
stop_training = trainer.on_epoch_end(log=log,
epoch=epoch) or stop_training
# handle logs
logs.append(log)
log = {
key: value
for key, value in log.items()
if not isinstance(value, np.ndarray)
}
logger.debug("Epoch: " + str(epoch) + " : " + str(log))
# if 'use_tensorboard_logger' in pipeline_config and pipeline_config['use_tensorboard_logger']:
# self.tensorboard_log(budget=budget, epoch=epoch, log=log, logdir=pipeline_config["result_logger_dir"])
if 'use_tensorboard_logger' in pipeline_config and pipeline_config[
'use_tensorboard_logger']:
self.tensorboard_log(
writer,
budget=budget,
epoch=epoch,
log=log,
logdir=pipeline_config["result_logger_dir"])
if stop_training:
break
epoch += 1
torch.cuda.empty_cache()
# wrap up
loss, final_log = self.wrap_up_training(
trainer=trainer,
logs=logs,
epoch=epoch,
train_loader=train_loader,
valid_loader=valid_loader,
budget=budget,
training_start_time=training_start_time,
fit_start_time=fit_start_time,
best_over_epochs=pipeline_config['best_over_epochs'],
refit=refit,
logger=logger)
return {'loss': loss, 'info': final_log}
def fit(self, hyperparameter_config, pipeline_config,
train_loader, valid_loader,
network, optimizer,
train_metric, additional_metrics,
log_functions,
budget,
loss_function,
training_techniques,
fit_start_time,
refit):
hyperparameter_config = ConfigWrapper(self.get_name(), hyperparameter_config)
logger = logging.getLogger('autonet')
logger.debug("Start train. Budget: " + str(budget))
if pipeline_config["torch_num_threads"] > 0:
torch.set_num_threads(pipeline_config["torch_num_threads"])
trainer = Trainer(
model=network,
loss_computation=self.batch_loss_computation_techniques[hyperparameter_config["batch_loss_computation_technique"]](),
metrics=[train_metric] + additional_metrics,
log_functions=log_functions,
criterion=loss_function,
budget=budget,
optimizer=optimizer,
training_techniques=training_techniques,
device=Trainer.get_device(pipeline_config),
logger=logger,
full_eval_each_epoch=pipeline_config["full_eval_each_epoch"])
trainer.prepare(pipeline_config, hyperparameter_config, fit_start_time)
logs = trainer.model.logs
epoch = trainer.model.epochs_trained
training_start_time = time.time()
while True:
# prepare epoch
log = dict()
trainer.on_epoch_start(log=log, epoch=epoch)
# training
train_metric_results, train_loss, stop_training = trainer.train(epoch + 1, train_loader)
if valid_loader is not None and trainer.eval_valid_each_epoch:
valid_metric_results = trainer.evaluate(valid_loader)
# evaluate
log['loss'] = train_loss
for i, metric in enumerate(trainer.metrics):
log['train_' + metric.__name__] = train_metric_results[i]
if valid_loader is not None and trainer.eval_valid_each_epoch:
log['val_' + metric.__name__] = valid_metric_results[i]
if trainer.eval_additional_logs_each_epoch:
for additional_log in trainer.log_functions:
log[additional_log.__name__] = additional_log(trainer.model, epoch)
# wrap up epoch
stop_training = trainer.on_epoch_end(log=log, epoch=epoch) or stop_training
# handle logs
logs.append(log)
log = {key: value for key, value in log.items() if not isinstance(value, np.ndarray)}
logger.debug("Epoch: " + str(epoch) + " : " + str(log))
if 'use_tensorboard_logger' in pipeline_config and pipeline_config['use_tensorboard_logger']:
self.tensorboard_log(budget=budget, epoch=epoch, log=log)
if stop_training:
break
epoch += 1
torch.cuda.empty_cache()
# wrap up
loss, final_log = self.wrap_up_training(trainer=trainer, logs=logs, epoch=epoch, minimize=pipeline_config['minimize'],
train_loader=train_loader, valid_loader=valid_loader, budget=budget, training_start_time=training_start_time, fit_start_time=fit_start_time,
best_over_epochs=pipeline_config['best_over_epochs'], refit=refit, logger=logger)
return {'loss': loss, 'info': final_log}