train_sf-yolo.py

# YOLOv5 🚀 by Ultralytics, AGPL-3.0 license
"""
Train a YOLOv5 model on a custom dataset.
Models and datasets download automatically from the latest YOLOv5 release.

Usage - Single-GPU training:
    $ python train.py --data coco128.yaml --weights yolov5s.pt --img 640  # from pretrained (recommended)
    $ python train.py --data coco128.yaml --weights '' --cfg yolov5s.yaml --img 640  # from scratch

Usage - Multi-GPU DDP training:
    $ python -m torch.distributed.run --nproc_per_node 4 --master_port 1 train.py --data coco128.yaml --weights yolov5s.pt --img 640 --device 0,1,2,3

Models:     https://github.com/ultralytics/yolov5/tree/master/models
Datasets:   https://github.com/ultralytics/yolov5/tree/master/data
Tutorial:   https://docs.ultralytics.com/yolov5/tutorials/train_custom_data
"""

import argparse
import math
import os
import random
import subprocess
import sys
import time
from copy import deepcopy
from datetime import datetime, timedelta
from pathlib import Path

try:
    import comet_ml  # must be imported before torch (if installed)
except ImportError:
    comet_ml = None

import numpy as np
import torch
import torch.distributed as dist
import torch.nn as nn
import yaml
from torch.optim import lr_scheduler
from tqdm import tqdm

FILE = Path(__file__).resolve()
ROOT = FILE.parents[0]  # YOLOv5 root directory
if str(ROOT) not in sys.path:
    sys.path.append(str(ROOT))  # add ROOT to PATH
ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative

import val as validate  # for end-of-epoch mAP
from models.experimental import attempt_load, load_model_weights
from models.yolo import Model
from utils.autoanchor import check_anchors
from utils.autobatch import check_train_batch_size
from utils.callbacks import Callbacks
from utils.dataloaders import create_dataloader
from utils.downloads import attempt_download, is_url
from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_git_info,
                           check_git_status, check_img_size, check_requirements, check_suffix, check_yaml, colorstr,
                           get_latest_run, increment_path, init_seeds, intersect_dicts, labels_to_class_weights,
                           labels_to_image_weights, methods, one_cycle, print_args, print_mutation, strip_optimizer,
                           yaml_save, non_max_suppression, xyxy2xywhn)
from utils.loggers import Loggers
from utils.loggers.comet.comet_utils import check_comet_resume
from utils.loss import ComputeLoss
from utils.metrics import fitness
from utils.plots import plot_evolve, plot_images
from utils.torch_utils import (EarlyStopping, ModelEMA, WeightEMA, de_parallel, select_device, smart_DDP, smart_optimizer,
                               smart_resume, torch_distributed_zero_first)
from TargetAugment.enhance_style import get_style_images
from TargetAugment.enhance_vgg16 import enhance_vgg16


LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
RANK = int(os.getenv('RANK', -1))
WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
GIT_INFO = check_git_info()

# Avoid excessive CPU usage on shared servers by limiting the number of threads
N_CORE = "16"
N_THREADS_TORCH = 16
os.environ["OMP_NUM_THREADS"] = N_CORE
os.environ["OPENBLAS_NUM_THREADS"] = N_CORE
os.environ["MKL_NUM_THREADS"] = N_CORE
os.environ["VECLIB_MAXIMUM_THREADS"] = N_CORE
os.environ["NUMEXPR_NUM_THREADS"] = N_CORE
torch.set_num_threads(N_THREADS_TORCH)

        
def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \
        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
    
    # Hyperparameters
    if isinstance(hyp, str):
        with open(hyp, errors='ignore') as f:
            hyp = yaml.safe_load(f)  # load hyps dict
    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
    
    teacher_alpha, conf_thres, iou_thres, max_gt_boxes,style_path, SSM_alpha = \
        opt.teacher_alpha, opt.conf_thres, opt.iou_thres, opt.max_gt_boxes, \
        opt.style_path, opt.SSM_alpha
        
    callbacks.run('on_pretrain_routine_start')

    # Directories
    w = save_dir / 'weights'  # weights dir
    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
    last_student, best_student = w / 'last_student.pt', w / 'best_student.pt'
    last_teacher, best_teacher = w / 'last_teacher.pt', w / 'best_teacher.pt'

    opt.random_style = style_path == ""
    opt.cuda = True
    opt.log_dir = save_dir / './enhance_style_samples'    
    
