car-detection-bayes/train.py

493 lines
22 KiB
Python
Raw Normal View History

2018-08-26 08:51:39 +00:00
import argparse
2019-07-25 11:19:26 +00:00
import torch.distributed as dist
2019-04-17 13:52:51 +00:00
import torch.optim as optim
2019-05-30 17:02:55 +00:00
import torch.optim.lr_scheduler as lr_scheduler
2019-03-21 12:48:40 +00:00
2019-06-24 11:43:17 +00:00
import test # import test.py to get mAP after each epoch
2018-08-26 08:51:39 +00:00
from models import *
from utils.datasets import *
from utils.utils import *
2019-07-24 16:02:26 +00:00
mixed_precision = True
try: # Mixed precision training https://github.com/NVIDIA/apex
from apex import amp
2019-08-01 16:29:57 +00:00
except:
mixed_precision = False # not installed
2019-07-24 16:02:26 +00:00
2019-09-17 22:38:49 +00:00
wdir = 'weights' + os.sep # weights dir
last = wdir + 'last.pt'
best = wdir + 'best.pt'
2019-09-17 22:54:07 +00:00
results_file = 'results.txt'
2019-09-17 22:38:49 +00:00
2019-11-14 23:07:27 +00:00
# Hyperparameters (k-series, 57.7 mAP yolov3-spp-416) https://github.com/ultralytics/yolov3/issues/310
2019-10-25 16:03:04 +00:00
hyp = {'giou': 3.31, # giou loss gain
2019-10-25 16:04:10 +00:00
'cls': 42.4, # cls loss gain
2019-10-25 16:03:04 +00:00
'cls_pw': 1.0, # cls BCELoss positive_weight
2019-11-18 20:01:17 +00:00
'obj': 52.0, # obj loss gain (*=img_size/320 if img_size != 320)
2019-10-25 16:03:04 +00:00
'obj_pw': 1.0, # obj BCELoss positive_weight
'iou_t': 0.213, # iou training threshold
'lr0': 0.00261, # initial learning rate (SGD=1E-3, Adam=9E-5)
2019-07-20 12:54:37 +00:00
'lrf': -4., # final LambdaLR learning rate = lr0 * (10 ** lrf)
2019-10-25 16:03:04 +00:00
'momentum': 0.949, # SGD momentum
'weight_decay': 0.000489, # optimizer weight decay
2019-09-10 09:35:46 +00:00
'fl_gamma': 0.5, # focal loss gamma
2019-10-25 16:03:04 +00:00
'hsv_h': 0.0103, # image HSV-Hue augmentation (fraction)
'hsv_s': 0.691, # image HSV-Saturation augmentation (fraction)
'hsv_v': 0.433, # image HSV-Value augmentation (fraction)
'degrees': 1.43, # image rotation (+/- deg)
'translate': 0.0663, # image translation (+/- fraction)
'scale': 0.11, # image scale (+/- gain)
'shear': 0.384} # image shear (+/- deg)
2019-08-18 00:08:47 +00:00
2019-09-11 12:00:57 +00:00
# Overwrite hyp with hyp*.txt (optional)
f = glob.glob('hyp*.txt')
if f:
for k, v in zip(hyp.keys(), np.loadtxt(f[0])):
hyp[k] = v
2019-09-11 11:15:16 +00:00
2019-08-18 11:05:32 +00:00
2019-08-23 11:25:27 +00:00
def train():
cfg = opt.cfg
data = opt.data
img_size = opt.img_size
2019-09-09 20:42:38 +00:00
epochs = 1 if opt.prebias else opt.epochs # 500200 batches at bs 64, 117263 images = 273 epochs
2019-08-23 11:25:27 +00:00
batch_size = opt.batch_size
accumulate = opt.accumulate # effective bs = batch_size * accumulate = 16 * 4 = 64
2019-08-23 13:17:17 +00:00
weights = opt.weights # initial training weights
2019-08-23 11:25:27 +00:00
2019-08-25 18:19:53 +00:00
if 'pw' not in opt.arc: # remove BCELoss positive weights
2019-08-27 10:57:19 +00:00
hyp['cls_pw'] = 1.
hyp['obj_pw'] = 1.
