model_to_optimize: nn.Module,

checkpoint_model: nn.Module,

checkpoint_file: str,

exclude: List[str],

) -> None:

dict_to_load = intersect_dicts(

checkpoint_model.float().state_dict(), # Checkpoint state_dict as FP32.

model_to_optimize.state_dict(),

exclude=exclude,

) # Intersect checkpoints.

model_to_optimize.load_state_dict(dict_to_load, strict=False) # load

LOGGER.info(

f'Transferred {len(dict_to_load)}/{len(model_to_optimize.state_dict())} '

f'items from {checkpoint_file}'

) # Report intersection statistics.

if len(dict_to_load) == 0:

weights: str,

device,

phase_name: str,

cfg,

opt,

nc: int,

hyp: Dict[str, Any],

resume: bool,

is_enot_phase: bool,

) -> Tuple[nn.Module, nn.Module, Optional[SearchSpaceModel], Dict[str, Any]]:

pretrained: bool = weights.endswith('.pt')

checkpoint: Optional[Dict[str, Any]] = None

checkpoint_config = cfg

checkpoint_model: Optional[nn.Module] = None

checkpoint_with_ss: bool = False

checkpoint_with_model_from_ss: bool = False

checkpoint_exclude: List[str] = []

if pretrained:

with torch_distributed_zero_first(RANK):

weights = attempt_download(weights) # Download if not found locally.

checkpoint = torch.load(weights, map_location=device) # Load checkpoint.

if 'phase_name' not in checkpoint: # When using checkpoints from ultralytics.

checkpoint['phase_name'] = 'train'

if checkpoint['phase_name'] == 'search':

raise RuntimeError('You should not use search phase checkpoints')

if 'ema' in checkpoint and checkpoint['ema'] is not None and not resume:

raise RuntimeError(

'To prevent hard-to-find bugs, we allow using only finalized checkpoints '

'for fine-tuning from pretrained weights\n'

'Finalized checkpoints can be obtained by calling "strip_optimizer" function '

'(BE CAREFUL, IT OVERWRITES YOUR CHECKPOINT FILE BY DEFAULT). This function '

'is automatically applied to your checkpoints at the end of training).'

)

checkpoint_model = checkpoint['model']

checkpoint_with_ss = isinstance(checkpoint_model, SearchSpaceModel)

checkpoint_with_model_from_ss = any(

isinstance(layer, SearchVariantsContainer)

for layer in checkpoint_model.modules()

) and not checkpoint_with_ss

checkpoint_config = checkpoint_model.original_model.yaml if checkpoint_with_ss else checkpoint_model.yaml

checkpoint_exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else [] # exclude keys

model_config = cfg or checkpoint_config

model = Model(model_config, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create

model_to_optimize: Union[nn.Module, SearchSpaceModel] = model

search_space: Optional[SearchSpaceModel] = None

if any(isinstance(layer, SearchVariantsContainer) for layer in model.modules()):

model_to_optimize = search_space = SearchSpaceModel(model).to(device)

# Verify that search space is provided for ENOT phases.

if is_enot_phase and search_space is None:

raise ValueError(

f'"{model_config}" model yaml file does not contain search blocks, '

f'but this is required for phase "{phase_name}"'

)

# Check whether we should extract single architecture from the search space or not.

has_architecture_indices = opt.architecture_indices != '[]'

extract_single_architecture = False

if is_enot_phase or search_space is None:

if has_architecture_indices:

raise ValueError(

f'--architecture-indices commandline argument should only be provided '

f'when training single model from search space.\n'

f'--architecture-indices should not be set for phase "{phase_name}" '

f'and "{model_config}" model yaml file.'

)

else:

if has_architecture_indices:

extract_single_architecture = True

else:

raise ValueError(

f'Training or tuning models from a search space requires setting '

f'--architecture-indices commandline argument.\n'

f'No architecture indices were provided for phase "{phase_name}" '

f'and "{model_config}" model yaml file.'

)

# Check that we can restore checkpoint weights in the first place.

if pretrained:

if is_enot_phase and not checkpoint_with_ss:

# We can only apply weights from SearchSpaceModel in ENOT phases (pretrain and search).

raise ValueError(

f'ENOT phase "{phase_name}" requires checkpoint with SearchSpaceModel, '

f'but checkpoint "{weights}" has a model with type {type(checkpoint_model)}.'