    # Save run settings
    if not evolve:
        yaml_save(save_dir / 'hyp.yaml', hyp)
        yaml_save(save_dir / 'opt.yaml', vars(opt))

    # Loggers
    data_dict = None
    if RANK in {-1, 0}:
        loggers = Loggers(save_dir, weights, opt, hyp, LOGGER)  # loggers instance

        # Register actions
        for k in methods(loggers):
            callbacks.register_action(k, callback=getattr(loggers, k))

        # Process custom dataset artifact link
        data_dict = loggers.remote_dataset
        if resume:  # If resuming runs from remote artifact
            weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size
    
    # Config
    plots = not evolve and not opt.noplots  # create plots
    cuda = device.type != 'cpu'
    init_seeds(opt.seed + 1 + RANK, deterministic=True)
    with torch_distributed_zero_first(LOCAL_RANK):
        data_dict = data_dict or check_dataset(data)  # check if None
    train_path, val_path = data_dict['train'], data_dict['val']
    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
    is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset

    # Model
    check_suffix(weights, '.pt')  # check weights
    pretrained = weights.endswith('.pt')
    if pretrained:
        with torch_distributed_zero_first(LOCAL_RANK):
            weights = attempt_download(weights)  # download if not found locally
        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
        model_student = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create the student model
        model_teacher = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create the teacher model
        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
        csd = intersect_dicts(csd, model_student.state_dict(), exclude=exclude)  # intersect
        model_student.load_state_dict(csd, strict=False)  # load student state_dict
        model_teacher.load_state_dict(csd.copy(), strict=False)  # load teacher state_dict
        LOGGER.info(f'Transferred {len(csd)}/{len(model_student.state_dict())} items from {weights}')  # report
    else:
        model_student = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create the student model
        model_teacher = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create the teacher model
    # amp = check_amp(model_student)  # check AMP
    amp = False

    # Freeze
    freeze_student = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
    for k, v in model_student.named_parameters():
        v.requires_grad = True  # train all layers
        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
        if any(x in k for x in freeze_student):
            LOGGER.info(f'freezing student {k}')
            v.requires_grad = False
            
    freeze_teacher = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
    for k, v in model_teacher.named_parameters():
        v.requires_grad = True  # train all layers
        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
        if any(x in k for x in freeze_teacher):
            LOGGER.info(f'freezing student {k}')
            v.requires_grad = False

    # Image size
    gs = max(int(model_student.stride.max()), 32)  # grid size (max stride)
    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple

    # Batch size
    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
        batch_size = check_train_batch_size(model_student, imgsz, amp)
        loggers.on_params_update({'batch_size': batch_size})

    # Optimizer
    nbs = 64  # nominal batch size
    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
    optimizer_student = smart_optimizer(model_student, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])

    # Unbiased Mean Teacher [github.com/kinredon/umt]
    def get_params(model, requires_grad=True, set_requires_grad=None):
        params = [value for key, value in model.named_parameters() if value.requires_grad == requires_grad]
        if set_requires_grad is not None:
            for param in params:
                param.requires_grad = set_requires_grad
        return params

    student_detection_params = get_params(model_student)
    teacher_detection_params = get_params(model_teacher, set_requires_grad=False)
    optimizer_teacher = WeightEMA(teacher_detection_params, student_detection_params, alpha=teacher_alpha)

    if opt.cos_lr:
        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
        scheduler = lr_scheduler.LambdaLR(optimizer_student, lr_lambda=lf) 
    elif opt.CosineAnnealingLR:
        lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp['lrf']) + hyp['lrf'] 
        scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer_student, T_max=10, eta_min=hyp['lr0'] * hyp['lrf'], last_epoch=-1)
    elif opt.CosineAnnealingWarmRestarts:
        lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp['lrf']) + hyp['lrf']  
        scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer_student, T_0=20, T_mult=2, eta_min=hyp['lr0'] * hyp['lrf'], last_epoch=-1)
    else:
        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  
        scheduler = lr_scheduler.LambdaLR(optimizer_student, lr_lambda=lf)  

