def generate_and_print_model_summary(config: ModelConfigBase, model: DeviceAwareModule) -> None: """ Writes a human readable summary of the present model to logging.info, and logs the number of trainable parameters to AzureML. :param config: The configuration for the model. :param model: The instantiated Pytorch model. """ random_state = RandomStateSnapshot.snapshot_random_state() # There appears to be a bug in apex, where previous use (in training for example) causes problems # when another model is later built on the CPU (for example, before loading from a checkpoint) # https://github.com/NVIDIA/apex/issues/694 # Hence, move the model to the GPU before doing model summary. if config.use_gpu: model = model.cuda() if isinstance(config, ScalarModelBase): # To generate the model summary, read the first item of the dataset. Then use the model's own # get_model_input function to convert the dataset item to input tensors, and feed them through the model. train_dataset = config.get_torch_dataset_for_inference(ModelExecutionMode.TRAIN) train_item_0 = next(iter(train_dataset.as_data_loader(shuffle=False, batch_size=1, num_dataload_workers=0))) model_inputs = get_scalar_model_inputs_and_labels(config, model, train_item_0).model_inputs # The model inputs may already be converted to float16, assuming that we would do mixed precision. # However, the model is not yet converted to float16 when this function is called, hence convert back to float32 summary = ModelSummary(model) summary.generate_summary(input_tensors=model_inputs, log_summaries_to_files=config.log_summaries_to_files) elif config.is_segmentation_model: summary_for_segmentation_models(config, model) assert model.summarizer summary = model.summarizer # type: ignore else: raise ValueError("Don't know how to generate a summary for this type of model?") RUN_CONTEXT.log(LoggingColumns.NumTrainableParameters, summary.n_trainable_params) random_state.restore_random_state()
def log_to_azure(self, label: str, metric: float) -> None: """ Logs a metric as a key/value pair to AzureML. """ if not is_offline_run_context(RUN_CONTEXT): metric_name = self.logging_prefix + label RUN_CONTEXT.log(metric_name, metric) # When running in a cross validation setting, log all metrics to the hyperdrive parent run too, # so that we can easily overlay graphs across runs. if self.log_to_parent_run and PARENT_RUN_CONTEXT: if self.cross_validation_split_index > DEFAULT_CROSS_VALIDATION_SPLIT_INDEX: PARENT_RUN_CONTEXT.log(f"{metric_name}_Split{self.cross_validation_split_index}", metric)
def run(self) -> None: """ Driver function to run a ML experiment. If an offline cross validation run is requested, then this function is recursively called for each cross validation split. """ if self.is_offline_cross_val_parent_run(): if self.model_config.is_segmentation_model: raise NotImplementedError( "Offline cross validation is only supported for classification models." ) self.spawn_offline_cross_val_classification_child_runs() return # Get the AzureML context in which the script is running if not self.model_config.is_offline_run and PARENT_RUN_CONTEXT is not None: logging.info("Setting tags from parent run.") self.set_run_tags_from_parent() self.save_build_info_for_dotnet_consumers() # Set data loader start method self.set_multiprocessing_start_method() # configure recovery container if provided checkpoint_handler = CheckpointHandler(model_config=self.model_config, azure_config=self.azure_config, project_root=self.project_root, run_context=RUN_CONTEXT) checkpoint_handler.discover_and_download_checkpoints_from_previous_runs( ) # do training and inference, unless the "only register" switch is set (which requires a run_recovery # to be valid). if not self.azure_config.register_model_only_for_epoch: # Set local_dataset to the mounted path specified in azure_runner.py, if any, or download it if that fails # and config.local_dataset was not already set. self.model_config.local_dataset = self.mount_or_download_dataset() self.model_config.write_args_file() logging.info(str(self.model_config)) # Ensure that training runs are fully reproducible - setting random seeds alone is not enough! make_pytorch_reproducible() # Check for existing dataset.csv file in the correct locations. Skip that if a dataset has already been # loaded (typically only during tests) if self.model_config.dataset_data_frame is None: assert self.model_config.local_dataset is not None ml_util.validate_dataset_paths(self.model_config.local_dataset) # train a new model if required if self.azure_config.train: with logging_section("Model training"): model_train(self.model_config, checkpoint_handler) else: self.model_config.write_dataset_files() self.create_activation_maps() # log