)

if not is_enot_phase:

if (search_space is not None) and (not checkpoint_with_ss) and (not checkpoint_with_model_from_ss):

# New model has search variant containers, but checkpoint contains

# regular model without search variants.

raise ValueError(

f'"{model_config}" model yaml file contains search variants, '

f'but the model from checkpoint "{weights}" is neither a search '

f'space instance nor an extracted model from a search space'

)

if (search_space is None) and (checkpoint_with_ss or checkpoint_with_model_from_ss):

# New model has no search variant containers, but checkpoint contains

# either a search space or a model from the search space.

raise ValueError(

f'"{model_config}" model yaml file contains no search variants, '

f'but the model from checkpoint "{weights}" is either a search '

f'space instance or an extracted model from a search space'

)

# For almost all cases, we can load model from our checkpoint to newly created model.

# The only exception is loading weights from a single model from the search space.

if pretrained and not checkpoint_with_model_from_ss:

set_checkpoint(model_to_optimize, checkpoint_model, weights, checkpoint_exclude)

if extract_single_architecture:

extracted_model = search_space.get_network_by_indexes(json.loads(opt.architecture_indices))

model = model_to_optimize = extracted_model

search_space = None

# Loading weights from a single model from the search space.

if checkpoint_with_model_from_ss:

set_checkpoint(model_to_optimize, checkpoint_model, weights, checkpoint_exclude)

# Optionally prune search space to remove rare or unused operations.

# from utils.prune_search_space_variants import prune_search_space_variants

# prune_indices = [[ 0, 2, 3, 7],

# [ 2, 4, 7, 9],

# [ 1, 7, 8, 10],

# [ 0, 1, 6, 9],

# [ 6, 7, 8, 9],

# [ 5, 6, 9, 10],

# [ 0, 3, 6, 8],

# [ 3, 5, 6, 9]]

# prune_search_space_variants(search_space, prune_indices=prune_indices)

opt,

device,

callbacks=Callbacks()

):

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

# This function prepares data from the dataloader to pass input images to the model during batch norm tuning.

def preprocess_data(x: Tuple[torch.Tensor]) -> Tuple[Tuple[torch.Tensor], Dict]:

return (x[0].to(device).float() / 255.0, ), {}

phase_name = opt.phase_name

is_pretrain = phase_name == 'pretrain'

is_search = phase_name == 'search'

is_enot_phase = is_pretrain or is_search

# Directories

w = save_dir / 'weights' # weights dir

w.mkdir(parents=True, exist_ok=True) # make dir

last, best = w / 'last.pt', w / 'best.pt'

# Hyperparameters

if isinstance(hyp, str):

with open(hyp) as f:

hyp = yaml.safe_load(f) # load hyps dict

LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))

# Save run settings

with open(save_dir / 'hyp.yaml', 'w') as f:

yaml.safe_dump(hyp, f, sort_keys=False)

with open(save_dir / 'opt.yaml', 'w') as f:

yaml.safe_dump(vars(opt), f, sort_keys=False)

data_dict = None

# Loggers

if RANK in [-1, 0]:

loggers = Loggers(save_dir, weights, opt, hyp, LOGGER) # loggers instance

if loggers.wandb:

data_dict = loggers.wandb.data_dict

if resume:

weights, epochs, hyp = opt.weights, opt.epochs, opt.hyp

# Register actions

for k in methods(loggers):

callbacks.register_action(k, callback=getattr(loggers, k))

# Config

plots = not evolve # create plots

cuda = device.type != 'cpu'

init_seeds(1 + RANK)

with torch_distributed_zero_first(RANK):

data_dict = data_dict or check_dataset(data) # check if None

train_path, val_path = data_dict[phase_name], data_dict['val']

nc = 1 if single_cls else int(data_dict['nc']) # number of classes

names = ['item'] if single_cls and len(data_dict['names']) != 1 else data_dict['names'] # class names

assert len(names) == nc, f'{len(names)} names found for nc={nc} dataset in {data}' # check

is_coco = data.endswith('coco.yaml') and nc == 80 # COCO dataset

# Model

pretrained = weights.endswith('.pt')

model_to_optimize, model, search_space, ckpt = create_and_prepare_model(

weights, device, phase_name, cfg, opt, nc, hyp, resume, is_enot_phase,

)

if not is_search and opt.grad_checkpoint:

LOGGER.info('Enabling gradient checkpoint for better GPU memory utilization')

enable_checkpoint(model_to_optimize)

# Freeze

freeze = [f'model.{x}.' for x in range(freeze)] # layers to freeze

for k, v in model.named_parameters():

v.requires_grad = True # train all layers

if any(x in k for x in freeze):

print(f'freezing {k}')

v.requires_grad = False

# Optimizer

if is_search: # We don't need to accumulate gradients during search procedure.

nbs = batch_size # nominal batch size

accumulate = 1 # accumulate loss before optimizing

else:

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

LOGGER.info(f"Scaled weight_decay = {hyp['weight_decay']}")

g0, g1, g2 = [], [], [] # optimizer parameter groups

if not is_search:

# Select parameters to optimize and later choose only params from this set.

params_to_optimize = set(

search_space.model_parameters()

if is_pretrain

else model_to_optimize.parameters()

)

for v in model_to_optimize.modules():

if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter) and v.bias in params_to_optimize: # bias

g2.append(v.bias)

if isinstance(v, nn.BatchNorm2d): # weight (no decay)

g0.append(v.weight)

elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter) and v.weight in params_to_optimize: # weight (with decay)

g1.append(v.weight)

else:

# Use architecture parameters during search procedure.

g0 = list(search_space.architecture_parameters())

if opt.adam:

optimizer = Adam(g0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum

else:

optimizer = SGD(g0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)

# These param groups are redundant for search procedure.

if not is_search:

optimizer.add_param_group({'params': g1, 'weight_decay': hyp['weight_decay']}) # add g1 with weight_decay

optimizer.add_param_group({'params': g2}) # add g2 (biases)

LOGGER.info(f"{colorstr('optimizer:')} {type(optimizer).__name__} with parameter groups "

f"{len(g0)} weight, {len(g1)} weight (no decay), {len(g2)} bias")

del g0, g1, g2

# Scheduler

if opt.linear_lr:

lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp['lrf'] # linear

else:

lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf']

scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs)

# EMA (should not be applied during search procedure)

ema = ModelEMA(model_to_optimize) if (RANK in [-1, 0] and not is_search) else None

# Resume

start_epoch, best_fitness = 0, 0.0

if pretrained:

if phase_name == ckpt['phase_name']:

LOGGER.info(f'Restoring train state for the same phase ({phase_name})')

# Optimizer

if ckpt['optimizer'] is not None:

optimizer.load_state_dict(ckpt['optimizer'])

best_fitness = ckpt['best_fitness']

# EMA

if ema and ckpt.get('ema'):

ema.ema.load_state_dict(ckpt['ema'].float().state_dict())

ema.updates = ckpt['updates']

# Epochs

start_epoch = ckpt['epoch'] + 1

if resume:

assert start_epoch > 0, f'{weights} training to {epochs} epochs is finished, nothing to resume.'

if epochs < start_epoch:

LOGGER.info(

f"{weights} has been trained for {ckpt['epoch']} epochs. Fine-tuning for {epochs} more epochs."

)

epochs += ckpt['epoch'] # finetune additional epochs

del ckpt

else:

LOGGER.info(f'Dropping training state because phase has changed ({ckpt["phase_name"]} -> {phase_name})')

# Image sizes

gs = max(int(model.stride.max()), 32) # grid size (max stride)

nl = model.model[-1].nl # number of detection layers (used for scaling hyp['obj'])

imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2) # verify imgsz is gs-multiple

# DP mode

if cuda and RANK == -1 and torch.cuda.device_count() > 1:

# DP is not supported in ENOT, raise.

raise RuntimeError('DP is not supported in ENOT framework, use torch.distributed.run instead.\n'

'See Multi-GPU Tutorial at https://github.com/ultralytics/yolov5/issues/475 to get started.')

# SyncBatchNorm

if opt.sync_bn and cuda and RANK != -1:

raise RuntimeError('SyncBatchNorm is supported, but you should contact ENOT developers to use it.')