    # EMA
    ema = ModelEMA(model_student) if RANK in {-1, 0} else None

    # Resume
    best_fitness, start_epoch = 0.0, 0
    if pretrained:
        if resume:
            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer_student, ema, weights, epochs, resume)
        del ckpt, csd

    # DP mode
    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
        LOGGER.warning(
            'WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.\n'
            'See Multi-GPU Tutorial at https://docs.ultralytics.com/yolov5/tutorials/multi_gpu_training to get started.'
        )
        model_student = torch.nn.DataParallel(model_student)
        model_teacher = torch.nn.DataParallel(model_teacher)

    # SyncBatchNorm
    if opt.sync_bn and cuda and RANK != -1:
        model_student = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model_student).to(device)
        model_teacher = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model_teacher).to(device)
        LOGGER.info('Using SyncBatchNorm()')

    # Trainloader
    train_loader, dataset = create_dataloader(train_path,
                                              imgsz,
                                              batch_size // WORLD_SIZE,
                                              gs,
                                              single_cls,
                                              hyp=hyp,
                                              augment=True,
                                              cache=None if opt.cache == 'val' else opt.cache,
                                              rect=opt.rect,
                                              rank=LOCAL_RANK,
                                              workers=workers,
                                              image_weights=opt.image_weights,
                                              quad=opt.quad,
                                              prefix=colorstr('train: '),
                                              shuffle=True,
                                              seed=opt.seed)
    labels = np.concatenate(dataset.labels, 0)
    mlc = int(labels[:, 0].max())  # max label class
    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'

    # Process 0
    if RANK in {-1, 0}:
        val_loader = create_dataloader(val_path,
                                       imgsz,
                                       batch_size // WORLD_SIZE * 2,
                                       gs,
                                       single_cls,
                                       hyp=hyp,
                                       cache=None if noval else opt.cache,
                                       rect=True,
                                       rank=-1,
                                       workers=workers * 2,
                                       pad=0.5,
                                       prefix=colorstr('val: '))[0]

        if not resume:
            if not opt.noautoanchor:
                check_anchors(dataset, model=model_student, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor student
                check_anchors(dataset, model=model_teacher, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor teacher
            model_student.half().float()  # pre-reduce anchor precision
            model_teacher.half().float()  # pre-reduce anchor precision

        callbacks.run('on_pretrain_routine_end', labels, names)

    # DDP mode
    if cuda and RANK != -1:
        model_student = smart_DDP(model_student)
        model_teacher = smart_DDP(model_teacher)

    # Model attributes
    nl = de_parallel(model_student).model[-1].nl  # number of detection layers (to scale hyps)
    hyp['box'] *= 3 / nl  # scale to layers
    hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
    hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
    hyp['label_smoothing'] = opt.label_smoothing
    model_student.nc = nc  # attach number of classes to model
    model_student.hyp = hyp  # attach hyperparameters to model
    model_student.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
    model_student.names = names
    model_teacher.nc = nc  # attach number of classes to model
    model_teacher.hyp = hyp  # attach hyperparameters to model
    model_teacher.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
    model_teacher.names = names

    # Start training
    t0 = time.time()
    nb = len(train_loader)  # number of batches
    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
    last_opt_step = -1
    maps = np.zeros(nc)  # mAP per class
    results = (0, 0, 0, 0, 0, 0, 0)  # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls)
    scheduler.last_epoch = start_epoch - 1  # do not move
    scaler = torch.cuda.amp.GradScaler(enabled=amp)
    stopper, stop = EarlyStopping(patience=opt.patience), False
    compute_loss = ComputeLoss(model_student)  # init loss class
    callbacks.run('on_train_start')
    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
                f"Logging results to {colorstr('bold', save_dir)}\n"
                f'Starting training for {epochs} epochs...')
                # if ni % (nb // 4) == 0 and ni!=nb:
    
    adain = enhance_vgg16(opt)
    
    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
        callbacks.run('on_train_epoch_start')
        model_student.train()

        # Update image weights (optional, single-GPU only)
        if opt.image_weights:
            cw = model_student.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx

        # Update mosaic border (optional)
        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders

        mloss = torch.zeros(3, device=device)  # mean losses
        # mloss = torch.zeros(4, device=device)  # mean losses
        if RANK != -1:
            train_loader.sampler.set_epoch(epoch)
        pbar = enumerate(train_loader)
        # LOGGER.info(('\n' + '%11s' * 8) % ('Epoch', 'GPU_mem', 'box_loss', 'obj_loss', 'cls_loss', 'gw_g_loss','Instances', 'Size'))
        LOGGER.info(('\n' + '%11s' * 7) % ('Epoch', 'GPU_mem', 'box_loss', 'obj_loss', 'cls_loss', 'Instances', 'Size'))
        if RANK in {-1, 0}:
            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
        optimizer_student.zero_grad()
        
        if SSM_alpha != 0.0:
            if epoch != start_epoch:
                # Get student and teacher state_dicts
                student_state_dict = model_student.state_dict()
                teacher_state_dict = model_teacher.state_dict()

                # Update student weights with EMA of teacher weights
                for name, param in student_state_dict.items():
                    if name in teacher_state_dict:
                        param.data.copy_((1.0 - SSM_alpha) * param.data + SSM_alpha * teacher_state_dict[name].data)
                LOGGER.info(f'Transferred teacher weights to student model at the start of epoch {epoch}, alpha={SSM_alpha}')
            
        for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------
            callbacks.run('on_train_batch_start')
            opt.imgs_paths = paths # path for target augment module
            ni = i + nb * epoch  # number integrated batches (since train start)
            imgs_255 = imgs.clone().to(torch.float32).to(device, non_blocking=True) 
            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0

            # Warmup
            if ni <= nw:
                xi = [0, nw]  # x interp
                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
                for j, x in enumerate(optimizer_student.param_groups):
                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
                    if 'momentum' in x:
                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])

            # Multi-scale
            if opt.multi_scale:
                sz = random.randrange(int(imgsz * 0.5), int(imgsz * 1.5) + gs) // gs * gs  # size
                sf = sz / max(imgs.shape[2:])  # scale factor
                if sf != 1:
                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)

            # Forward
            with torch.cuda.amp.autocast(amp):
                imgs_style = get_style_images(imgs_255, opt, adain) / 255
                    
                # Teacher forward 
                model_teacher.eval()
                pred_teacher, _ = model_teacher(imgs)  # inference on teacher to obtain pseudo labels
                pred_teacher_nms = non_max_suppression(pred_teacher, conf_thres=conf_thres, iou_thres=iou_thres, multi_label=True, agnostic=single_cls, max_det=max_gt_boxes)
                            
                # Student forward
                model_student.zero_grad()
                pred_student = model_student(imgs_style)  # forward
                    
                # Generate pseudo labels using the teacher's prediction in order to obtain student loss
                batch_size, ch, img_height, img_width = imgs.shape
                pseudo_labels_batch_nms = []

                # Loop over each image in the batch
                for image_id in range(batch_size):
                    if pred_teacher_nms[image_id].shape[0]:
                        # Remove the confidence of the predicted tensor -> [cls x y x y]
                        label_data = torch.cat((pred_teacher_nms[image_id][:, 5].unsqueeze(-1), pred_teacher_nms[image_id][:, :4]), dim=1)
                
                        # Convert bounding box format from xyxy to normalized xywh
                        label_data[:, 1:5] = xyxy2xywhn(label_data[:, 1:5], w=img_width, h=img_height)
                        
                        # Concatenate the image ids (within the current batch) with the pseudo-labels list
                        pseudo_label = torch.cat((torch.full((label_data.shape[0], 1), image_id, device=device), label_data), dim=1)
                        
                        pseudo_labels_batch_nms.append(pseudo_label)

                # Concatenate all labels, or create a default label if there are none
                pseudo_labels_final = torch.cat(pseudo_labels_batch_nms, dim=0) if pseudo_labels_batch_nms else torch.tensor([[0, 0, 0.5, 0.7, random.uniform(0.3, 0.9), random.uniform(0.3, 0.9)]], device=device)
    
                # Compute loss using teacher's prediction as pseudo labels
                loss, loss_items = compute_loss(pred_student, pseudo_labels_final.to(device), model_teacher, model_student)  # loss 
                
                if RANK != -1:
                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
                if opt.quad:
                    loss *= 4.

