train_net.py

# TRY MY BEST TO SEED
import random
import numpy as np
import torch
SEED = 0
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
torch.cuda.manual_seed_all(SEED)
# torch.autograd.set_detect_anomaly(True)
# torch.use_deterministic_algorithms(True)
# torch.backends.cudnn.deterministic = True

import copy
import itertools
import logging
import os
from collections import OrderedDict
from typing import Any, Dict, List, Set

import detectron2.utils.comm as comm
from detectron2.checkpoint import DetectionCheckpointer
from detectron2.data import build_detection_train_loader, build_detection_test_loader
from detectron2.engine import DefaultTrainer, default_argument_parser, default_setup, launch
from detectron2.engine import hooks
from detectron2.evaluation import (
    DatasetEvaluators,
    verify_results,
)
from detectron2.projects.deeplab import build_lr_scheduler
from detectron2.solver.build import maybe_add_gradient_clipping


from evaluator.evaluator import TrajectoryEvaluator
from dataloader.dataset_mappers.dataset_mapper import DatasetMapper

from model import setup

from common.envs import *


class Trainer(DefaultTrainer):
    def __init__(self, cfg):
        super().__init__(cfg)
        self.eval_period = cfg.TEST.EVAL_PERIOD
        self.training_evaluators = []
        for e in self.training_evaluators:
            e.reset()

    def run_step(self):
        self._trainer.iter = self.iter
        self._trainer.run_step()
        if self._trainer.iter % self.eval_period == 0:
            for e in self.training_evaluators:
                e.evaluate()
            for e in self.training_evaluators:
                e.reset()
        for e in self.training_evaluators:
            e.process(self._trainer.model.training_input, self._trainer.model.training_inference)

    @classmethod
    def build_evaluator(cls, cfg, dataset_name, output_folder=None):
        if output_folder is None:
            output_folder = os.path.join(cfg.OUTPUT_DIR, "inference")
        evaluator_list = []
        if os.getenv('DEPLOY_SEQS') is not None:
            evaluator_list.append(TrajectoryEvaluator(dataset_name, distributed=True, output_dir=output_folder))
        return DatasetEvaluators(evaluator_list)

    @classmethod
    def build_train_loader(cls, cfg):
        mapper = DatasetMapper(cfg, is_train=True)
        return build_detection_train_loader(cfg, mapper=mapper)

    @classmethod
    def build_test_loader(cls, cfg, dataset_name):
        mapper = DatasetMapper(cfg, is_train=False)
        # mapper = TrajectoryPredictionDatasetMapper(cfg, is_train=False)
        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 torch.optim.lr_scheduler.ExponentialLR(optimizer, 0.9999839709739701)  # ** 3200 = 0.95
        return build_lr_scheduler(cfg, optimizer)

    @classmethod
    def build_optimizer(cls, cfg, model):
        # return torch.optim.Adam(model.parameters(), lr=cfg.SOLVER.BASE_LR*cfg.SOLVER.BACKBONE_MULTIPLIER)
        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 "sidecar_bone" in module_name:
                    hyperparams["lr"] = 0.00001
                if "interfere_bone" in module_name:
                    hyperparams["lr"] = 0.00001
                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(cls, cfg, model, evaluators=None):
        if not cfg.SOLVER.AMP.ENABLED:
            results = super(Trainer, cls).test(cfg, model, evaluators)
        else:
            with torch.cuda.amp.autocast(enabled=True):
                results = super(Trainer, cls).test(cfg, model, evaluators)
        return results

    @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 = ModelWithTTA(cfg, model)
        evaluators = [
            cls.build_evaluator(
                cfg, name, output_folder=os.path.join(cfg.OUTPUT_DIR, "inference_TTA")
            )
            for name in cfg.DATASETS.TEST
        ]
        if not cfg.SOLVER.AMP.ENABLED:
            res = cls.test(cfg, model, evaluators)
        else:
            with torch.cuda.amp.autocast(enabled=True):
                res = cls.test(cfg, model, evaluators)
        res = OrderedDict({k + "_TTA": v for k, v in res.items()})
        return res

    def build_hooks(self):
        ret = super(Trainer, self).build_hooks()
        cfg = self.cfg.clone()
        if comm.is_main_process():
            ret.append(hooks.BestCheckpointer(cfg.TEST.EVAL_PERIOD, self.checkpointer, "metrics"))
        return ret


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,),
    )