# Trainloader

augment = not is_search # We should not augment search data during search procedure.

pad = 0.5 if is_search else 0.0 # Use the same padding as in validation data loader.

use_rect = opt.rect or is_search # Use rectangular images as in validation data loader.

train_loader, dataset = create_dataloader(train_path, imgsz, batch_size // WORLD_SIZE, gs, single_cls,

hyp=hyp, augment=augment, cache=opt.cache, rect=use_rect, rank=RANK,

workers=workers, image_weights=opt.image_weights, quad=opt.quad,

pad=pad, prefix=colorstr(f'{phase_name}: '))

bn_tune_loader = train_loader # Use pretrain data loader for batch norm tuning.

mlc = int(np.concatenate(dataset.labels, 0)[:, 0].max()) # max label class

nb = len(train_loader) # number of batches

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, pad=0.5,

prefix=colorstr('val: '))[0]

if not resume:

labels = np.concatenate(dataset.labels, 0)

# c = torch.tensor(labels[:, 0]) # classes

# cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency

# model._initialize_biases(cf.to(device))

if plots:

plot_labels(labels, names, save_dir)

# Anchors

if not opt.noautoanchor:

check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)

model.half().float() # pre-reduce anchor precision

callbacks.on_pretrain_routine_end()

# Not necessary with ENOT framework, only keep for train phase.

# DDP mode

if cuda and RANK != -1 and not is_enot_phase:

model_to_optimize = DDP(model_to_optimize, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK)

# Try to load latency container for latency-aware search.

# You can find latency containers in yolov5/latency folder.

if is_search:

latency_file = opt.latency_file

if latency_file == '' and opt.max_latency_value != 0.0:

raise ValueError('Latency file is not provided when max-latency-value is positive')

if latency_file != '':

latency_container = SearchSpaceLatencyContainer.load_from_file(latency_file)

latency_scale = max_latency(latency_container)

latency_type = latency_container.latency_type

constant_latency = latency_container.constant_latency / latency_scale

operations_latencies = (np.array(latency_container.operations_latencies) / latency_scale).tolist()

latency_container = SearchSpaceLatencyContainer(

latency_type,

constant_latency,

operations_latencies,

)

search_space.apply_latency_container(latency_container)

opt.max_latency_value /= latency_scale

additional_kwargs = {}

if is_search:

# Use pretrain data loader for batch norm tuning (with all augmentations as during pre-train).

bn_tune_loader = create_dataloader(

data_dict['pretrain'], 640, batch_size // WORLD_SIZE * 2, gs, single_cls,

hyp=hyp, augment=True, cache=opt.cache, rect=opt.rect, rank=RANK,

workers=workers, image_weights=opt.image_weights, quad=opt.quad,

pad=0,

)[0]

# Define additional keyword arguments for ENOT optimizer which are necessary for yolov5 search procedure.

additional_kwargs = {

'bn_tune_dataloader': bn_tune_loader,

'bn_tune_batches': 5,

'bn_validation_tune_batches': 0,

'sample_to_model_inputs': preprocess_data,

}

if opt.max_latency_value != 0.0:

additional_kwargs['max_latency_value'] = opt.max_latency_value

# Construct ENOT optimizer.

optimizer = build_enot_optimizer(phase_name, model=model_to_optimize, optimizer=optimizer, **additional_kwargs)

bn_tune_loader = BatchNormTuneDataLoaderWrapper(bn_tune_loader) # Wrap to make PyTorch happy.

# Perform output distribution optimization (recommended for convnets).

if is_pretrain:

sample_input = torch.zeros(1, 3, 256, 256).cuda()

search_space.initialize_output_distribution_optimization(sample_input)

# Model parameters

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.nc = nc # attach number of classes to model

model.hyp = hyp # attach hyperparameters to model

model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc # attach class weights

model.names = names

# Start training

t0 = time.time()

nw = max(round(hyp['warmup_epochs'] * nb), 1000) # number of warmup iterations, max(3 epochs, 1k 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 = amp.GradScaler(enabled=cuda)

compute_loss = ComputeLoss(model) # init loss class

LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'

f'Using {train_loader.num_workers} dataloader workers\n'

f'Logging results to {save_dir}\n'

f'Starting training for {epochs} epochs...')

for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------

model.train()

# Update image weights (optional)

if opt.image_weights and not is_search: # image_weights are not tested.