the number of epochs used for model training RUN_CONTEXT.log(name="Train epochs", value=self.model_config.num_epochs) # We specify the ModelProcessing as DEFAULT here even if the run_recovery points to an ensemble run, because # the current run is a single one. See the documentation of ModelProcessing for more details. best_epoch = self.run_inference_and_register_model( checkpoint_handler, ModelProcessing.DEFAULT) # Generate report if best_epoch: Runner.generate_report(self.model_config, best_epoch, ModelProcessing.DEFAULT) elif self.model_config.is_scalar_model and len( self.model_config.get_test_epochs()) == 1: # We don't register scalar models but still want to create a report if we have run inference. Runner.generate_report(self.model_config, self.model_config.get_test_epochs()[0], ModelProcessing.DEFAULT)
def model_train(config: ModelConfigBase, checkpoint_handler: CheckpointHandler) -> ModelTrainingResults: """ The main training loop. It creates the model, dataset, optimizer_type, and criterion, then proceeds to train the model. If a checkpoint was specified, then it loads the checkpoint before resuming training. :param config: The arguments which specify all required information. :param checkpoint_handler: Checkpoint handler object to find checkpoint paths for model initialization :raises TypeError: If the arguments are of the wrong type. :raises ValueError: When there are issues loading a previous checkpoint. """ # Save the dataset files for later use in cross validation analysis config.write_dataset_files() # set the random seed for all libraries ml_util.set_random_seed(config.get_effective_random_seed(), "Patch visualization") # Visualize how patches are sampled for segmentation models. This changes the random generator, but we don't # want training to depend on how many patients we visualized, and hence set the random seed again right after. with logging_section( "Visualizing the effect of sampling random crops for training"): visualize_random_crops_for_dataset(config) ml_util.set_random_seed(config.get_effective_random_seed(), "Model training") logging.debug("Creating the PyTorch model.") # Create the train loader and validation loader to load images from the dataset data_loaders = config.create_data_loaders() # Get the path to the checkpoint to recover from checkpoint_path = checkpoint_handler.get_recovery_path_train() models_and_optimizer = ModelAndInfo( config=config, model_execution_mode=ModelExecutionMode.TRAIN, checkpoint_path=checkpoint_path) # Create the main model # If continuing from a previous run at a specific epoch, then load the previous model. model_loaded = models_and_optimizer.try_create_model_and_load_from_checkpoint( ) if not model_loaded: raise ValueError( "There was no checkpoint file available for the model for given start_epoch {}" .format(config.start_epoch)) # Print out a detailed breakdown of layers, memory consumption and time. generate_and_print_model_summary(config, models_and_optimizer.model) # Move model to GPU and adjust for multiple GPUs models_and_optimizer.adjust_model_for_gpus() # Create the mean teacher model and move to GPU if config.compute_mean_teacher_model: mean_teacher_model_loaded = models_and_optimizer.try_create_mean_teacher_model_load_from_checkpoint_and_adjust( ) if not mean_teacher_model_loaded: raise ValueError( "There was no checkpoint file available for the mean teacher model " f"for given start_epoch {config.start_epoch}") # Create optimizer models_and_optimizer.create_optimizer() if checkpoint_handler.should_load_optimizer_checkpoint(): optimizer_loaded = models_and_optimizer.try_load_checkpoint_for_optimizer( ) if not optimizer_loaded: raise ValueError( f"There was no checkpoint file available for the optimizer for given start_epoch " f"{config.start_epoch}") # Create checkpoint directory for this run if it doesn't already exist logging.info(f"Models are saved at {config.checkpoint_folder}") if not config.checkpoint_folder.is_dir(): config.checkpoint_folder.mkdir() # Create the SummaryWriters for Tensorboard writers = create_summary_writers(config) config.create_dataframe_loggers() # Create LR scheduler l_rate_scheduler = SchedulerWithWarmUp(config, models_and_optimizer.optimizer) # Training loop logging.info("Starting training") train_results_per_epoch, val_results_per_epoch, learning_rates_per_epoch = [], [], [] resource_monitor = None if config.monitoring_interval_seconds > 0: # initialize and start GPU monitoring diagnostics_events = config.logs_folder / "diagnostics" logging.info( f"Starting resource monitor, outputting to {diagnostics_events}") resource_monitor = ResourceMonitor( interval_seconds=config.monitoring_interval_seconds, tensorboard_folder=diagnostics_events) resource_monitor.start() gradient_scaler = GradScaler( ) if config.use_gpu and config.use_mixed_precision else None optimal_temperature_scale_values = [] for epoch in config.get_train_epochs(): logging.info("Starting epoch {}".format(epoch)) save_epoch = config.should_save_epoch( epoch) and models_and_optimizer.optimizer is not None # store the learning rates used for each epoch epoch_lrs = l_rate_scheduler.get_last_lr() learning_rates_per_epoch.append(epoch_lrs) train_val_params: TrainValidateParameters = \ TrainValidateParameters(data_loader=data_loaders[ModelExecutionMode.TRAIN], model=models_and_optimizer.model, mean_teacher_model=models_and_optimizer.mean_teacher_model, epoch=epoch, optimizer=models_and_optimizer.optimizer, gradient_scaler=gradient_scaler, epoch_learning_rate=epoch_lrs, summary_writers=writers, dataframe_loggers=config.metrics_data_frame_loggers, in_training_mode=True) training_steps = create_model_training_steps(config, train_val_params) train_epoch_results = train_or_validate_epoch(training_steps) train_results_per_epoch.append(train_epoch_results.metrics) metrics.validate_and_store_model_parameters(writers.train, epoch, models_and_optimizer.model) # Run without adjusting weights on the validation set train_val_params.in_training_mode = False train_val_params.data_loader = data_loaders[ModelExecutionMode.VAL] # if temperature scaling is enabled then do not save validation metrics for the checkpoint epochs # as these will be re-computed after performing temperature scaling on the validation set. if isinstance(config, SequenceModelBase): train_val_params.save_metrics = not ( save_epoch and config.temperature_scaling_config) training_steps = create_model_training_steps(config, train_val_params) val_epoch_results = train_or_validate_epoch(training_steps) val_results_per_epoch.append(val_epoch_results.metrics) if config.is_segmentation_model: metrics.store_epoch_stats_for_segmentation( config.outputs_folder, epoch, epoch_lrs, train_epoch_results.metrics, val_epoch_results.metrics) if save_epoch: # perform temperature scaling if required if isinstance( config, SequenceModelBase) and config.temperature_scaling_config: optimal_temperature, scaled_val_results = \ temperature_scaling_steps(config, train_val_params, val_epoch_results) optimal_temperature_scale_values.append(optimal_temperature) # overwrite the metrics for the epoch with the metrics from the temperature scaled model val_results_per_epoch[-1] = scaled_val_results.metrics models_and_optimizer.save_checkpoint(epoch) # Updating the learning rate should happen at the end of the training loop, so that the # initial learning rate will be used for the very first epoch. l_rate_scheduler.step() model_training_results = ModelTrainingResults( train_results_per_epoch=train_results_per_epoch, val_results_per_epoch=val_results_per_epoch, learning_rates_per_epoch=learning_rates_per_epoch, optimal_temperature_scale_values_per_checkpoint_epoch= optimal_temperature_scale_values) logging.info("Finished training") # Since we have trained the model further, let the checkpoint_handler object know so it can handle # checkpoints correctly. checkpoint_handler.additional_training_done() # Upload visualization directory to AML run context to be able to see it # in the Azure UI. if config.max_batch_grad_cam > 0 and config.visualization_folder.exists(): RUN_CONTEXT.upload_folder(name=VISUALIZATION_FOLDER, path=str(config.visualization_folder)) writers.close_all() config.metrics_data_frame_loggers.close_all() if resource_monitor: # stop the resource monitoring process logging.info( "Shutting down the resource monitor process. Aggregate resource utilization:" ) for name, value in resource_monitor.read_aggregate_metrics(): logging.info(f"{name}: {value}") if not is_offline_run_context(RUN_CONTEXT): RUN_CONTEXT.log(name, value) resource_monitor.kill() return model_training_results
def run(self) -> None: """ Driver function to run a ML experiment. If an offline cross validation run is requested, then this function is recursively called for each cross validation split. """ if self.is_offline_cross_val_parent_run(): if self.model_config.is_segmentation_model: raise NotImplementedError("Offline cross validation is only supported for classification models.") self.spawn_offline_cross_val_classification_child_runs() return # Get the AzureML context in which the script is running if not self.model_config.is_offline_run and PARENT_RUN_CONTEXT is not None: logging.info("Setting tags from parent run.") self.set_run_tags_from_parent() self.save_build_info_for_dotnet_consumers() # Set data loader start method self.set_multiprocessing_start_method() # configure recovery container if