2019-08-25 18:19:53 +00:00
2019-07-15 15:00:04 +00:00
# Initialize
2019-04-17 15:27:51 +00:00
init_seeds()
2019-07-10 18:47:05 +00:00
multi_scale = opt.multi_scale
2019-06-13 16:13:30 +00:00
if multi_scale:
2019-11-20 21:24:50 +00:00
img_sz_min = round(img_size / 32 / 1.5)
img_sz_max = round(img_size / 32 * 1.5)
2019-07-20 11:20:01 +00:00
img_size = img_sz_max * 32 # initiate with maximum multi_scale size
2019-08-04 00:50:35 +00:00
print('Using multi-scale %g - %g' % (img_sz_min * 32, img_size))
2018-08-26 08:51:39 +00:00
# Configure run
2019-07-20 13:10:31 +00:00
data_dict = parse_data_cfg(data)
2019-04-27 15:57:07 +00:00
train_path = data_dict['train']
nc = int(data_dict['classes']) # number of classes
2018-08-26 08:51:39 +00:00
# Remove previous results
2019-09-18 00:25:09 +00:00
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
2018-08-26 08:51:39 +00:00
# Initialize model
2019-08-23 15:18:59 +00:00
model = Darknet(cfg, arc=opt.arc).to(device)
2018-08-26 08:51:39 +00:00
# Optimizer
2019-08-26 12:47:36 +00:00
pg0, pg1 = [], [] # optimizer parameter groups
for k, v in dict(model.named_parameters()).items():
if 'Conv2d.weight' in k:
pg1 += [v] # parameter group 1 (apply weight_decay)
else:
pg0 += [v] # parameter group 0
2019-09-11 12:25:48 +00:00
if opt.adam:
optimizer = optim.Adam(pg0, lr=hyp['lr0'])
# optimizer = AdaBound(pg0, lr=hyp['lr0'], final_lr=0.1)
else:
optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
2019-08-26 12:47:36 +00:00
optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay
del pg0, pg1
2018-08-26 08:51:39 +00:00
2019-02-21 14:57:18 +00:00
cutoff = -1 # backbone reaches to cutoff layer
2019-02-22 15:15:20 +00:00
start_epoch = 0
2019-10-15 23:36:13 +00:00
best_fitness = float('inf')
2019-09-19 16:05:04 +00:00
attempt_download(weights)
2019-08-23 13:17:17 +00:00
if weights.endswith('.pt'): # pytorch format
2019-11-15 01:22:09 +00:00
# possible weights are '*.pt', 'yolov3-spp.pt', 'yolov3-tiny.pt' etc.
2019-08-23 13:17:17 +00:00
chkpt = torch.load(weights, map_location=device)
2019-04-02 16:04:04 +00:00
2019-08-23 13:17:17 +00:00
# load model
2019-11-25 21:45:28 +00:00
try:
chkpt['model'] = {k: v for k, v in chkpt['model'].items() if model.state_dict()[k].numel() == v.numel()}
model.load_state_dict(chkpt['model'], strict=False)
# model.load_state_dict(chkpt['model'])
except KeyError as e:
s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s. " \
"See https://github.com/ultralytics/yolov3/issues/657" % (opt.weights, opt.cfg, opt.weights)
raise KeyError(s) from e
2019-04-02 16:04:04 +00:00
2019-08-23 13:17:17 +00:00
# load optimizer
2019-04-02 16:04:04 +00:00
if chkpt['optimizer'] is not None:
optimizer.load_state_dict(chkpt['optimizer'])
2019-07-02 16:21:28 +00:00
best_fitness = chkpt['best_fitness']
2019-07-08 16:00:19 +00:00
2019-08-23 13:17:17 +00:00
# load results
if chkpt.get('training_results') is not None:
2019-09-18 00:25:09 +00:00
with open(results_file, 'w') as file:
