import random
from pathlib import Path
from importlib.util import spec_from_file_location, module_from_spec

################################################################################
# Objects365 V2 · 365-class · YOLOv8-X Config (MMYOLO 0.6+)   2025-06-27
################################################################################

# ────── A. 데이터 경로 ───────────────────────────────────────────────
DATA_ROOT   = '/data/ltb/ultralytics/objects365_mmyolo_coco'
TRAIN_JSON  = f'{DATA_ROOT}/train_coco.json'
VAL_JSON    = f'{DATA_ROOT}/val_coco.json'
LOAD_FROM   = None

# 사전 학습 weights 로딩 (없으면 None)
load_from = LOAD_FROM

# ────── B. 학습 관련 ───────────────────────────────────────────────
EPOCHS              = 100
SWITCH_EPOCH        = 80
CLOSE_MOSAIC_EPOCHS = 10

# ────── C. 배치 사이즈 ─────────────────────────────────────────────
TRAIN_BATCH_SIZE = 32
VAL_BATCH_SIZE   = 32
TEST_BATCH_SIZE  = 32

# ────── D. 하이퍼파라미터 ───────────────────────────────────────────
BASE_LR      = 0.00125
MIXUP_PROB   = 0.15
WEIGHT_DECAY = 0.0005
MOMENTUM     = 0.937
ALBU_PROB    = 0.01

# ────── E. DataLoader 워커 수 ───────────────────────────────────────
NUM_WORKERS_TRAIN = 32
NUM_WORKERS_VAL   = 32
NUM_WORKERS_TEST  = 32

# ────── F. 클래스·팔레트 로딩 ───────────────────────────────────────
CLASSES_PY = Path('/data/ltb/mmyolo/objects365_classes.py')
if not CLASSES_PY.exists():
    raise FileNotFoundError(f'{CLASSES_PY} not found')
spec = spec_from_file_location('obj365', CLASSES_PY)
obj365 = module_from_spec(spec)
spec.loader.exec_module(obj365)

CLASS_NAMES = tuple(obj365.OBJECTS365_CLASSES)
NUM_CLASSES  = len(CLASS_NAMES)
assert NUM_CLASSES == 365, f'Expected 365 classes, got {NUM_CLASSES}'

# 고정 팔레트(시각화용)
if hasattr(obj365, 'PALETTE'):
    PALETTE = list(obj365.PALETTE)
else:
    random.seed(42)
    PALETTE = [tuple(random.randint(0,255) for _ in range(3))
               for _ in range(NUM_CLASSES)]

CLASS_WEIGHTS = [1.0] * NUM_CLASSES

# 모듈 참조 제거
del spec, obj365, module_from_spec, spec_from_file_location, random

# ────── G. Optimizer & Scheduler ───────────────────────────────────────
optim_wrapper = dict(
    type='AmpOptimWrapper',
    constructor='YOLOv5OptimizerConstructor',
    loss_scale='dynamic',
    optimizer=dict(
        type='SGD',
        lr=BASE_LR,
        momentum=MOMENTUM,
        nesterov=True,
        weight_decay=WEIGHT_DECAY,
        batch_size_per_gpu=TRAIN_BATCH_SIZE,
        paramwise_cfg=dict(custom_keys={'backbone': dict(lr_mult=0.1)})
    ),
    clip_grad=dict(max_norm=5.0)
)

param_scheduler = [
    dict(
        type='YOLOv5ParamSchedulerHook',
        scheduler_type='linear',
        lr_factor=0.01,
        max_epochs=EPOCHS
    ),
    dict(
        type='StepLR',
        by_epoch=True,
        begin=0,
        end=EPOCHS,
        step_size=10,
        gamma=10.0,
        param_group='backbone'
    )
]