provided checkpoint_handler = CheckpointHandler(model_config=self.model_config, azure_config=self.azure_config, project_root=self.project_root, run_context=RUN_CONTEXT) checkpoint_handler.download_recovery_checkpoints_or_weights() # do training and inference, unless the "only register" switch is set (which requires a run_recovery # to be valid). if not self.azure_config.only_register_model: # Set local_dataset to the mounted path specified in azure_runner.py, if any, or download it if that fails # and config.local_dataset was not already set. self.model_config.local_dataset = self.mount_or_download_dataset() # Check for existing dataset.csv file in the correct locations. Skip that if a dataset has already been # loaded (typically only during tests) if self.model_config.dataset_data_frame is None: assert self.model_config.local_dataset is not None ml_util.validate_dataset_paths( self.model_config.local_dataset, self.model_config.dataset_csv) # train a new model if required if self.azure_config.train: with logging_section("Model training"): model_train(self.model_config, checkpoint_handler, num_nodes=self.azure_config.num_nodes) else: self.model_config.write_dataset_files() self.create_activation_maps() # log the number of epochs used for model training RUN_CONTEXT.log(name="Train epochs", value=self.model_config.num_epochs) # We specify the ModelProcessing as DEFAULT here even if the run_recovery points to an ensemble run, because # the current run is a single one. See the documentation of ModelProcessing for more details. self.run_inference_and_register_model(checkpoint_handler, ModelProcessing.DEFAULT) if self.model_config.generate_report: self.generate_report(ModelProcessing.DEFAULT) # If this is an cross validation run, and the present run is child run 0, then wait for the sibling runs, # build the ensemble model, and write a report for that. if self.model_config.number_of_cross_validation_splits > 0: if self.model_config.should_wait_for_other_cross_val_child_runs(): self.wait_for_runs_to_finish() self.create_ensemble_model()
def log_metrics(self, metrics: Dict[str, float], step: Optional[int] = None) -> None: if self.is_azureml_run: for key, value in metrics.items(): RUN_CONTEXT.log(key, value)
def model_train(config: ModelConfigBase, checkpoint_handler: CheckpointHandler, num_nodes: int = 1) -> ModelTrainingResults: """ The main training loop. It creates the Pytorch model based on the configuration options passed in, creates a Pytorch Lightning trainer, and trains the model. If a checkpoint was specified, then it loads the checkpoint before resuming training. :param config: The arguments which specify all required information. :param checkpoint_handler: Checkpoint handler object to find checkpoint paths for model initialization :param num_nodes: The number of nodes to use in distributed training. """ # Get the path to the checkpoint to recover from checkpoint_path = checkpoint_handler.get_recovery_path_train() # This reads the dataset file, and possibly sets required pre-processing objects, like one-hot encoder # for categorical features, that need to be available before creating the model. config.read_dataset_if_needed() # Create the trainer object. Backup the environment variables before doing that, in case we need to run a second # training in the unit tests.d old_environ = dict(os.environ) seed_everything(config.get_effective_random_seed()) trainer, storing_logger = create_lightning_trainer(config, checkpoint_path, num_nodes=num_nodes) logging.info(f"GLOBAL_RANK: {os.getenv('GLOBAL_RANK')}, LOCAL_RANK {os.getenv('LOCAL_RANK')}. " f"trainer.global_rank: {trainer.global_rank}") logging.debug("Creating the PyTorch model.") lightning_model = create_lightning_model(config) lightning_model.storing_logger = storing_logger resource_monitor = None # Execute some bookkeeping tasks only once if running distributed: if is_rank_zero(): config.write_args_file() logging.info(str(config)) # Save the dataset files for later use in cross validation analysis config.write_dataset_files() logging.info(f"Model checkpoints are saved at {config.checkpoint_folder}") # set the random seed for all libraries ml_util.set_random_seed(config.get_effective_random_seed(), "Patch visualization") # Visualize how patches are sampled for segmentation models. This changes the random generator, but we don't # want training to depend on how many patients we visualized, and hence set the random seed again right after. with logging_section("Visualizing the effect of sampling random crops for training"): visualize_random_crops_for_dataset(config) # Print out a detailed breakdown of layers, memory consumption and time. generate_and_print_model_summary(config, lightning_model.model) if