2019-07-08 17:26:46 +00:00
file.write(chkpt['training_results']) # write results.txt
2019-07-08 16:00:19 +00:00
start_epoch = chkpt['epoch'] + 1
2019-04-02 16:04:04 +00:00
del chkpt
2018-10-30 14:18:52 +00:00
2019-08-23 13:37:25 +00:00
elif len(weights) > 0: # darknet format
2019-11-15 01:22:09 +00:00
# possible weights are '*.weights', 'yolov3-tiny.conv.15', 'darknet53.conv.74' etc.
2019-08-23 13:17:17 +00:00
cutoff = load_darknet_weights(model, weights)
2018-10-30 14:18:52 +00:00
2019-08-24 18:55:01 +00:00
if opt.transfer or opt.prebias: # transfer learning edge (yolo) layers
2019-08-23 13:17:17 +00:00
nf = int(model.module_defs[model.yolo_layers[0] - 1]['filters']) # yolo layer size (i.e. 255)
2019-08-23 13:43:16 +00:00
2019-09-26 10:08:40 +00:00
if opt.prebias:
for p in optimizer.param_groups:
# lower param count allows more aggressive training settings: i.e. SGD ~0.1 lr0, ~0.9 momentum
p['lr'] *= 100 # lr gain
if p.get('momentum') is not None: # for SGD but not Adam
p['momentum'] *= 0.9
2019-08-23 13:43:16 +00:00
2019-08-23 13:17:17 +00:00
for p in model.parameters():
2019-08-26 12:47:36 +00:00
if opt.prebias and p.numel() == nf: # train (yolo biases)
2019-08-24 18:55:01 +00:00
p.requires_grad = True
2019-08-26 12:47:36 +00:00
elif opt.transfer and p.shape[0] == nf: # train (yolo biases+weights)
2019-08-24 18:55:01 +00:00
p.requires_grad = True
2019-08-26 12:47:36 +00:00
else: # freeze layer
2019-08-24 18:55:01 +00:00
p.requires_grad = False
2019-06-16 21:17:40 +00:00
2019-04-24 10:58:14 +00:00
# Scheduler https://github.com/ultralytics/yolov3/issues/238
2019-04-17 13:52:51 +00:00
# lf = lambda x: 1 - x / epochs # linear ramp to zero
2019-04-24 11:30:24 +00:00
# lf = lambda x: 10 ** (hyp['lrf'] * x / epochs) # exp ramp
2019-05-30 17:02:55 +00:00
# lf = lambda x: 1 - 10 ** (hyp['lrf'] * (1 - x / epochs)) # inverse exp ramp
# scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
2019-09-18 22:36:11 +00:00
# scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=range(59, 70, 1), gamma=0.8) # gradual fall to 0.1*lr0
2019-07-30 16:25:53 +00:00
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[round(opt.epochs * x) for x in [0.8, 0.9]], gamma=0.1)
2019-05-30 17:02:55 +00:00
scheduler.last_epoch = start_epoch - 1
2019-04-18 19:56:50 +00:00
2019-04-24 10:58:14 +00:00
# # Plot lr schedule
2019-04-18 19:44:57 +00:00
# y = []
# for _ in range(epochs):
# scheduler.step()
# y.append(optimizer.param_groups[0]['lr'])
2019-04-24 10:58:14 +00:00
# plt.plot(y, label='LambdaLR')
# plt.xlabel('epoch')
2019-06-21 11:19:23 +00:00
# plt.ylabel('LR')
2019-04-24 10:58:14 +00:00
# plt.tight_layout()
# plt.savefig('LR.png', dpi=300)
2019-04-17 14:15:08 +00:00
2019-07-24 16:02:26 +00:00
# Mixed precision training https://github.com/NVIDIA/apex
if mixed_precision:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
# Initialize distributed training
2019-11-20 21:14:24 +00:00
if device.type != 'cpu' and torch.cuda.device_count() > 1:
2019-07-24 16:02:26 +00:00
dist.init_process_group(backend='nccl', # 'distributed backend'
init_method='tcp://127.0.0.1:9999', # distributed training init method
world_size=1, # number of nodes for distributed training
rank=0) # distributed training node rank
model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)
2019-08-05 15:25:50 +00:00
model.yolo_layers = model.module.yolo_layers # move yolo layer indices to top level
2019-07-24 16:02:26 +00:00
# Dataset
2019-05-21 15:37:34 +00:00
