car-detection-bayes/train.py

194 lines
7.3 KiB
Python
Raw Normal View History

2018-08-26 08:51:39 +00:00
import argparse
import time
from models import *
from utils.datasets import *
from utils.utils import *
parser = argparse.ArgumentParser()
2018-09-24 01:34:12 +00:00
parser.add_argument('-epochs', type=int, default=1, help='number of epochs')
2018-08-26 08:51:39 +00:00
parser.add_argument('-batch_size', type=int, default=12, help='size of each image batch')
parser.add_argument('-data_config_path', type=str, default='cfg/coco.data', help='data config file path')
parser.add_argument('-cfg', type=str, default='cfg/yolov3.cfg', help='cfg file path')
parser.add_argument('-img_size', type=int, default=32 * 13, help='size of each image dimension')
2018-09-24 01:34:12 +00:00
parser.add_argument('-resume', default=True, help='resume training flag')
2018-08-26 08:51:39 +00:00
opt = parser.parse_args()
print(opt)
cuda = torch.cuda.is_available()
device = torch.device('cuda:0' if cuda else 'cpu')
random.seed(0)
np.random.seed(0)
torch.manual_seed(0)
if cuda:
torch.cuda.manual_seed(0)
torch.cuda.manual_seed_all(0)
torch.backends.cudnn.benchmark = True
2018-08-26 15:09:10 +00:00
2018-08-26 08:51:39 +00:00
def main(opt):
os.makedirs('checkpoints', exist_ok=True)
# Configure run
data_config = parse_data_config(opt.data_config_path)
num_classes = int(data_config['classes'])
2018-09-01 16:41:05 +00:00
if platform == 'darwin': # MacOS (local)
2018-09-02 22:35:46 +00:00
train_path = data_config['train']
2018-09-01 16:41:05 +00:00
else: # linux (cloud, i.e. gcp)
2018-08-26 18:30:47 +00:00
train_path = '../coco/trainvalno5k.part'
2018-08-26 08:51:39 +00:00
# Initialize model
model = Darknet(opt.cfg, opt.img_size)
# Get dataloader
2018-09-02 09:38:39 +00:00
dataloader = load_images_and_labels(train_path, batch_size=opt.batch_size, img_size=opt.img_size, augment=True)
2018-08-26 08:51:39 +00:00
2018-09-24 01:06:04 +00:00
# Reload saved optimizer state
2018-08-26 08:51:39 +00:00
start_epoch = 0
best_loss = float('inf')
if opt.resume:
2018-09-24 01:34:12 +00:00
checkpoint = torch.load('checkpoints/yolov3.pt', map_location='cpu')
2018-08-26 08:51:39 +00:00
model.load_state_dict(checkpoint['model'])
if torch.cuda.device_count() > 1:
print('Using ', torch.cuda.device_count(), ' GPUs')
model = nn.DataParallel(model)
model.to(device).train()
# # Transfer learning
# for i, (name, p) in enumerate(model.named_parameters()):
# #name = name.replace('module_list.', '')
# #print('%4g %70s %9s %12g %20s %12g %12g' % (
# # i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))
# if p.shape[0] != 650: # not YOLO layer
# p.requires_grad = False
# Set optimizer
# optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, model.parameters()))
2018-09-24 01:06:04 +00:00
optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=1e-3,
momentum=.9, weight_decay=5e-4, nesterov=True)
2018-08-26 08:51:39 +00:00
2018-09-24 01:06:04 +00:00
if checkpoint['optimizer'] is not None:
optimizer.load_state_dict(checkpoint['optimizer'])
start_epoch = checkpoint['epoch'] + 1
best_loss = checkpoint['best_loss']
2018-08-26 08:51:39 +00:00
del checkpoint # current, saved
else:
if torch.cuda.device_count() > 1:
print('Using ', torch.cuda.device_count(), ' GPUs')
model = nn.DataParallel(model)
model.to(device).train()
2018-09-20 16:03:19 +00:00
# Set optimizer
# optimizer = torch.optim.Adam(model.parameters(), lr=1e-4, weight_decay=5e-4)
optimizer = torch.optim.SGD(model.parameters(), lr=1e-3, momentum=.9, weight_decay=5e-4, nesterov=True)
2018-08-26 08:51:39 +00:00
# Set scheduler
# scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.99082, last_epoch=start_epoch - 1)
modelinfo(model)
t0, t1 = time.time(), time.time()
