# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ MaskFormer Training Script. This script is a simplified version of the training script in detectron2/tools. """ import copy import itertools import logging import os from collections import OrderedDict from typing import Any, Dict, List, Set import torch import detectron2.utils.comm as comm from detectron2.checkpoint import DetectionCheckpointer from detectron2.config import get_cfg from detectron2.data import MetadataCatalog, build_detection_train_loader from detectron2.engine import DefaultTrainer, default_argument_parser, default_setup, launch from detectron2.evaluation import CityscapesInstanceEvaluator, CityscapesSemSegEvaluator, \ COCOEvaluator, COCOPanopticEvaluator, DatasetEvaluators, SemSegEvaluator, verify_results, \ DatasetEvaluator from detectron2.projects.deeplab import add_deeplab_config, build_lr_scheduler from detectron2.solver.build import maybe_add_gradient_clipping from detectron2.utils.logger import setup_logger from detectron2.utils.file_io import PathManager import numpy as np from PIL import Image import glob import pycocotools.mask as mask_util from detectron2.data import DatasetCatalog, MetadataCatalog from detectron2.utils.comm import all_gather, is_main_process, synchronize import json from torch.nn.parallel import DistributedDataParallel from detectron2.engine.train_loop import AMPTrainer, SimpleTrainer, TrainerBase, HookBase import weakref from detectron2.utils.events import EventStorage from detectron2.utils.logger import _log_api_usage # from detectron2.evaluation import SemSegGzeroEvaluator # from mask_former.evaluation.sem_seg_evaluation_gzero import SemSegGzeroEvaluator class SemSegGzeroEvaluator(DatasetEvaluator): """ Evaluate semantic segmentation metrics. """ def __init__( self, dataset_name, distributed, output_dir=None, *, num_classes=None, ignore_label=None ): """ Args: dataset_name (str): name of the dataset to be evaluated. distributed (True): if True, will collect results from all ranks for evaluation. Otherwise, will evaluate the results in the current process. output_dir (str): an output directory to dump results. num_classes, ignore_label: deprecated argument """ self._logger = logging.getLogger(__name__) if num_classes is not None: self._logger.warn( "SemSegEvaluator(num_classes) is deprecated! It should be obtained from metadata." ) if ignore_label is not None: self._logger.warn( "SemSegEvaluator(ignore_label) is deprecated! It should be obtained from metadata." ) self._dataset_name = dataset_name self._distributed = distributed self._output_dir = output_dir self._cpu_device = torch.device("cpu") self.input_file_to_gt_file = { dataset_record["file_name"]: dataset_record["sem_seg_file_name"] for dataset_record in DatasetCatalog.get(dataset_name) } meta = MetadataCatalog.get(dataset_name) # Dict that maps contiguous training ids to COCO category ids try: c2d = meta.stuff_dataset_id_to_contiguous_id self._contiguous_id_to_dataset_id = {v: k for k, v in c2d.items()} except AttributeError: self._contiguous_id_to_dataset_id = None self._class_names = meta.stuff_classes self._val_extra_classes = meta.val_extra_classes self._num_classes = len(meta.stuff_classes) if num_classes is not None: assert self._num_classes == num_classes, f"{self._num_classes} != {num_classes}" self._ignore_label = ignore_label if ignore_label is not None else meta.ignore_label def reset(self): self._conf_matrix = np.zeros((self._num_classes + 1, self._num_classes + 1), dtype=np.int64) self._predictions = [] def process(self, inputs, outputs): """ Args: inputs: the inputs to a model. It is a list of dicts. Each dict corresponds to an image and contains keys like "height", "width", "file_name". outputs: the outputs of a model. It is either list of semantic segmentation predictions (Tensor [H, W]) or list of dicts with key "sem_seg" that contains semantic segmentation prediction in the same format. """ for input, output in zip(inputs, outputs): output = output["sem_seg"].argmax(dim=0).to(self._cpu_device) pred = np.array(output, dtype=np.int) with PathManager.open(self.input_file_to_gt_file[input["file_name"]], "rb") as f: gt = np.array(Image.open(f), dtype=np.int) gt[gt == self._ignore_label] = self._num_classes self._conf_matrix += np.bincount( (self._num_classes + 1) * pred.reshape(-1) + gt.reshape(-1), minlength=self._conf_matrix.size, ).reshape(self._conf_matrix.shape) self._predictions.extend(self.encode_json_sem_seg(pred, input["file_name"])) def evaluate(self): """ Evaluates standard semantic segmentation metrics (http://cocodataset.org/#stuff-eval): * Mean intersection-over-union averaged across classes (mIoU) * Frequency Weighted IoU (fwIoU) * Mean pixel accuracy averaged across classes (mACC) * Pixel Accuracy (pACC) """ if self._distributed: synchronize() conf_matrix_list = all_gather(self._conf_matrix) self._predictions = all_gather(self._predictions) self._predictions = list(itertools.chain(*self._predictions)) if not is_main_process(): return self._conf_matrix = np.zeros_like(self._conf_matrix) for conf_matrix in conf_matrix_list: self._conf_matrix += conf_matrix if self._output_dir: PathManager.mkdirs(self._output_dir) file_path = os.path.join(self._output_dir, "sem_seg_predictions.json") with PathManager.open(file_path, "w") as f: f.write(json.dumps(self._predictions)) acc = np.full(self._num_classes, np.nan, dtype=np.float) iou = np.full(self._num_classes, np.nan, dtype=np.float) tp = self._conf_matrix.diagonal()[:-1].astype(np.float) pos_gt = np.sum(self._conf_matrix[:-1, :-1], axis=0).astype(np.float) class_weights = pos_gt / np.sum(pos_gt) pos_pred = np.sum(self._conf_matrix[:-1, :-1], axis=1).astype(np.float) acc_valid = pos_gt > 0 acc[acc_valid] = tp[acc_valid] / pos_gt[acc_valid] iou_valid = (pos_gt + pos_pred) > 0 union = pos_gt + pos_pred - tp iou[acc_valid] = tp[acc_valid] / union[acc_valid] macc = np.sum(acc[acc_valid]) / np.sum(acc_valid) miou = np.sum(iou[acc_valid]) / np.sum(iou_valid) fiou = np.sum(iou[acc_valid] * class_weights[acc_valid]) pacc = np.sum(tp) / np.sum(pos_gt) seen_IoU = 0 unseen_IoU = 0 seen_acc = 0 unseen_acc = 0 res = {} res["mIoU"] = 100 * miou res["fwIoU"] = 100 * fiou for i, name in enumerate(self._class_names): res["IoU-{}".format(name)] = 100 * iou[i] if name in self._val_extra_classes: unseen_IoU = unseen_IoU + 100 * iou[i] else: seen_IoU = seen_IoU + 100 * iou[i] unseen_IoU = unseen_IoU / len(self._val_extra_classes) seen_IoU = seen_IoU / (self._num_classes - len(self._val_extra_classes)) res["mACC"] = 100 * macc res["pACC"] = 100 * pacc for i, name in enumerate(self._class_names): res["ACC-{}".format(name)] = 100 * acc[i] if name in self._val_extra_classes: unseen_acc = unseen_acc + 100 * iou[i] else: seen_acc = seen_acc + 100 * iou[i] unseen_acc = unseen_acc / len(self._val_extra_classes) seen_acc = seen_acc / (self._num_classes - len(self._val_extra_classes)) res["seen_IoU"] = seen_IoU