dataset = LoadImagesAndLabels(train_path,
img_size,
batch_size,
augment=True,
2019-07-20 12:54:37 +00:00
hyp=hyp, # augmentation hyperparameters
2019-07-30 15:51:19 +00:00
rect=opt.rect, # rectangular training
2019-08-07 14:45:13 +00:00
image_weights=opt.img_weights,
2019-09-16 17:53:44 +00:00
cache_labels=True if epochs > 10 else False,
2019-08-28 14:50:34 +00:00
cache_images=False if opt.prebias else opt.cache_images)
# Dataloader
2019-11-21 03:34:22 +00:00
batch_size = min(batch_size, len(dataset))
2019-07-24 13:56:10 +00:00
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
2019-11-21 03:34:22 +00:00
num_workers=min([os.cpu_count(), batch_size if batch_size > 1 else 0, 16]),
2019-07-24 13:56:10 +00:00
shuffle=not opt.rect, # Shuffle=True unless rectangular training is used
pin_memory=True,
collate_fn=dataset.collate_fn)
2018-08-26 08:51:39 +00:00
2019-03-07 16:16:38 +00:00
# Start training
2019-08-05 14:59:32 +00:00
model.nc = nc # attach number of classes to model
2019-08-23 15:37:29 +00:00
model.arc = opt.arc # attach yolo architecture
2019-04-17 13:52:51 +00:00
model.hyp = hyp # attach hyperparameters to model
2019-11-20 21:36:15 +00:00
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights
2019-10-06 22:50:47 +00:00
torch_utils.model_info(model, report='summary') # 'full' or 'summary'
2019-04-17 13:52:51 +00:00
nb = len(dataloader)
2019-05-10 12:15:09 +00:00
maps = np.zeros(nc) # mAP per class
2019-11-26 03:24:05 +00:00
# torch.autograd.set_detect_anomaly(True)
2019-08-24 15:16:20 +00:00
results = (0, 0, 0, 0, 0, 0, 0) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
2019-07-16 15:56:39 +00:00
t0 = time.time()
2019-08-28 14:50:34 +00:00
print('Starting %s for %g epochs...' % ('prebias' if opt.prebias else 'training', epochs))
2019-08-22 22:36:48 +00:00
for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------
model.train()
2019-08-24 14:43:43 +00:00
print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size'))
2018-09-20 16:03:19 +00:00
2019-05-23 10:32:11 +00:00
# Freeze backbone at epoch 0, unfreeze at epoch 1 (optional)
2019-07-15 15:00:04 +00:00
freeze_backbone = False
2019-03-19 08:38:32 +00:00
if freeze_backbone and epoch < 2:
for name, p in model.named_parameters():
2019-02-21 14:57:18 +00:00
if int(name.split('.')[1]) < cutoff: # if layer < 75
p.requires_grad = False if epoch == 0 else True
2018-11-27 17:14:48 +00:00
2019-07-30 15:51:19 +00:00
# Update image weights (optional)
if dataset.image_weights:
2019-08-01 23:33:24 +00:00
w = model.class_weights.cpu().numpy() * (1 - maps) ** 2 # class weights
2019-07-30 15:51:19 +00:00
image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w)
dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n) # rand weighted idx
2019-05-10 12:15:09 +00:00
2019-08-24 14:43:43 +00:00
mloss = torch.zeros(4).to(device) # mean losses
2019-06-30 15:34:29 +00:00
pbar = tqdm(enumerate(dataloader), total=nb) # progress bar
2019-08-22 22:36:48 +00:00
for i, (imgs, targets, paths, _) in pbar: # batch -------------------------------------------------------------
2019-08-23 11:39:43 +00:00
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device)
targets = targets.to(device)
2018-09-19 02:21:46 +00:00
2019-08-08 17:49:15 +00:00
# Multi-Scale training
2019-06-13 16:13:30 +00:00
if multi_scale:
2019-08-08 18:16:32 +00:00
if ni / accumulate % 10 == 0: #  adjust (67% - 150%) every 10 batches
2019-07-20 11:20:01 +00:00
img_size = random.randrange(img_sz_min, img_sz_max + 1) * 32
