def extract_clusters( cfg: AttrDict, dist_run_id: str, checkpoint_folder: str, local_rank: int = 0, node_id: int = 0, ): """ Sets up and executes model visualisation extraction workflow on one node """ # setup the environment variables set_env_vars(local_rank, node_id, cfg) dist_rank = int(os.environ["RANK"]) # setup logging setup_logging(__name__, output_dir=checkpoint_folder, rank=dist_rank) logging.info(f"Env set for rank: {local_rank}, dist_rank: {dist_rank}") # print the environment info for the current node if local_rank == 0: current_env = os.environ.copy() print_system_env_info(current_env) # setup the multiprocessing to be forkserver. # See https://fb.quip.com/CphdAGUaM5Wf setup_multiprocessing_method(cfg.MULTI_PROCESSING_METHOD) # set seeds logging.info("Setting seed....") set_seeds(cfg, dist_rank) # We set the CUDA device here as well as a safe solution for all downstream # `torch.cuda.current_device()` calls to return correct device. if cfg.MACHINE.DEVICE == "gpu" and torch.cuda.is_available(): local_rank, _ = get_machine_local_and_dist_rank() torch.cuda.set_device(local_rank) # print the training settings and system settings if local_rank == 0: print_cfg(cfg) logging.info("System config:\n{}".format(collect_env_info())) # Build the SSL trainer to set up distributed training and then # extract the cluster assignments for all entries in the dataset trainer = SelfSupervisionTrainer(cfg, dist_run_id) cluster_assignments = trainer.extract_clusters() # Save the cluster assignments in the output folder if dist_rank == 0: ClusterAssignmentLoader.save_cluster_assignment( output_dir=get_checkpoint_folder(cfg), assignments=ClusterAssignment( config=cfg, cluster_assignments=cluster_assignments), ) # close the logging streams including the file handlers logging.info("All Done!") shutdown_logging()
def extract_main( cfg: AttrDict, dist_run_id: str, checkpoint_folder: str, local_rank: int = 0, node_id: int = 0, ): """ Sets up and executes feature extraction workflow per machine. Args: cfg (AttrDict): user specified input config that has optimizer, loss, meters etc settings relevant to the training dist_run_id (str): For multi-gpu training with PyTorch, we have to specify how the gpus are going to rendezvous. This requires specifying the communication method: file, tcp and the unique rendezvous run_id that is specific to 1 run. We recommend: 1) for 1node: use init_method=tcp and run_id=auto 2) for multi-node, use init_method=tcp and specify run_id={master_node}:{port} local_rank (int): id of the current device on the machine. If using gpus, local_rank = gpu number on the current machine node_id (int): id of the current machine. starts from 0. valid for multi-gpu """ # setup the environment variables set_env_vars(local_rank, node_id, cfg) dist_rank = int(os.environ["RANK"]) # setup logging setup_logging(__name__, output_dir=checkpoint_folder, rank=dist_rank) logging.info(f"Env set for rank: {local_rank}, dist_rank: {dist_rank}") # print the environment info for the current node if local_rank == 0: current_env = os.environ.copy() print_system_env_info(current_env) # setup the multiprocessing to be forkserver. # See https://fb.quip.com/CphdAGUaM5Wf setup_multiprocessing_method(cfg.MULTI_PROCESSING_METHOD) # set seeds logging.info("Setting seed....") set_seeds(cfg, dist_rank) # We set the CUDA device here as well as a safe solution for all downstream # `torch.cuda.current_device()` calls to return correct device. if cfg.MACHINE.DEVICE == "gpu" and torch.cuda.is_available(): local_rank, _ = get_machine_local_and_dist_rank() torch.cuda.set_device(local_rank) # print the training settings and system settings if local_rank == 0: print_cfg(cfg) logging.info("System config:\n{}".format(collect_env_info())) trainer = SelfSupervisionTrainer(cfg, dist_run_id) features = trainer.extract() for split in features.keys(): logging.info(f"============== Split: {split} =======================") for layer_name, layer_features in features[split].items(): out_feat_file = os.path.join( checkpoint_folder, f"rank{dist_rank}_{split}_{layer_name}_features.npy") out_target_file = os.path.join( checkpoint_folder, f"rank{dist_rank}_{split}_{layer_name}_targets.npy") out_inds_file = os.path.join( checkpoint_folder, f"rank{dist_rank}_{split}_{layer_name}_inds.npy") feat_shape = layer_features["features"].shape logging.info( f"Saving extracted features of {layer_name} with shape {feat_shape} to: {out_feat_file}" ) save_file(layer_features["features"], out_feat_file) logging.info( f"Saving extracted targets of {layer_name} to: {out_target_file}" ) save_file(layer_features["targets"], out_target_file) logging.info( f"Saving extracted indices of {layer_name} to: {out_inds_file}" ) save_file(layer_features["inds"], out_inds_file) logging.info("All Done!") # close the logging streams including the filehandlers shutdown_logging()