            # Backward
            scaler.scale(loss).backward()

            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
            if ni - last_opt_step >= accumulate:
                scaler.unscale_(optimizer_student)  # unscale gradients
                torch.nn.utils.clip_grad_norm_(model_student.parameters(), max_norm=10.0)  # clip gradients
                scaler.step(optimizer_student)  # optimizer.step
                scaler.update()
                optimizer_student.zero_grad()
                if ema:
                    ema.update(model_student)
                last_opt_step = ni

                model_teacher.train()
                model_teacher.zero_grad()
                optimizer_teacher.step()

            # Log
            if RANK in {-1, 0}:
                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
                # pbar.set_description(('%11s' * 2 + '%11.4g' * 6) %
                #                      (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
                pbar.set_description(('%11s' * 2 + '%11.4g' * 5) %
                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
                # Plot                   
                callbacks.run('on_train_batch_end', model_student, ni, imgs, targets, paths, list(mloss), epoch, pseudo_labels_final, names, pbar)
                if callbacks.stop_training:
                    return

        # Scheduler
        lr = [x['lr'] for x in optimizer_student.param_groups]  # for loggers
        scheduler.step()

        if RANK in {-1, 0}:
            # mAP
            callbacks.run('on_train_epoch_end', epoch=epoch)
            ema.update_attr(model_student, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
            if not noval or final_epoch:  # Calculate mAP
                results, maps, _ = validate.run(data_dict,
                                                batch_size=batch_size // WORLD_SIZE * 2,
                                                imgsz=imgsz,
                                                half=amp,
                                                model=model_teacher,
                                                single_cls=single_cls,
                                                dataloader=val_loader,
                                                save_dir=save_dir,
                                                plots=False,
                                                callbacks=callbacks,
                                                compute_loss=compute_loss,
                                                verbose=False)

            # Update best mAP
            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, mAP@.5, mAP@.5-.95]
            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
            if fi > best_fitness:
                best_fitness = fi
            log_vals = list(mloss) + list(results) + lr
            callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)

            # Save model
            if (not nosave) or (final_epoch and not evolve):  # if save
                ckpt_student = {
                    'epoch': epoch,
                    'best_fitness': best_fitness,
                    'model': deepcopy(de_parallel(model_student)).half(),
                    'ema': deepcopy(ema.ema).half(),
                    'updates': ema.updates,
                    'optimizer': optimizer_student.state_dict(),
                    'opt': vars(opt),
                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
                    'date': datetime.now().isoformat()}
                
                ckpt_teacher = {
                    'epoch': epoch,
                    'best_fitness': best_fitness,
                    'model': deepcopy(de_parallel(model_teacher)).half(),
                    'opt': vars(opt),
                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
                    'date': datetime.now().isoformat()}

                # Save last, best and delete
                torch.save(ckpt_teacher, last_teacher)
                if best_fitness == fi:
                    torch.save(ckpt_teacher, best_teacher)
                if opt.save_period > 0 and epoch % opt.save_period == 0:
                    torch.save(ckpt_teacher, w / f'epoch{epoch}_teacher.pt')
                del ckpt_student, ckpt_teacher
                callbacks.run('on_model_save', last_teacher, epoch, final_epoch, best_fitness, fi)

        # EarlyStopping
        if RANK != -1:  # if DDP training
            broadcast_list = [stop if RANK == 0 else None]
            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
            if RANK != 0:
                stop = broadcast_list[0]
        if stop:
            break  # must break all DDP ranks

        # end epoch ----------------------------------------------------------------------------------------------------
        # optimizer_teacher.step() 
    # end training -----------------------------------------------------------------------------------------------------
    if RANK in {-1, 0}:
        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
        for f in last_teacher, best_teacher:
            if f.exists():
                strip_optimizer(f)  # strip optimizers
                if f is best_teacher:
                    LOGGER.info(f'\nValidating {f}...')
                    results, _, _ = validate.run(
                        data_dict,
                        batch_size=batch_size // WORLD_SIZE * 2,
                        imgsz=imgsz,
                        model=attempt_load(f, device).half(),
                        iou_thres=0.65 if is_coco else 0.60,  # best pycocotools at iou 0.65
                        single_cls=single_cls,
                        dataloader=val_loader,
                        save_dir=save_dir,
                        save_json=is_coco,
                        verbose=True,
                        plots=plots,
                        callbacks=callbacks,
                        compute_loss=compute_loss)  # val best model with plots
                    if is_coco:
                        callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)