sf = img_size / max(imgs.shape[2:]) # scale factor
if sf != 1:
2019-08-08 17:49:15 +00:00
ns = [math.ceil(x * sf / 32.) * 32 for x in imgs.shape[2:]] # new shape (stretched to 32-multiple)
2019-07-20 11:20:01 +00:00
imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
2019-06-12 11:04:58 +00:00
2019-03-21 20:41:12 +00:00
# Plot images with bounding boxes
2019-08-23 11:31:32 +00:00
if ni == 0:
2019-08-09 14:37:19 +00:00
fname = 'train_batch%g.jpg' % i
2019-08-09 17:29:36 +00:00
plot_images(imgs=imgs, targets=targets, paths=paths, fname=fname)
2019-08-09 14:37:19 +00:00
if tb_writer:
2019-08-09 17:29:36 +00:00
tb_writer.add_image(fname, cv2.imread(fname)[:, :, ::-1], dataformats='HWC')
2019-03-21 20:41:12 +00:00
2019-08-08 17:49:15 +00:00
# Hyperparameter burn-in
2019-08-08 19:16:09 +00:00
# n_burn = nb - 1 # min(nb // 5 + 1, 1000) # number of burn-in batches
2019-08-08 18:16:32 +00:00
# if ni <= n_burn:
2019-08-08 17:49:15 +00:00
# for m in model.named_modules():
# if m[0].endswith('BatchNorm2d'):
2019-08-08 18:16:32 +00:00
# m[1].momentum = 1 - i / n_burn * 0.99 # BatchNorm2d momentum falls from 1 - 0.01
# g = (i / n_burn) ** 4 # gain rises from 0 - 1
2019-07-22 23:35:03 +00:00
# for x in optimizer.param_groups:
# x['lr'] = hyp['lr0'] * g
# x['weight_decay'] = hyp['weight_decay'] * g
2018-09-20 16:03:19 +00:00
# Run model
pred = model(imgs)
2019-03-07 16:16:38 +00:00
# Compute loss
2019-08-23 15:37:29 +00:00
loss, loss_items = compute_loss(pred, targets, model)
2019-08-31 15:55:19 +00:00
if not torch.isfinite(loss):
print('WARNING: non-finite loss, ending training ', loss_items)
return results
2019-03-07 16:16:38 +00:00
2019-08-26 14:24:19 +00:00
# Scale loss by nominal batch_size of 64
loss *= batch_size / 64
2019-08-24 21:58:08 +00:00
2019-03-07 16:16:38 +00:00
# Compute gradient
2019-04-13 14:02:45 +00:00
if mixed_precision:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
2018-10-09 17:22:33 +00:00
2019-03-07 16:16:38 +00:00
# Accumulate gradient for x batches before optimizing
2019-08-23 11:31:32 +00:00
if ni % accumulate == 0:
2018-12-16 14:16:19 +00:00
optimizer.step()
optimizer.zero_grad()
2018-09-19 02:21:46 +00:00
2019-04-15 11:55:52 +00:00
# Print batch results
2019-05-23 10:32:11 +00:00
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
2019-07-16 15:56:39 +00:00
mem = torch.cuda.memory_cached() / 1E9 if torch.cuda.is_available() else 0 # (GB)
2019-08-24 14:43:43 +00:00
s = ('%10s' * 2 + '%10.3g' * 6) % (
2019-07-16 16:18:08 +00:00
'%g/%g' % (epoch, epochs - 1), '%.3gG' % mem, *mloss, len(targets), img_size)
2019-08-29 12:29:07 +00:00
pbar.set_description(s)
2019-08-24 19:20:25 +00:00
2019-08-29 12:29:07 +00:00
# end batch ------------------------------------------------------------------------------------------------
# Update scheduler
scheduler.step()
2019-08-29 12:29:07 +00:00
# Process epoch results
2019-08-24 19:35:56 +00:00
final_epoch = epoch + 1 == epochs
2019-08-24 19:20:25 +00:00
if opt.prebias:
print_model_biases(model)
else:
# Calculate mAP (always test final epoch, skip first 10 if opt.nosave)
if not (opt.notest or (opt.nosave and epoch < 10)) or final_epoch:
with torch.no_grad():
results, maps = test.test(cfg,
data,
batch_size=batch_size,
img_size=opt.img_size,
model=model,
conf_thres=0.001 if final_epoch and epoch > 0 else 0.1, # 0.1 for speed
save_json=final_epoch and epoch > 0 and 'coco.data' in data)
2019-04-05 13:34:42 +00:00
# Write epoch results
2019-09-17 22:54:07 +00:00