res["unseen_IoU"] = unseen_IoU res["harmonic mean"] = 2 * seen_IoU * unseen_IoU / (seen_IoU + unseen_IoU) # res["unseen_acc"] = unseen_acc # res["seen_acc"] = seen_acc if self._output_dir: file_path = os.path.join(self._output_dir, "sem_seg_evaluation.pth") with PathManager.open(file_path, "wb") as f: torch.save(res, f) results = OrderedDict({"sem_seg": res}) self._logger.info(results) return results def encode_json_sem_seg(self, sem_seg, input_file_name): """ Convert semantic segmentation to COCO stuff format with segments encoded as RLEs. See http://cocodataset.org/#format-results """ json_list = [] for label in np.unique(sem_seg): if self._contiguous_id_to_dataset_id is not None: # import ipdb; ipdb.set_trace() assert ( label in self._contiguous_id_to_dataset_id ), "Label {} is not in the metadata info for {}".format(label, self._dataset_name) dataset_id = self._contiguous_id_to_dataset_id[label] else: dataset_id = int(label) mask = (sem_seg == label).astype(np.uint8) mask_rle = mask_util.encode(np.array(mask[:, :, None], order="F"))[0] mask_rle["counts"] = mask_rle["counts"].decode("utf-8") json_list.append( {"file_name": input_file_name, "category_id": dataset_id, "segmentation": mask_rle} ) return json_list # MaskFormer from cat_seg import ( DETRPanopticDatasetMapper, MaskFormerPanopticDatasetMapper, MaskFormerSemanticDatasetMapper, SemanticSegmentorWithTTA, add_mask_former_config, ) def create_ddp_model(model, *, fp16_compression=False, **kwargs): """ Create a DistributedDataParallel model if there are >1 processes. Args: model: a torch.nn.Module fp16_compression: add fp16 compression hooks to the ddp object. See more at https://pytorch.org/docs/stable/ddp_comm_hooks.html#torch.distributed.algorithms.ddp_comm_hooks.default_hooks.fp16_compress_hook kwargs: other arguments of :module:`torch.nn.parallel.DistributedDataParallel`. """ # noqa if comm.get_world_size() == 1: return model if "device_ids" not in kwargs: kwargs["device_ids"] = [comm.get_local_rank()] ddp = DistributedDataParallel(model, **kwargs) if fp16_compression: from torch.distributed.algorithms.ddp_comm_hooks import default as comm_hooks ddp.register_comm_hook(state=None, hook=comm_hooks.fp16_compress_hook) return ddp class Trainer(DefaultTrainer): """ Extension of the Trainer class adapted to DETR. """ def __init__(self, cfg): # super().__init__(cfg) self._hooks: List[HookBase] = [] self.iter: int = 0 self.start_iter: int = 0 self.max_iter: int self.storage: EventStorage _log_api_usage("trainer." + self.__class__.__name__) logger = logging.getLogger("detectron2") if not logger.isEnabledFor(logging.INFO): # setup_logger is not called for d2 setup_logger() cfg = DefaultTrainer.auto_scale_workers(cfg, comm.get_world_size()) # Assume these objects must be constructed in this order. model = self.build_model(cfg) optimizer = self.build_optimizer(cfg, model) data_loader = self.build_train_loader(cfg) model = create_ddp_model(model, broadcast_buffers=False, find_unused_parameters=True) self._trainer = (AMPTrainer if cfg.SOLVER.AMP.ENABLED else SimpleTrainer)( model, data_loader, optimizer ) self.scheduler = self.build_lr_scheduler(cfg, optimizer) self.checkpointer = DetectionCheckpointer( # Assume you want to save checkpoints together with logs/statistics model, cfg.OUTPUT_DIR, trainer=weakref.proxy(self), ) self.start_iter = 0 self.max_iter = cfg.SOLVER.MAX_ITER self.cfg = cfg self.register_hooks(self.build_hooks()) @classmethod def