        callbacks.run('on_train_end', last_teacher, best_teacher, epoch, results)


    torch.cuda.empty_cache()
    return results


def parse_opt(known=False):
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='initial weights path')
    parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path')
    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=960, help='train, val image size (pixels)')
    parser.add_argument('--rect', action='store_true', help='rectangular training')
    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
    parser.add_argument('--noplots', action='store_true', help='save no plot files')
    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
    parser.add_argument('--device', default='0', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW'], default='SGD', help='optimizer')
    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
    parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
    parser.add_argument('--name', default='exp', help='save to project/name')
    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
    parser.add_argument('--quad', action='store_true', help='quad dataloader')
    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
    parser.add_argument('--CosineAnnealingLR', action='store_true', help='CosineAnnealingLR scheduler')
    parser.add_argument('--CosineAnnealingWarmRestarts', action='store_true', help='CosineAnnealingWarmRestarts  scheduler')
    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
    parser.add_argument('--conf_thres', type=float, default=0.4, help='Confidence threshold for pseudo labels generation by the teacher model')
    parser.add_argument('--iou_thres', type=float, default=0.3, help='IoU threshold for pseudo labels NMS by the teacher model')
    parser.add_argument('--max_gt_boxes', type=int, default=20, help='Maximum number of ground truth boxes for pseudo labels generation by the teacher model')
    parser.add_argument('--teacher_alpha', type=float, default=0.999, help='Teacher alpha for MT EMA')
    parser.add_argument('--plot_save_interval', type=int, default=100, help='Interval of epoch to save plots (useful to debug)')

    # Logger arguments
    parser.add_argument('--entity', default=None, help='Entity')
    parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='Upload data, "val" option')
    parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval')
    parser.add_argument('--artifact_alias', type=str, default='latest', help='Version of dataset artifact to use')

    parser.add_argument('--decoder_path', type=str, help='Decoder path')
    parser.add_argument('--encoder_path', type=str, help='Encoder path')
    parser.add_argument('--fc1', type=str, help='fc1 path')
    parser.add_argument('--fc2', type=str, help='fc2 path')
    parser.add_argument('--style_path', type=str, default="", help='Path to the style image, if not specified, random style will be used')
    parser.add_argument('--style_add_alpha', type=float, default=1, help='The amount of style to add to the image (between 0 and 1)')        
    parser.add_argument('--save_style_samples', action='store_true', help='Save style samples images (useful to debug)')
    parser.add_argument('--SSM_alpha', type=float, default=0.0, help='SSM momentum')
    
    return parser.parse_known_args()[0] if known else parser.parse_args()


def main(opt, callbacks=Callbacks()):
    # Checks
    if RANK in {-1, 0}:
        print_args(vars(opt))
        check_git_status()
        check_requirements(ROOT / 'requirements.txt')

    # Resume (from specified or most recent last.pt)
    if opt.resume and not check_comet_resume(opt) and not opt.evolve:
        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
        opt_data = opt.data  # original dataset
        if opt_yaml.is_file():
            with open(opt_yaml, errors='ignore') as f:
                d = yaml.safe_load(f)
        else:
            d = torch.load(last, map_location='cpu')['opt']
        opt = argparse.Namespace(**d)  # replace
        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
        if is_url(opt_data):
            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
    else:
        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
        if opt.evolve:
            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
                opt.project = str(ROOT / 'runs/evolve')
            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
        if opt.name == 'cfg':
            opt.name = Path(opt.cfg).stem  # use model.yaml as name
        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))

    # DDP mode
    device = select_device(opt.device, batch_size=opt.batch_size)
    if LOCAL_RANK != -1:
        msg = 'is not compatible with YOLOv5 Multi-GPU DDP training'
        assert not opt.image_weights, f'--image-weights {msg}'
        assert not opt.evolve, f'--evolve {msg}'
        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
        torch.cuda.set_device(LOCAL_RANK)
        device = torch.device('cuda', LOCAL_RANK)
        dist.init_process_group(backend='nccl' if dist.is_nccl_available() else 'gloo',
                                timeout=timedelta(seconds=10800))