with open(results_file, 'a') as f:
f.write(s + '%10.3g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
2019-11-19 20:42:12 +00:00
if len(opt.name) and opt.bucket and not opt.prebias:
2019-11-20 21:26:50 +00:00
os.system('gsutil cp results.txt gs://%s/results%s.txt' % (opt.bucket, opt.name))
2019-04-05 13:34:42 +00:00
2019-08-08 20:30:34 +00:00
# Write Tensorboard results
2019-08-09 14:37:19 +00:00
if tb_writer:
2019-08-24 14:43:43 +00:00
x = list(mloss) + list(results)
titles = ['GIoU', 'Objectness', 'Classification', 'Train loss',
'Precision', 'Recall', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification']
2019-08-09 14:37:19 +00:00
for xi, title in zip(x, titles):
tb_writer.add_scalar(title, xi, epoch)
2019-08-08 20:30:34 +00:00
2019-08-24 19:20:25 +00:00
# Update best mAP
2019-10-15 23:36:13 +00:00
fitness = sum(results[4:]) # total loss
if fitness < best_fitness:
2019-07-02 16:21:28 +00:00
best_fitness = fitness
2019-03-19 08:38:32 +00:00
# Save training results
2019-09-11 21:11:01 +00:00
save = (not opt.nosave) or (final_epoch and not opt.evolve) or opt.prebias
if save:
2019-09-17 22:54:07 +00:00
with open(results_file, 'r') as f:
2019-07-08 16:00:19 +00:00
# Create checkpoint
chkpt = {'epoch': epoch,
'best_fitness': best_fitness,
2019-09-17 22:54:07 +00:00
'training_results': f.read(),
2019-07-08 16:00:19 +00:00
'model': model.module.state_dict() if type(
model) is nn.parallel.DistributedDataParallel else model.state_dict(),
2019-08-23 10:57:26 +00:00
'optimizer': None if final_epoch else optimizer.state_dict()}
2019-04-05 13:34:42 +00:00
2019-07-15 15:54:31 +00:00
# Save last checkpoint
torch.save(chkpt, last)
# Save best checkpoint
2019-07-02 16:21:28 +00:00
if best_fitness == fitness:
2019-04-02 16:04:04 +00:00
torch.save(chkpt, best)
2019-04-02 16:04:04 +00:00
# Save backup every 10 epochs (optional)
2019-04-02 12:07:14 +00:00
if epoch > 0 and epoch % 10 == 0:
2019-08-23 13:17:17 +00:00
torch.save(chkpt, wdir + 'backup%g.pt' % epoch)
2019-04-02 14:33:52 +00:00
2019-04-05 13:34:42 +00:00
# Delete checkpoint
2019-08-29 12:29:07 +00:00
del chkpt
# end epoch ----------------------------------------------------------------------------------------------------
2018-08-26 08:51:39 +00:00
2019-09-09 20:42:38 +00:00
# end training
2019-11-03 02:54:14 +00:00
if len(opt.name) and not opt.prebias:
2019-11-18 02:48:50 +00:00
fresults, flast, fbest = 'results%s.txt' % opt.name, 'last%s.pt' % opt.name, 'best%s.pt' % opt.name
os.rename('results.txt', fresults)
os.rename(wdir + 'last.pt', wdir + flast) if os.path.exists(wdir + 'last.pt') else None
os.rename(wdir + 'best.pt', wdir + fbest) if os.path.exists(wdir + 'best.pt') else None
# save to cloud
if opt.bucket:
2019-11-18 03:00:12 +00:00
os.system('gsutil cp %s %s gs://%s' % (fresults, wdir + flast, opt.bucket))
2019-11-18 02:48:50 +00:00
2019-08-23 11:45:49 +00:00
plot_results() # save as results.png
2019-08-24 19:39:25 +00:00
print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
2019-07-24 17:31:38 +00:00
dist.destroy_process_group() if torch.cuda.device_count() > 1 else None
2019-07-23 22:22:07 +00:00
torch.cuda.empty_cache()
2019-10-15 23:40:40 +00:00
2019-04-17 14:15:08 +00:00
return results
2018-08-26 08:51:39 +00:00
2019-10-05 10:45:10 +00:00
def prebias():
# trains output bias layers for 1 epoch and creates new backbone
if opt.prebias:
2019-11-20 21:24:50 +00:00
a = opt.img_weights # save settings
opt.img_weights = False # disable settings