# Generate indices

if RANK in [-1, 0]:

cw = model.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

# Broadcast if DDP

if RANK != -1:

indices = (torch.tensor(dataset.indices) if RANK == 0 else torch.zeros(dataset.n)).int()

dist.broadcast(indices, 0)

if RANK != 0:

dataset.indices = indices.cpu().numpy()

# Update mosaic border

# 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

if RANK != -1:

train_loader.sampler.set_epoch(epoch)

pbar = enumerate(train_loader)

LOGGER.info(('\n' + '%10s' * 7) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'labels', 'img_size'))

if RANK in [-1, 0]:

pbar = tqdm(pbar, total=nb) # progress bar

optimizer.zero_grad()

for i, (imgs, targets, paths, _) in pbar: # batch -------------------------------------------------------------

ni = i + nb * epoch # number integrated batches (since train start)

imgs = imgs.to(device, non_blocking=True).float() / 255.0 # uint8 to float32, 0-255 to 0.0-1.0

# Warmup

if ni <= nw and not is_search: # Warmup is not needed during search procedure

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.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 == 2 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 and not is_search: # Multi-scale search is senseless. Run separate search runs instead.

sz = random.randrange(imgsz * 0.5, 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)

loss, loss_items = None, None

# Wrap forward+backward with closure.

def closure():

# Forward

with amp.autocast(enabled=cuda):

model.model[-1].training = True

pred = model_to_optimize(imgs) # forward

nonlocal loss, loss_items

loss, loss_items = compute_loss(pred, targets.to(device)) # loss scaled by batch_size

if RANK != -1:

loss *= WORLD_SIZE # gradient averaged between devices in DDP mode

if opt.quad:

loss *= 4.

if is_search: # For latency optimization.

loss = optimizer.modify_loss(loss)

# Backward

scaler.scale(loss).backward()

# For gradient accumulation.

optimizer.model_step(closure)

# Optimize

if ni - last_opt_step >= accumulate:

scaler.step(optimizer) # optimizer.step

scaler.update()

optimizer.zero_grad()

if ema:

ema.update(model_to_optimize)

last_opt_step = ni

# 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(('%10s' * 2 + '%10.4g' * 5) % (

f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))

callbacks.on_train_batch_end(ni, model, imgs, targets, paths, plots)

# end batch ------------------------------------------------------------------------------------------------

# Scheduler

lr = [x['lr'] for x in optimizer.param_groups] # for loggers

if is_search: # Fix that during search we have only one parameter group.

lr = [optimizer.param_groups[0]['lr'], 0.0, 0.0]

scheduler.step()

if RANK in [-1, 0]:

# mAP

callbacks.on_train_epoch_end(epoch=epoch)

if ema is not None:

ema.update_attr(model_to_optimize, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])

if is_enot_phase:

if is_pretrain: # Testing baseline architecture during pretrain phase.

arch_to_test = [0] * len(search_space.search_blocks)

else: # Testing the best current architecture during search phase.

np.set_printoptions(precision=4, linewidth=120)

arch_to_test = optimizer.prepare_validation_model()

LOGGER.info(f'Testing architecture: {arch_to_test}')

LOGGER.info(f'It\'s latency: {current_latency(search_space)}')

LOGGER.info(

f'Architecture probabilities:\n'

f'{np.array(search_space.architecture_probabilities)}'

)

np.save(str(save_dir / f'p_{epoch + 1}.bin'), np.array(search_space.architecture_probabilities))

np.save(str(save_dir / f'a_{epoch + 1}.bin'), np.array(arch_to_test, dtype=np.int64))

test_search_space = ema.ema if ema is not None else search_space # Use EMA model if exists.

test_model = test_search_space.get_network_by_indexes(arch_to_test).cuda() # Extract single arch.

tune_bn_stats( # Separate batch norm tuning. This gives better network performance estimation.

test_model,

bn_tune_loader,

reset_bns=True,

set_momentums_none=True,

n_steps=50, # You can reduce this value for faster execution.