config.monitoring_interval_seconds > 0: # initialize and start GPU monitoring diagnostics_events = config.logs_folder / "diagnostics" logging.info(f"Starting resource monitor, outputting to {diagnostics_events}") resource_monitor = ResourceMonitor(interval_seconds=config.monitoring_interval_seconds, tensorboard_folder=diagnostics_events) resource_monitor.start() # Training loop logging.info("Starting training") lightning_data = TrainingAndValidationDataLightning(config) # type: ignore # When trying to store the config object in the constructor, it does not appear to get stored at all, later # reference of the object simply fail. Hence, have to set explicitly here. lightning_data.config = config trainer.fit(lightning_model, datamodule=lightning_data) trainer.logger.close() # type: ignore lightning_model.close_all_loggers() world_size = getattr(trainer, "world_size", 0) is_azureml_run = not config.is_offline_run # Per-subject model outputs for regression models are written per rank, and need to be aggregated here. # Each thread per rank will come here, and upload its files to the run outputs. Rank 0 will later download them. if is_azureml_run and world_size > 1 and isinstance(lightning_model, ScalarLightning): upload_output_file_as_temp(lightning_model.train_subject_outputs_logger.csv_path, config.outputs_folder) upload_output_file_as_temp(lightning_model.val_subject_outputs_logger.csv_path, config.outputs_folder) # DDP will start multiple instances of the runner, one for each GPU. Those should terminate here after training. # We can now use the global_rank of the Lightining model, rather than environment variables, because DDP has set # all necessary properties. if lightning_model.global_rank != 0: logging.info(f"Terminating training thread with rank {lightning_model.global_rank}.") sys.exit() logging.info("Choosing the best checkpoint and removing redundant files.") cleanup_checkpoint_folder(config.checkpoint_folder) # Lightning modifies a ton of environment variables. If we first run training and then the test suite, # those environment variables will mislead the training runs in the test suite, and make them crash. # Hence, restore the original environment after training. os.environ.clear() os.environ.update(old_environ) if world_size and isinstance(lightning_model, ScalarLightning): if is_azureml_run and world_size > 1: # In a DDP run on the local box, all ranks will write to local disk, hence no download needed. # In a multi-node DDP, each rank would upload to AzureML, and rank 0 will now download all results and # concatenate for rank in range(world_size): for mode in [ModelExecutionMode.TRAIN, ModelExecutionMode.VAL]: file = mode.value + "/" + get_subject_output_file_per_rank(rank) RUN_CONTEXT.download_file(name=TEMP_PREFIX + file, output_file_path=config.outputs_folder / file) # Concatenate all temporary file per execution mode for mode in [ModelExecutionMode.TRAIN, ModelExecutionMode.VAL]: temp_files = (config.outputs_folder / mode.value).rglob(SUBJECT_OUTPUT_PER_RANK_PREFIX + "*") result_file = config.outputs_folder / mode.value / SUBJECT_METRICS_FILE_NAME for i, file in enumerate(temp_files): temp_file_contents = file.read_text() if i == 0: # Copy the first file as-is, including the first line with the column headers result_file.write_text(temp_file_contents) else: # For all files but the first one, cut off the header line. result_file.write_text(os.linesep.join(temp_file_contents.splitlines()[1:])) model_training_results = ModelTrainingResults( train_results_per_epoch=list(storing_logger.to_metrics_dicts(prefix_filter=TRAIN_PREFIX).values()), val_results_per_epoch=list(storing_logger.to_metrics_dicts(prefix_filter=VALIDATION_PREFIX).values()), train_diagnostics=lightning_model.train_diagnostics, val_diagnostics=lightning_model.val_diagnostics, optimal_temperature_scale_values_per_checkpoint_epoch=[] ) logging.info("Finished training") # Since we have trained the model further, let the checkpoint_handler object know so it can handle # checkpoints correctly. checkpoint_handler.additional_training_done() # Upload visualization directory to AML run context to be able to see it # in the Azure UI. if config.max_batch_grad_cam > 0 and config.visualization_folder.exists(): RUN_CONTEXT.upload_folder(name=VISUALIZATION_FOLDER, path=str(config.visualization_folder)) if resource_monitor: # stop the resource monitoring process logging.info("Shutting down the resource monitor process. Aggregate resource utilization:") for name, value in resource_monitor.read_aggregate_metrics(): logging.info(f"{name}: {value}") if not config.is_offline_run: RUN_CONTEXT.log(name, value) resource_monitor.kill() return model_training_results