2019-10-05 10:45:10 +00:00
train() # transfer-learn yolo biases for 1 epoch
create_backbone(last) # saved results as backbone.pt
2019-11-20 21:24:50 +00:00
2019-10-05 10:45:10 +00:00
opt.weights = wdir + 'backbone.pt' # assign backbone
opt.prebias = False # disable prebias
2019-11-20 21:24:50 +00:00
opt.img_weights = a # reset settings
2019-10-05 10:45:10 +00:00
2018-08-26 08:51:39 +00:00
if __name__ == '__main__':
parser = argparse.ArgumentParser()
2019-08-23 11:25:27 +00:00
parser.add_argument('--epochs', type=int, default=273) # 500200 batches at bs 16, 117263 images = 273 epochs
parser.add_argument('--batch-size', type=int, default=32) # effective bs = batch_size * accumulate = 16 * 4 = 64
parser.add_argument('--accumulate', type=int, default=2, help='batches to accumulate before optimizing')
2019-10-05 13:29:27 +00:00
parser.add_argument('--cfg', type=str, default='cfg/yolov3-spp.cfg', help='cfg file path')
2019-10-06 22:51:10 +00:00
parser.add_argument('--data', type=str, default='data/coco.data', help='*.data file path')
2019-08-23 11:31:32 +00:00
parser.add_argument('--multi-scale', action='store_true', help='adjust (67% - 150%) img_size every 10 batches')
2019-08-06 12:57:12 +00:00
parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)')
2019-07-08 13:02:20 +00:00
parser.add_argument('--rect', action='store_true', help='rectangular training')
2019-08-23 11:25:27 +00:00
parser.add_argument('--resume', action='store_true', help='resume training from last.pt')
parser.add_argument('--transfer', action='store_true', help='transfer learning')
2019-06-24 12:46:00 +00:00
parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
2019-04-17 15:27:51 +00:00
parser.add_argument('--notest', action='store_true', help='only test final epoch')
2019-07-01 15:17:29 +00:00
parser.add_argument('--evolve', action='store_true', help='evolve hyperparameters')
2019-07-08 16:32:31 +00:00
parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
2019-07-30 16:27:37 +00:00
parser.add_argument('--img-weights', action='store_true', help='select training images by weight')
2019-08-07 14:45:13 +00:00
parser.add_argument('--cache-images', action='store_true', help='cache images for faster training')
2019-11-20 21:40:24 +00:00
parser.add_argument('--weights', type=str, default='weights/ultralytics49.pt', help='initial weights')
2019-10-12 11:03:57 +00:00
parser.add_argument('--arc', type=str, default='default', help='yolo architecture') # defaultpw, uCE, uBCE
2019-08-24 18:55:01 +00:00
parser.add_argument('--prebias', action='store_true', help='transfer-learn yolo biases prior to training')
2019-09-09 20:42:38 +00:00
parser.add_argument('--name', default='', help='renames results.txt to results_name.txt if supplied')
2019-11-25 04:38:30 +00:00
parser.add_argument('--device', default='', help='device id (i.e. 0 or 0,1 or cpu)')
2019-09-11 12:25:48 +00:00
parser.add_argument('--adam', action='store_true', help='use adam optimizer')
parser.add_argument('--var', type=float, help='debug variable')
opt = parser.parse_args()
2019-09-17 22:38:49 +00:00
opt.weights = last if opt.resume else opt.weights
2019-05-03 16:14:16 +00:00
print(opt)
2019-11-25 04:29:29 +00:00
device = torch_utils.select_device(opt.device, apex=mixed_precision, batch_size=opt.batch_size)
2019-11-20 21:14:24 +00:00
if device.type == 'cpu':
mixed_precision = False
2019-11-18 20:01:17 +00:00
# scale hyp['obj'] by img_size (evolved at 416)
hyp['obj'] *= opt.img_size / 416.