build_evaluator(cls, cfg, dataset_name, output_folder=None): """ Create evaluator(s) for a given dataset. This uses the special metadata "evaluator_type" associated with each builtin dataset. For your own dataset, you can simply create an evaluator manually in your script and do not have to worry about the hacky if-else logic here. """ if output_folder is None: output_folder = os.path.join(cfg.OUTPUT_DIR, "inference") evaluator_list = [] evaluator_type = MetadataCatalog.get(dataset_name).evaluator_type if evaluator_type in ["sem_seg", "ade20k_panoptic_seg"]: evaluator_list.append( SemSegEvaluator( dataset_name, distributed=True, output_dir=output_folder, ) ) # import pdb; pdb.set_trace() if evaluator_type == "sem_seg_gzero": evaluator_list.append( SemSegGzeroEvaluator( dataset_name, distributed=True, output_dir=output_folder, ) ) if evaluator_type == "coco": evaluator_list.append(COCOEvaluator(dataset_name, output_dir=output_folder)) if evaluator_type in [ "coco_panoptic_seg", "ade20k_panoptic_seg", "cityscapes_panoptic_seg", ]: evaluator_list.append(COCOPanopticEvaluator(dataset_name, output_folder)) if evaluator_type == "cityscapes_instance": assert ( torch.cuda.device_count() >= comm.get_rank() ), "CityscapesEvaluator currently do not work with multiple machines." return CityscapesInstanceEvaluator(dataset_name) if evaluator_type == "cityscapes_sem_seg": assert ( torch.cuda.device_count() >= comm.get_rank() ), "CityscapesEvaluator currently do not work with multiple machines." return CityscapesSemSegEvaluator(dataset_name) if evaluator_type == "cityscapes_panoptic_seg": assert ( torch.cuda.device_count() >= comm.get_rank() ), "CityscapesEvaluator currently do not work with multiple machines." evaluator_list.append(CityscapesSemSegEvaluator(dataset_name)) if len(evaluator_list) == 0: raise NotImplementedError( "no Evaluator for the dataset {} with the type {}".format( dataset_name, evaluator_type ) ) elif len(evaluator_list) == 1: return evaluator_list[0] return DatasetEvaluators(evaluator_list) @classmethod def build_train_loader(cls, cfg): # Semantic segmentation dataset mapper if cfg.INPUT.DATASET_MAPPER_NAME == "mask_former_semantic": mapper = MaskFormerSemanticDatasetMapper(cfg, True) # Panoptic segmentation dataset mapper elif cfg.INPUT.DATASET_MAPPER_NAME == "mask_former_panoptic": mapper = MaskFormerPanopticDatasetMapper(cfg, True) # DETR-style dataset mapper for COCO panoptic segmentation elif cfg.INPUT.DATASET_MAPPER_NAME == "detr_panoptic": mapper = DETRPanopticDatasetMapper(cfg, True) else: mapper = None return build_detection_train_loader(cfg, mapper=mapper) @classmethod def build_lr_scheduler(cls, cfg, optimizer): """ It now calls :func:`detectron2.solver.build_lr_scheduler`. Overwrite it if you'd like a different scheduler. """ return build_lr_scheduler(cfg, optimizer) @classmethod def build_optimizer(cls, cfg, model): weight_decay_norm = cfg.SOLVER.WEIGHT_DECAY_NORM weight_decay_embed = cfg.SOLVER.WEIGHT_DECAY_EMBED defaults = {} defaults["lr"] = cfg.SOLVER.BASE_LR defaults["weight_decay"] = cfg.SOLVER.WEIGHT_DECAY norm_module_types = ( torch.nn.BatchNorm1d, torch.nn.BatchNorm2d, torch.nn.BatchNorm3d, torch.nn.SyncBatchNorm, # NaiveSyncBatchNorm inherits from BatchNorm2d torch.nn.GroupNorm, torch.nn.InstanceNorm1d, torch.nn.InstanceNorm2d, torch.nn.InstanceNorm3d, torch.nn.LayerNorm, torch.nn.LocalResponseNorm, ) params: List[Dict[str, Any]] = [] memo: Set[torch.nn.parameter.Parameter] = set() for