sample_to_model_inputs=preprocess_data,

)

else:

test_model = ema.ema if ema is not None else model

final_epoch = epoch + 1 == epochs

if not noval or final_epoch: # Calculate mAP

results, maps, _ = val.run(data_dict,

batch_size=batch_size // WORLD_SIZE * 2,

imgsz=imgsz,

model=test_model,

single_cls=single_cls,

dataloader=val_loader,

save_dir=save_dir,

save_json=is_coco and final_epoch,

verbose=nc < 50 and final_epoch,

plots=plots and final_epoch,

callbacks=callbacks,

compute_loss=compute_loss)

# Update best mAP

fi = fitness(np.array(results).reshape(1, -1)) # weighted combination of [P, R, mAP@.5, mAP@.5-.95]

if fi > best_fitness:

best_fitness = fi

log_vals = list(mloss) + list(results) + lr

callbacks.on_fit_epoch_end(log_vals, epoch, best_fitness, fi)

# Save model

if (not nosave) or (final_epoch and not evolve): # if save

ckpt = {'epoch': epoch,

'best_fitness': best_fitness,

'model': deepcopy(de_parallel(model_to_optimize)).half(), # Save search space or model.

'ema': deepcopy(ema.ema).half() if ema is not None else None,

'updates': ema.updates if ema is not None else None,

'optimizer': optimizer.state_dict(),

'wandb_id': loggers.wandb.wandb_run.id if loggers.wandb else None,

'phase_name': phase_name} # Also save phase name to use in phase transitions.

# Save last, best and delete

torch.save(ckpt, last)

if best_fitness == fi:

torch.save(ckpt, best)

del ckpt

callbacks.on_model_save(last, epoch, final_epoch, best_fitness, fi)

# end epoch ----------------------------------------------------------------------------------------------------

# end training -----------------------------------------------------------------------------------------------------

if RANK in [-1, 0]:

LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')

if not evolve:

# Not necessary to call it here.

# if is_coco: # COCO dataset

# for m in [last, best] if best.exists() else [last]: # speed, mAP tests

# results, _, _ = val.run(data_dict,

# batch_size=batch_size // WORLD_SIZE * 2,

# imgsz=imgsz,

# model=attempt_load(m, device).half(),

# iou_thres=0.7, # NMS IoU threshold for best pycocotools results

# single_cls=single_cls,

# dataloader=val_loader,

# save_dir=save_dir,

# save_json=True,

# plots=False)

# Strip optimizers

for f in last, best:

if f.exists():

strip_optimizer(f) # strip optimizers

callbacks.on_train_end(last, best, plots, epoch)

LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}")

torch.cuda.empty_cache()

parser = argparse.ArgumentParser()

parser.add_argument('--weights', type=str, default='yolov5s.pt', help='initial weights path')

parser.add_argument('--cfg', type=str, default='', help='model.yaml path')

parser.add_argument('--data', type=str, default='data/coco128.yaml', help='dataset.yaml path')

parser.add_argument('--hyp', type=str, default='data/hyps/hyp.scratch.yaml', help='hyperparameters path')

parser.add_argument('--epochs', type=int, default=300)

parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs')

parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, 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 check')

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='--cache images in "ram" (default) or "disk"')

parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')

parser.add_argument('--device', default='', 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('--adam', action='store_true', help='use torch.optim.Adam() 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='maximum number of dataloader workers')

parser.add_argument('--project', default='runs/train', help='save to project/name')

parser.add_argument('--entity', default=None, help='W&B entity')

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('--linear-lr', action='store_true', help='linear LR')

parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')

parser.add_argument('--upload_dataset', action='store_true', help='Upload dataset as W&B artifact table')

parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval for W&B')

parser.add_argument('--save_period', type=int, default=-1, help='Log model after every "save_period" epoch')

parser.add_argument('--artifact_alias', type=str, default="latest", help='version of dataset artifact to be used')

parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify')

parser.add_argument('--freeze', type=int, default=0, help='Number of layers to freeze. backbone=10, all=24')

parser.add_argument('--phase-name', type=str, default='train', help='ENOT phase name')

parser.add_argument('--max-latency-value', type=float, default=0.0, help='Latency constraint for search procedure')

parser.add_argument('--architecture-indices', type=str, default='[]', help='Train model architecture indices')

parser.add_argument('--latency-file', type=str, default='', help='Path to search space latency container file')

parser.add_argument('--grad-checkpoint', action='store_true', help='Activate gradient checkpointing')

opt = parser.parse_known_args()[0] if known else parser.parse_args()

