train_net.py

from typing import Any, Dict, List, Set
import copy
import itertools
import os
import argparse
import torch
import torch.distributed as dist
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, build_detection_test_loader
from detectron2.engine import (
    DefaultTrainer,
    default_setup,
)
from detectron2.evaluation import DatasetEvaluators, verify_results
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 uni_3d import (
    Front3DDPSEvaluator,
    Front3DEvaluator,
    MatterportDPSEvaluator,
    Front3DDPSDatasetMapper,
    Front3DDatasetMapper,
    Front3DTestDatasetMapper,
    add_mask2former_config,
    add_uni_3d_config,
)


class Trainer(DefaultTrainer):
    """
    Extension of the Trainer class adapted to MaskFormer.
    """

    @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 == "matterport_dps":
            assert (
                torch.cuda.device_count() > comm.get_rank()
            ), "MatterportDPSEvaluator currently do not work with multiple machines."
            evaluator_list.append(MatterportDPSEvaluator(dataset_name, output_dir=output_folder))
        if evaluator_type == "front3d_dps":
            assert (
                torch.cuda.device_count() > comm.get_rank()
            ), "Front3DDPSEvaluator currently do not work with multiple machines."
            evaluator_list.append(Front3DDPSEvaluator(dataset_name, output_dir=output_folder))
        if evaluator_type == "front3d":
            evaluator_list.append(Front3DEvaluator(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)
    
    def before_step(self):
        super().before_step()
        if self.cfg.GLOBAL.EMPTY_CACHE_BEFORE_STEP:
            torch.cuda.empty_cache()

    @classmethod
    def build_train_loader(cls, cfg):
        if cfg.INPUT.DATASET_MAPPER_NAME == "front3d_dps":
            mapper = Front3DDPSDatasetMapper(cfg, True)
            return build_detection_train_loader(cfg, mapper=mapper)
        elif cfg.INPUT.DATASET_MAPPER_NAME == "front3d":
            mapper = Front3DDatasetMapper(cfg, True)
            return build_detection_train_loader(cfg, mapper=mapper)
        else:
            mapper = None
            return build_detection_train_loader(cfg, mapper=mapper)
        
    @classmethod
    def build_test_loader(cls, cfg, dataset_name):
        if cfg.INPUT.DATASET_MAPPER_NAME == "front3d":
            mapper = Front3DTestDatasetMapper(cfg)
        else:
            mapper = None
        return build_detection_test_loader(cfg, dataset_name, 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


def setup(args):
    """
    Create configs and perform basic setups.
    """
    cfg = get_cfg()
    # for poly lr schedule
    add_deeplab_config(cfg)
    add_mask2former_config(cfg)
    add_uni_3d_config(cfg)
    cfg.merge_from_file(args.config_file)
    cfg.merge_from_list(args.opts)
    cfg.freeze()
    default_setup(cfg, args)
    setup_logger(output=cfg.OUTPUT_DIR, distributed_rank=comm.get_rank(), name="uni_3d")
    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 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__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--config-file", default="", metavar="FILE", help="path to config file")
    parser.add_argument(
        "--resume",
        action="store_true",
        help="Whether to attempt to resume from the checkpoint directory. "
        "See documentation of `DefaultTrainer.resume_or_load()` for what it means.",
    )
    parser.add_argument("--eval-only", action="store_true", help="perform evaluation only")
    parser.add_argument(
        "opts",
        help="""
Modify config options at the end of the command. For Yacs configs, use
space-separated "PATH.KEY VALUE" pairs.
For python-based LazyConfig, use "path.key=value".
        """.strip(),
        default=None,
        nargs=argparse.REMAINDER,
    )
    args = parser.parse_args()
    print("Command Line Args:", args)
    
    world_size = int(os.environ.get("WORLD_SIZE", 1))
    if world_size > 1:
        dist.init_process_group()
        comm.create_local_process_group(int(os.environ["LOCAL_WORLD_SIZE"]))

        if torch.cuda.is_available():
            torch.cuda.set_device(comm.get_local_rank())
    comm.synchronize()
    main(args)