    # Train
    if not opt.evolve:
        train(opt.hyp, opt, device, callbacks)

    # Evolve hyperparameters (optional)
    else:
        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
        meta = {         
            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
            'momentum': (1, 0.6, 0.98),  # SGD momentum/Adam beta1
            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
            'warmup_epochs': (0, 0.0, 5.0),  # warmup epochs (fractions ok)
            'warmup_momentum': (0, 0.0, 0.95),  # warmup initial momentum
            'warmup_bias_lr': (0, 0.0, 0.2),  # warmup initial bias lr
            'box': (0, 0.02, 0.2),  # box loss gain
            'cls': (0, 0.2, 4.0),  # cls loss gain
            'cls_pw': (0, 0.5, 2.0),  # cls BCELoss positive_weight
            'obj': (0, 0.2, 4.0),  # obj loss gain (scale with pixels)
            'obj_pw': (0, 0.5, 2.0),  # obj BCELoss positive_weight
            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
            'anchor_t': (0, 2.0, 8.0),  # anchor-multiple threshold
            # 'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
            'hsv_h': (0, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
            'hsv_s': (0, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
            'hsv_v': (0, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
            'degrees': (0, 0.0, 45.0),  # image rotation (+/- deg)
            'translate': (0, 0.0, 0.9),  # image translation (+/- fraction)
            'scale': (0, 0.0, 0.9),  # image scale (+/- gain)
            'shear': (0, 0.0, 10.0),  # image shear (+/- deg)
            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
            'flipud': (0, 0.0, 1.0),  # image flip up-down (probability)
            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
            'mosaic': (0, 0.0, 1.0),  # image mixup (probability)
            'mixup': (0, 0.0, 1.0),  # image mixup (probability)
            'copy_paste': (0, 0.0, 1.0)}  # segment copy-paste (probability)

        with open(opt.hyp, errors='ignore') as f:
            hyp = yaml.safe_load(f)  # load hyps dict
        #     if 'anchors' not in hyp:  # anchors commented in hyp.yaml
        #         hyp['anchors'] = 3
        # if opt.noautoanchor:
        #     del hyp['anchors'], meta['anchors']
        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
        if opt.bucket:
            # download evolve.csv if exists
            subprocess.run([
                'gsutil',
                'cp',
                f'gs://{opt.bucket}/evolve.csv',
                str(evolve_csv), ])

        for _ in range(opt.evolve):  # generations to evolve
            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
                # Select parent(s)
                parent = 'single'  # parent selection method: 'single' or 'weighted'
                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
                print(x)
                n = min(5, len(x))  # number of previous results to consider
                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
                if parent == 'single' or len(x) == 1:
                    # x = x[random.randint(0, n - 1)]  # random selection
                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
                elif parent == 'weighted':
                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination

                # Mutate
                mp, s = 0.8, 0.2  # mutation probability, sigma
                npr = np.random
                npr.seed(int(time.time()))
                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
                ng = len(meta)
                v = np.ones(ng)
                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
                    # print(i, k)
                    hyp[k] = float(x[i + 7] * v[i])  # mutate

            # Constrain to limits
            for k, v in meta.items():
                hyp[k] = max(hyp[k], v[1])  # lower limit
                hyp[k] = min(hyp[k], v[2])  # upper limit
                hyp[k] = round(hyp[k], 5)  # significant digits

            # Train mutation
            results = train(hyp.copy(), opt, device, callbacks)
            callbacks = Callbacks()
            # Write mutation results
            keys = ('metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss',
                    'val/obj_loss', 'val/cls_loss')
            print_mutation(keys, results, hyp.copy(), save_dir, opt.bucket)

        # Plot results
        plot_evolve(evolve_csv)
        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
                    f"Results saved to {colorstr('bold', save_dir)}\n"
                    f'Usage example: $ python train.py --hyp {evolve_yaml}')


def run(**kwargs):
    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolov5m.pt')
    opt = parse_opt(True)
    for k, v in kwargs.items():
        setattr(opt, k, v)
    main(opt)
    return opt


if __name__ == '__main__':
    opt = parse_opt()
    main(opt)