print('%10s' * 16 % (
2018-09-19 02:21:46 +00:00
'Epoch', 'Batch', 'x', 'y', 'w', 'h', 'conf', 'cls', 'total', 'P', 'R', 'nTargets', 'TP', 'FP', 'FN', 'time'))
2018-08-26 08:51:39 +00:00
for epoch in range(opt.epochs):
epoch += start_epoch
2018-09-20 16:03:19 +00:00
# Multi-Scale YOLO Training
# img_size = random.choice(range(10, 20)) * 32 # 320 - 608 pixels
2018-09-02 12:01:31 +00:00
# dataloader = load_images_and_labels(train_path, batch_size=opt.batch_size, img_size=img_size, augment=True)
2018-09-01 17:13:12 +00:00
# print('Running this epoch with image size %g' % img_size)
2018-08-26 08:51:39 +00:00
2018-09-20 16:03:19 +00:00
# Update scheduler (automatic)
2018-08-26 08:51:39 +00:00
# scheduler.step()
2018-09-20 16:03:19 +00:00
# Update scheduler (manual)
2018-08-26 08:51:39 +00:00
# for g in optimizer.param_groups:
2018-09-20 16:03:19 +00:00
# g['lr'] = 1e-3 * (g ** epoch) # 1/10th every [30, 50, 100, 250] epochs using g = [.926, .955, .977, .992]
2018-08-26 08:51:39 +00:00
ui = -1
rloss = defaultdict(float) # running loss
metrics = torch.zeros(4, num_classes)
for i, (imgs, targets) in enumerate(dataloader):
2018-09-19 02:21:46 +00:00
if sum([len(x) for x in targets]) < 1: # if no targets continue
continue
2018-09-20 16:03:19 +00:00
# SGD burn-in
2018-09-24 01:34:12 +00:00
# if (epoch == 0) & (i <= 1000):
# power = 4
# lr = 1e-3 * (i / 1000) ** power
# for g in optimizer.param_groups:
# g['lr'] = lr
# # print('SGD Burn-In LR = %9.5g' % lr, end='')
2018-09-20 16:03:19 +00:00
# Compute loss, compute gradient, update parameters
loss = model(imgs.to(device), targets, requestPrecision=True)
2018-09-19 02:21:46 +00:00
optimizer.zero_grad()
loss.backward()
optimizer.step()
2018-09-20 16:03:19 +00:00
# Compute running epoch-means of tracked metrics
2018-09-19 02:21:46 +00:00
ui += 1
metrics += model.losses['metrics']
for key, val in model.losses.items():
rloss[key] = (rloss[key] * ui + val) / (ui + 1)
# Precision
precision = metrics[0] / (metrics[0] + metrics[1] + 1e-16)
k = (metrics[0] + metrics[1]) > 0
if k.sum() > 0:
mean_precision = precision[k].mean()
else:
mean_precision = 0
# Recall
recall = metrics[0] / (metrics[0] + metrics[2] + 1e-16)
k = (metrics[0] + metrics[2]) > 0
if k.sum() > 0:
mean_recall = recall[k].mean()
else:
mean_recall = 0
s = ('%10s%10s' + '%10.3g' * 14) % (
2018-09-19 02:32:16 +00:00
'%g/%g' % (epoch, opt.epochs - 1), '%g/%g' % (i, len(dataloader) - 1), rloss['x'],
2018-09-19 02:21:46 +00:00
rloss['y'], rloss['w'], rloss['h'], rloss['conf'], rloss['cls'],
rloss['loss'], mean_precision, mean_recall, model.losses['nT'], model.losses['TP'],
model.losses['FP'], model.losses['FN'], time.time() - t1)
t1 = time.time()
print(s)
2018-08-26 08:51:39 +00:00
2018-08-26 09:42:34 +00:00
# Write epoch results
with open('results.txt', 'a') as file:
file.write(s + '\n')
2018-08-26 08:51:39 +00:00
# Update best loss
2018-09-19 02:21:46 +00:00
loss_per_target = rloss['loss'] / rloss['nT']
2018-08-26 08:51:39 +00:00
if loss_per_target < best_loss:
best_loss = loss_per_target
# Save latest checkpoint
checkpoint = {'epoch': epoch,
'best_loss': best_loss,
'model': model.state_dict(),
'optimizer': optimizer.state_dict()}
torch.save(checkpoint, 'checkpoints/latest.pt')
# Save best checkpoint
if best_loss == loss_per_target:
os.system('cp checkpoints/latest.pt checkpoints/best.pt')
# Save backup checkpoint
2018-08-26 22:00:25 +00:00
if (epoch > 0) & (epoch % 5 == 0):
2018-08-26 08:51:39 +00:00
os.system('cp checkpoints/latest.pt checkpoints/backup' + str(epoch) + '.pt')
# Save final model
dt = time.time() - t0
print('Finished %g epochs in %.2fs (%.2fs/epoch)' % (epoch, dt, dt / (epoch + 1)))
if __name__ == '__main__':
torch.cuda.empty_cache()
main(opt)
torch.cuda.empty_cache()