module_name, module in model.named_modules(): for module_param_name, value in module.named_parameters(recurse=False): if not value.requires_grad: continue # Avoid duplicating parameters if value in memo: continue memo.add(value) hyperparams = copy.copy(defaults) if "backbone" in module_name: hyperparams["lr"] = hyperparams["lr"] * cfg.SOLVER.BACKBONE_MULTIPLIER if ( "relative_position_bias_table" in module_param_name or "absolute_pos_embed" in module_param_name ): print(module_param_name) hyperparams["weight_decay"] = 0.0 if isinstance(module, norm_module_types): hyperparams["weight_decay"] = weight_decay_norm if isinstance(module, torch.nn.Embedding): hyperparams["weight_decay"] = weight_decay_embed params.append({"params": [value], **hyperparams}) def maybe_add_full_model_gradient_clipping(optim): # detectron2 doesn't have full model gradient clipping now clip_norm_val = cfg.SOLVER.CLIP_GRADIENTS.CLIP_VALUE enable = ( cfg.SOLVER.CLIP_GRADIENTS.ENABLED and cfg.SOLVER.CLIP_GRADIENTS.CLIP_TYPE == "full_model" and clip_norm_val > 0.0 ) class FullModelGradientClippingOptimizer(optim): def step(self, closure=None): all_params = itertools.chain(*[x["params"] for x in self.param_groups]) torch.nn.utils.clip_grad_norm_(all_params, clip_norm_val) super().step(closure=closure) return FullModelGradientClippingOptimizer if enable else optim optimizer_type = cfg.SOLVER.OPTIMIZER if optimizer_type == "SGD": optimizer = maybe_add_full_model_gradient_clipping(torch.optim.SGD)( params, cfg.SOLVER.BASE_LR, momentum=cfg.SOLVER.MOMENTUM ) elif optimizer_type == "ADAMW": optimizer = maybe_add_full_model_gradient_clipping(torch.optim.AdamW)( params, cfg.SOLVER.BASE_LR ) else: raise NotImplementedError(f"no optimizer type {optimizer_type}") if not cfg.SOLVER.CLIP_GRADIENTS.CLIP_TYPE == "full_model": optimizer = maybe_add_gradient_clipping(cfg, optimizer) return optimizer @classmethod def test_with_TTA(cls, cfg, model): logger = logging.getLogger("detectron2.trainer") # In the end of training, run an evaluation with TTA. logger.info("Running inference with test-time augmentation ...") model = SemanticSegmentorWithTTA(cfg, model) evaluators = [ cls.build_evaluator( cfg, name, output_folder=os.path.join(cfg.OUTPUT_DIR, "inference_TTA") ) for name in cfg.DATASETS.TEST ] res = cls.test(cfg, model, evaluators) res = OrderedDict({k + "_TTA": v for k, v in res.items()}) return res def setup(args): """ Create configs and perform basic setups. """ cfg = get_cfg() # for poly lr schedule add_deeplab_config(cfg) add_mask_former_config(cfg) cfg.merge_from_file(args.config_file) cfg.merge_from_list(args.opts) cfg.freeze() default_setup(cfg, args) # Setup logger for "mask_former" module setup_logger(output=cfg.OUTPUT_DIR, distributed_rank=comm.get_rank(), name="mask_former") return cfg def main(args): cfg = setup(args) if args.eval_only: model = Trainer.build_model(cfg) DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load( cfg.MODEL.WEIGHTS, resume=args.resume ) res = Trainer.test(cfg, model) if cfg.TEST.AUG.ENABLED: res.update(Trainer.test_with_TTA(cfg, model)) if comm.is_main_process(): verify_results(cfg, res) return res trainer = Trainer(cfg) trainer.resume_or_load(resume=args.resume) return trainer.train() if __name__ == "__main__": args = default_argument_parser().parse_args() print("Command Line Args:", args) launch( main, args.num_gpus, num_machines=args.num_machines, machine_rank=args.machine_rank, dist_url=args.dist_url, args=(args,), )