2019-11-09 18:56:38 +00:00
2019-08-09 17:35:02 +00:00
tb_writer = None
2019-07-24 17:02:24 +00:00
if not opt.evolve: # Train normally
2019-08-09 14:37:19 +00:00
try:
# Start Tensorboard with "tensorboard --logdir=runs", view at http://localhost:6006/
from torch.utils.tensorboard import SummaryWriter
tb_writer = SummaryWriter()
except:
2019-08-09 17:35:02 +00:00
pass
2019-08-08 20:30:34 +00:00
2019-10-05 10:45:10 +00:00
prebias() # optional
2019-08-24 19:20:25 +00:00
train() # train normally
2019-07-24 17:02:24 +00:00
else: # Evolve hyperparameters (optional)
2019-06-24 12:46:00 +00:00
opt.notest = True # only test final epoch
opt.nosave = True # only save final checkpoint
2019-07-24 17:02:24 +00:00
if opt.bucket:
os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket) # download evolve.txt if exists
2019-04-17 15:51:39 +00:00
2019-09-10 15:28:36 +00:00
for _ in range(1): # generations to evolve
2019-07-24 18:16:35 +00:00
if os.path.exists('evolve.txt'): # if evolve.txt exists: select best hyps and mutate
2019-09-18 11:23:37 +00:00
# Select parent(s)
2019-07-24 17:02:24 +00:00
x = np.loadtxt('evolve.txt', ndmin=2)
2019-09-20 18:31:37 +00:00
parent = 'weighted' # parent selection method: 'single' or 'weighted'
if parent == 'single' or len(x) == 1:
x = x[fitness(x).argmax()]
elif parent == 'weighted': # weighted combination
n = min(10, x.shape[0]) # number to merge
x = x[np.argsort(-fitness(x))][:n] # top n mutations
w = fitness(x) - fitness(x).min() # weights
x = (x[:n] * w.reshape(n, 1)).sum(0) / w.sum() # new parent
2019-07-24 17:02:24 +00:00
for i, k in enumerate(hyp.keys()):
2019-08-06 12:36:12 +00:00
hyp[k] = x[i + 7]
2019-07-24 17:02:24 +00:00
# Mutate
2019-09-17 13:16:29 +00:00
np.random.seed(int(time.time()))
2019-10-08 22:14:27 +00:00
s = [.2, .2, .2, .2, .2, .2, .2, .0, .02, .2, .2, .2, .2, .2, .2, .2, .2, .2] # sigmas
2019-07-24 17:02:24 +00:00
for i, k in enumerate(hyp.keys()):
x = (np.random.randn(1) * s[i] + 1) ** 2.0 # plt.hist(x.ravel(), 300)
hyp[k] *= float(x) # vary by sigmas
2019-04-17 15:27:51 +00:00
2019-04-24 12:09:15 +00:00
# Clip to limits
2019-09-10 09:35:46 +00:00
keys = ['lr0', 'iou_t', 'momentum', 'weight_decay', 'hsv_s', 'hsv_v', 'translate', 'scale', 'fl_gamma']
2019-09-10 12:04:16 +00:00
limits = [(1e-5, 1e-2), (0.00, 0.70), (0.60, 0.98), (0, 0.001), (0, .9), (0, .9), (0, .9), (0, .9), (0, 3)]
2019-04-24 12:09:15 +00:00
for k, v in zip(keys, limits):
hyp[k] = np.clip(hyp[k], v[0], v[1])
2019-04-17 17:04:01 +00:00
2019-07-01 15:14:42 +00:00
# Train mutation
2019-10-05 10:45:10 +00:00
prebias()
2019-08-23 11:25:27 +00:00
results = train()
2019-04-17 15:27:51 +00:00
# Write mutation results
2019-07-25 15:49:54 +00:00
print_mutation(hyp, results, opt.bucket)
# Plot results
2019-07-26 10:00:43 +00:00
# plot_evolution_results(hyp)