# ────── H. Model ────────────────────────────────────────────────────────
model = dict(
    type='YOLODetector',
    data_preprocessor=dict(
        type='YOLOv5DetDataPreprocessor',
        bgr_to_rgb=True, mean=[0.0,0.0,0.0],
        std=[255.0,255.0,255.0]
    ),
    backbone=dict(
        type='YOLOv8CSPDarknet', arch='P5',
        deepen_factor=1.0, widen_factor=1.0,
        norm_cfg=dict(type='BN', momentum=0.03, eps=1e-3),
        act_cfg=dict(type='SiLU', inplace=True)
    ),
    neck=dict(
        type='YOLOv8PAFPN',
        deepen_factor=1.0, widen_factor=1.0,
        in_channels=[256,512,1024], out_channels=[256,512,1024],
        num_csp_blocks=3,
        norm_cfg=dict(type='BN', momentum=0.03, eps=1e-3),
        act_cfg=dict(type='SiLU', inplace=True)
    ),
    bbox_head=dict(
        type='YOLOv8Head',
        head_module=dict(
            type='YOLOv8HeadModule',
            in_channels=[256,512,1024],
            featmap_strides=[8,16,32],
            num_classes=NUM_CLASSES,
            reg_max=16,
            norm_cfg=dict(type='BN', momentum=0.03, eps=1e-3),
            act_cfg=dict(type='SiLU', inplace=True)
        ),
        bbox_coder=dict(type='DistancePointBBoxCoder'),
        loss_bbox=dict(
            type='IoULoss', bbox_format='xywh', iou_mode='ciou',
            loss_weight=7.5, reduction='sum', return_iou=False
        ),
        loss_cls=dict(
            type='CrossEntropyCustomLoss_mmyolo',
            use_sigmoid=True, num_classes=NUM_CLASSES,
            reduction='none', loss_weight=0.75,
            class_weight=CLASS_WEIGHTS
        ),
        loss_dfl=dict(
            type='mmdet.DistributionFocalLoss',
            reduction='mean', loss_weight=0.375
        ),
        prior_generator=dict(
            type='mmdet.MlvlPointGenerator',
            strides=[8,16,32], offset=0.5
        )
    ),
    train_cfg=dict(
        assigner=dict(
            type='BatchTaskAlignedAssigner',
            num_classes=NUM_CLASSES,
            topk=10, alpha=0.5, beta=6.0, eps=1e-9, use_ciou=True
        )
    ),
    test_cfg=dict(
        score_thr=0.01,
        nms_pre=30000, multi_label=True,
        nms=dict(type='nms', iou_threshold=0.5),
        max_per_img=500
    )
)

# ────── I. Hooks & Environment ─────────────────────────────────────────
default_hooks = dict(
    checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=10, save_best='auto'),
    logger=dict(type='LoggerHook', interval=50),
    param_scheduler=dict(type='YOLOv5ParamSchedulerHook'),
    timer=dict(type='IterTimerHook'),
    sampler_seed=dict(type='DistSamplerSeedHook'),
    visualization=dict(type='mmdet.DetVisualizationHook')
)

custom_hooks = [
    dict(type='EMAHook', ema_type='ExpMomentumEMA', momentum=1e-4, priority=49, strict_load=False, update_buffers=True),
    dict(type='mmdet.PipelineSwitchHook', switch_epoch=SWITCH_EPOCH, switch_pipeline='train_pipeline_stage2')
]

env_cfg = dict(cudnn_benchmark=True, dist_cfg=dict(backend='nccl'), mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0))

# ────── J. Train 파이프라인 정의 (초기/Stage1) ───────────────────────────
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='YOLOv5RandomAffine', max_rotate_degree=10.0, max_shear_degree=2.0, scaling_ratio_range=(0.8,1.2), border_val=(114,114,114)),
    dict(type='Mosaic', img_scale=(960,960), pad_val=114.0,
         pre_transform=[dict(type='LoadImageFromFile'), dict(type='LoadAnnotations', with_bbox=True)]),
    dict(type='YOLOv5RandomAffine', border=(-480,-480), border_val=(114,114,114), max_aspect_ratio=100, max_rotate_degree=0.0, max_shear_degree=0.0, scaling_ratio_range=(0.1,1.9)),
    dict(type='mmdet.Albu',
         transforms=[dict(type='Blur', p=ALBU_PROB), dict(type='MedianBlur', p=ALBU_PROB), dict(type='ToGray', p=ALBU_PROB), dict(type='CLAHE', p=ALBU_PROB)],
         bbox_params=dict(format='pascal_voc', label_fields=['gt_bboxes_labels','gt_ignore_flags'], type='BboxParams'),
         keymap=dict(img='image', gt_bboxes='bboxes')),
    dict(type='YOLOv5HSVRandomAug'),
    dict(type='mmdet.RandomFlip', prob=0.5),
    dict(type='mmdet.PackDetInputs', meta_keys=('img_id','img_path','ori_shape','img_shape','flip','flip_direction'))
]