# Checks

set_logging(RANK)

if RANK in [-1, 0]:

print(colorstr('train: ') + ', '.join(f'{k}={v}' for k, v in vars(opt).items()))

# check_git_status() # We changed the repository!

check_requirements(requirements=FILE.parent / 'requirements.txt', exclude=['thop'])

# Resume

if opt.resume and not check_wandb_resume(opt) and not opt.evolve: # resume an interrupted run

ckpt = opt.resume if isinstance(opt.resume, str) else get_latest_run() # specified or most recent path

assert os.path.isfile(ckpt), 'ERROR: --resume checkpoint does not exist'

with open(Path(ckpt).parent.parent / 'opt.yaml') as f:

opt = argparse.Namespace(**yaml.safe_load(f)) # replace

opt.cfg, opt.weights, opt.resume = '', ckpt, True # reinstate

LOGGER.info(f'Resuming training from {ckpt}')

else:

opt.data, opt.cfg, opt.hyp = check_file(opt.data), check_file(opt.cfg), check_file(opt.hyp) # check files

assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'

if opt.evolve:

opt.project = 'runs/evolve'

opt.exist_ok = opt.resume

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:

from datetime import timedelta

assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'

assert opt.batch_size % WORLD_SIZE == 0, '--batch-size must be multiple of CUDA device count'

assert not opt.image_weights, '--image-weights argument is not compatible with DDP training'

assert not opt.evolve, '--evolve argument is not compatible with DDP training'

assert not opt.sync_bn, '--sync-bn known training issue, see https://github.com/ultralytics/yolov5/issues/3998'

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")

# Train

if not opt.evolve:

train(opt.hyp, opt, device)

if WORLD_SIZE > 1 and RANK == 0:

_ = [print('Destroying process group... ', end=''), dist.destroy_process_group(), print('Done.')]

# 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': (0.3, 0.6, 0.98), # SGD momentum/Adam beta1

'weight_decay': (1, 0.0, 0.001), # optimizer weight decay

'warmup_epochs': (1, 0.0, 5.0), # warmup epochs (fractions ok)

'warmup_momentum': (1, 0.0, 0.95), # warmup initial momentum

'warmup_bias_lr': (1, 0.0, 0.2), # warmup initial bias lr

'box': (1, 0.02, 0.2), # box loss gain

'cls': (1, 0.2, 4.0), # cls loss gain

'cls_pw': (1, 0.5, 2.0), # cls BCELoss positive_weight

'obj': (1, 0.2, 4.0), # obj loss gain (scale with pixels)

'obj_pw': (1, 0.5, 2.0), # obj BCELoss positive_weight

'iou_t': (0, 0.1, 0.7), # IoU training threshold

'anchor_t': (1, 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': (1, 0.0, 0.1), # image HSV-Hue augmentation (fraction)

'hsv_s': (1, 0.0, 0.9), # image HSV-Saturation augmentation (fraction)

'hsv_v': (1, 0.0, 0.9), # image HSV-Value augmentation (fraction)

'degrees': (1, 0.0, 45.0), # image rotation (+/- deg)

'translate': (1, 0.0, 0.9), # image translation (+/- fraction)

'scale': (1, 0.0, 0.9), # image scale (+/- gain)

'shear': (1, 0.0, 10.0), # image shear (+/- deg)

'perspective': (0, 0.0, 0.001), # image perspective (+/- fraction), range 0-0.001

'flipud': (1, 0.0, 1.0), # image flip up-down (probability)

'fliplr': (0, 0.0, 1.0), # image flip left-right (probability)

'mosaic': (1, 0.0, 1.0), # image mixup (probability)

'mixup': (1, 0.0, 1.0), # image mixup (probability)

'copy_paste': (1, 0.0, 1.0)} # segment copy-paste (probability)

with open(opt.hyp) as f:

hyp = yaml.safe_load(f) # load hyps dict

if 'anchors' not in hyp: # anchors commented in hyp.yaml

hyp['anchors'] = 3

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'