def extract_label_predictions_main( cfg: AttrDict, dist_run_id: str, checkpoint_folder: str, local_rank: int = 0, node_id: int = 0, ): """ Sets up and executes label predictions workflow per machine. Runs the model in eval mode only to extract the label predicted per class. Args: cfg (AttrDict): user specified input config that has optimizer, loss, meters etc settings relevant for the feature extraction. dist_run_id (str): For multi-gpu training with PyTorch, we have to specify how the gpus are going to rendezvous. This requires specifying the communication method: file, tcp and the unique rendezvous run_id that is specific to 1 run. We recommend: 1) for 1node: use init_method=tcp and run_id=auto 2) for multi-node, use init_method=tcp and specify run_id={master_node}:{port} local_rank (int): id of the current device on the machine. If using gpus, local_rank = gpu number on the current machine node_id (int): id of the current machine. starts from 0. valid for multi-gpu """ # setup the environment variables set_env_vars(local_rank, node_id, cfg) dist_rank = int(os.environ["RANK"]) # setup logging setup_logging(__name__, output_dir=checkpoint_folder, rank=dist_rank) # setup the multiprocessing to be forkserver. See https://fb.quip.com/CphdAGUaM5Wf logging.info( f"Setting multiprocessing method: {cfg.MULTI_PROCESSING_METHOD}") setup_multiprocessing_method(cfg.MULTI_PROCESSING_METHOD) # set seeds logging.info("Setting seed....") set_seeds(cfg, dist_rank) # We set the CUDA device here as well as a safe solution for all downstream # `torch.cuda.current_device()` calls to return correct device. if cfg.MACHINE.DEVICE == "gpu" and torch.cuda.is_available(): local_rank, _ = get_machine_local_and_dist_rank() torch.cuda.set_device(local_rank) # print the training settings and system settings # print the environment info for the current node logging.info(f"Env set for rank: {local_rank}, dist_rank: {dist_rank}") if local_rank == 0: current_env = os.environ.copy() print_system_env_info(current_env) print_cfg(cfg) logging.info(f"System config:\n{collect_env_info()}") # Identify the hooks to run for the extract label engine # TODO - we need to plug this better with the engine registry # - we either need to use the global hooks registry # - or we need to create specific hook registry by engine hooks = extract_label_hook_generator(cfg) trainer = SelfSupervisionTrainer(cfg, dist_run_id, hooks=hooks) trainer.extract( output_folder=cfg.EXTRACT_FEATURES.OUTPUT_DIR or checkpoint_folder, extract_features=False, extract_predictions=True, ) logging.info("All Done!") # close the logging streams including the filehandlers shutdown_logging()
def train_main( cfg: AttrDict, dist_run_id: str, checkpoint_path: str, checkpoint_folder: str, local_rank: int = 0, node_id: int = 0, hook_generator: Callable[[Any], List[ClassyHook]] = default_hook_generator, ): """ Sets up and executes training workflow per machine. Args: cfg (AttrDict): user specified input config that has optimizer, loss, meters etc settings relevant to the training dist_run_id (str): For multi-gpu training with PyTorch, we have to specify how the gpus are going to rendezvous. This requires specifying the communication method: file, tcp and the unique rendezvous run_id that is specific to 1 run. We recommend: 1) for 1node: use init_method=tcp and run_id=auto 2) for multi-node, use init_method=tcp and specify run_id={master_node}:{port} checkpoint_path (str): if the training is being resumed from a checkpoint, path to the checkpoint. The tools/run_distributed_engines.py automatically looks for the checkpoint in the checkpoint directory. checkpoint_folder (str): what directory to use for checkpointing. The tools/run_distributed_engines.py creates the directory based on user input in the yaml config file. local_rank (int): id of the current device on the machine. If using gpus, local_rank = gpu number on the current machine node_id (int): id of the current machine. starts from 0. valid for multi-gpu hook_generator (Callable): The utility function that prepares all the hoooks that will be used in training based on user selection. Some basic hooks are used by default. """ # setup the environment variables set_env_vars(local_rank, node_id, cfg) dist_rank = int(os.environ["RANK"]) # setup logging setup_logging(__name__, output_dir=checkpoint_folder, rank=dist_rank) logging.info(f"Env set for rank: {local_rank}, dist_rank: {dist_rank}") # print the environment info for the current node if local_rank == 0: current_env = os.environ.copy() print_system_env_info(current_env) # setup the multiprocessing to be