# ────── K. Train DataLoader ────────────────────────────────────────────
train_cfg = dict(
    type='EpochBasedTrainLoop',
    max_epochs=EPOCHS,
    val_interval=1,
    dynamic_intervals=[(499, 1)]
)

train_data_prefix = ''
train_dataloader = dict(
    batch_size=TRAIN_BATCH_SIZE,
    num_workers=NUM_WORKERS_TRAIN,
    persistent_workers=False,
    pin_memory=True,
    sampler=dict(type='DefaultSampler', shuffle=True),
    collate_fn=dict(type='yolov5_collate'),
    dataset=dict(
        type='RepeatDataset', times=3,
        dataset=dict(
            type='YOLOv5CocoDataset',
            data_root=DATA_ROOT,
            ann_file=TRAIN_JSON,
            data_prefix=dict(img='images/'),
            metainfo=dict(classes=CLASS_NAMES, palette=PALETTE),
            pipeline=train_pipeline
        )
    )
)

# ────── L. Validation DataLoader & Evaluator ─────────────────────────
val_cfg = dict(type='ValLoop')
val_data_prefix = ''
val_dataloader = dict(
    batch_size=VAL_BATCH_SIZE,
    num_workers=NUM_WORKERS_VAL,
    persistent_workers=False,
    pin_memory=True,
    sampler=dict(type='DefaultSampler', shuffle=False),
    dataset=dict(
        type='YOLOv5CocoDataset',
        data_root=DATA_ROOT,
        ann_file=VAL_JSON,
        data_prefix=dict(img='images/'),
        metainfo=dict(classes=CLASS_NAMES, palette=PALETTE),
        pipeline=[
            dict(type='LoadImageFromFile'),
            dict(type='YOLOv5KeepRatioResize', scale=(960,960)),
            dict(type='LetterResize', scale=(960,960), allow_scale_up=False, pad_val=dict(img=114)),
            dict(type='mmdet.PackDetInputs', meta_keys=('img_id','img_path','ori_shape','img_shape','scale_factor','pad_param'))
        ],
        test_mode=True
    )
)

val_evaluator = dict(type='mmdet.CocoMetric', ann_file=VAL_JSON, metric='bbox', proposal_nums=(100,200,500))

# ────── M. Test DataLoader & Evaluator ────────────────────────────────
test_cfg = dict(type='TestLoop')
test_dataloader = dict(
    batch_size=TEST_BATCH_SIZE,
    num_workers=NUM_WORKERS_TEST,
    persistent_workers=False,
    pin_memory=True,
    sampler=dict(type='DefaultSampler', shuffle=False),
    dataset=dict(
        type='YOLOv5CocoDataset',
        data_root=DATA_ROOT,
        ann_file=VAL_JSON,
        data_prefix=dict(img='images/'),
        metainfo=dict(classes=CLASS_NAMES, palette=PALETTE),
        pipeline=[
            dict(type='LoadImageFromFile'),
            dict(type='YOLOv5KeepRatioResize', scale=(640,640)),
            dict(type='LetterResize', scale=(640,640), allow_scale_up=False, pad_val=dict(img=114)),
            dict(type='mmdet.PackDetInputs', meta_keys=('img_id','img_path','ori_shape','img_shape','scale_factor','pad_param'))
        ],
        test_mode=True
    )
)

test_evaluator = val_evaluator

# # ────── N. 기타 설정 ─────────────────────────────────────────────────
# train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=EPOCHS, val_interval=1, dynamic_intervals=[(EPOCHS-1,1)])
# val_cfg   = dict(type='ValLoop')
# test_cfg  = dict(type='TestLoop')

# 저장 디렉토리
work_dir = 'work_dirs/250627_yolov8_objects365_test_01'

# 로그 레벨
log_level = 'INFO'