forkserver. # See https://fb.quip.com/CphdAGUaM5Wf setup_multiprocessing_method(cfg.MULTI_PROCESSING_METHOD) # set seeds logging.info("Setting seed....") set_seeds(cfg, dist_rank) # We set the CUDA device here as well as a safe solution for all downstream # `torch.cuda.current_device()` calls to return correct device. if cfg.MACHINE.DEVICE == "gpu" and torch.cuda.is_available(): local_rank, _ = get_machine_local_and_dist_rank() torch.cuda.set_device(local_rank) # print the training settings and system settings if local_rank == 0: print_cfg(cfg) logging.info("System config:\n{}".format(collect_env_info())) # get the hooks - these hooks are executed per replica hooks = hook_generator(cfg) # build the SSL trainer. The trainer first prepares a "task" object which # acts as a container for various things needed in a training: datasets, # dataloader, optimizers, losses, hooks, etc. "Task" will also have information # about phases (train, test) both. The trainer then sets up distributed # training. trainer = SelfSupervisionTrainer( cfg, dist_run_id, checkpoint_path, checkpoint_folder, hooks ) trainer.train() logging.info("All Done!") # close the logging streams including the filehandlers shutdown_logging()
def extract_features_main( cfg: AttrDict, dist_run_id: str, checkpoint_folder: str, local_rank: int = 0, node_id: int = 0, ): """ Sets up and executes feature extraction workflow per machine. Args: cfg (AttrDict): user specified input config that has optimizer, loss, meters etc settings relevant to the training dist_run_id (str): For multi-gpu training with PyTorch, we have to specify how the gpus are going to rendezvous. This requires specifying the communication method: file, tcp and the unique rendezvous run_id that is specific to 1 run. We recommend: 1) for 1node: use init_method=tcp and run_id=auto 2) for multi-node, use init_method=tcp and specify run_id={master_node}:{port} checkpoint_folder (str): what directory to use for checkpointing. This folder will be used to output the extracted features as well in case config.EXTRACT_FEATURES.OUTPUT_DIR is not set local_rank (int): id of the current device on the machine. If using gpus, local_rank = gpu number on the current machine node_id (int): id of the current machine. starts from 0. valid for multi-gpu """ # setup the environment variables set_env_vars(local_rank, node_id, cfg) dist_rank = int(os.environ["RANK"]) # setup logging setup_logging(__name__, output_dir=checkpoint_folder, rank=dist_rank) logging.info(f"Env set for rank: {local_rank}, dist_rank: {dist_rank}") # print the environment info for the current node if local_rank == 0: current_env = os.environ.copy() print_system_env_info(current_env) # setup the multiprocessing to be forkserver. # See https://fb.quip.com/CphdAGUaM5Wf setup_multiprocessing_method(cfg.MULTI_PROCESSING_METHOD) # set seeds logging.info("Setting seed....") set_seeds(cfg, dist_rank) # We set the CUDA device here as well as a safe solution for all downstream # `torch.cuda.current_device()` calls to return correct device. if cfg.MACHINE.DEVICE == "gpu" and torch.cuda.is_available(): local_rank, _ = get_machine_local_and_dist_rank() torch.cuda.set_device(local_rank) # print the training settings and system settings if local_rank == 0: print_cfg(cfg) logging.info("System config:\n{}".format(collect_env_info())) # Identify the hooks to run for the extract label engine # TODO - we need to plug this better with the engine registry # - we either need to use the global hooks registry # - or we need to create specific hook registry by engine hooks = extract_features_hook_generator(cfg) # Run the label prediction extraction trainer = SelfSupervisionTrainer(cfg, dist_run_id, hooks=hooks) output_dir = cfg.EXTRACT_FEATURES.OUTPUT_DIR or checkpoint_folder trainer.extract( output_folder=cfg.EXTRACT_FEATURES.OUTPUT_DIR or checkpoint_folder, extract_features=True, extract_predictions=False, ) # TODO (prigoyal): merge this function with _extract_features if dist_rank == 0 and cfg.EXTRACT_FEATURES.MAP_FEATURES_TO_IMG_NAME: # Get the names of the features that we extracted features for. If user doesn't # specify the features to evaluate, we get the full model output and freeze # head/trunk both as caution. layers = get_trunk_output_feature_names(cfg.MODEL) if len(layers) == 0: layers = ["heads"] available_splits = [ item.lower() for item in trainer.task.available_splits ] for split in available_splits: image_paths = trainer.task.datasets[split].get_image_paths()[0] for layer in layers: ExtractedFeaturesLoader.map_features_to_img_filepath( image_paths=image_paths, input_dir=output_dir, split=split, layer=layer, ) logging.info("All Done!") # close the logging streams including the filehandlers shutdown_logging()