CONFIG["BATCH_SIZE"],
shuffle=True,
collate_fn=ssd_utils.detection_collate)
if FROM_TRAIN_ITER > 1:
net.load_state_dict(t.load("outputs/SSD_%03d.pth" % (FROM_TRAIN_ITER - 1)))
index = 0
step_index = 0
# predict = j_m_ssd.SSDPredict(CONFIG["CLASSES"])
bar = j_bar.ProgressBar(CONFIG["EPOCH"], len(train_loader),
"Loss : %.3f; Total Loss : %.3f")
predict = ssd_predict.SSDPredict(CONFIG["CLASSES"])
net.train()
log = logger.Logger(CONFIG["LOG_DIR"])
for epoch in range(FROM_TRAIN_ITER, CONFIG["EPOCH"] + 1):
total_loss = 0.
t.cuda.empty_cache()
for i, (images, targets) in enumerate(train_loader):
index += i
if epoch >= 30:
LEARNING_RATE = 0.0005
if epoch >= 50:
LEARNING_RATE = 0.00025
if epoch >= 80:
LEARNING_RATE = 0.00001
if epoch >= 100:
LEARNING_RATE = 0.000005
for param_group in optimizer.param_groups:
param_group['lr'] = LEARNING_RATE
print("learning rate down")
lr = lr * scale
if batch == steps[i]:
break
else:
break
for param_group in optimizer.param_groups:
param_group['lr'] = lr / batch_size
return lr
processed_batches = 0
if FROM_TRAIN_ITER > 1:
model.load_state_dict(
torch.load("outputs/YOLOV3_%03d.pth" % (FROM_TRAIN_ITER - 1)))
log = logger.Logger("logs/")
predict = yolo_predict.YoloV3Predict(CONFIG["CLASSES"])
LEARNING_RATE = learning_rate
bar = j_bar.ProgressBar(max_epochs, len(train_loader),
"Loss:%.3f;Total Loss:%.3f")
for epoch in range(FROM_TRAIN_ITER, max_epochs + 1):
model.train()
total_loss = 0
torch.cuda.empty_cache()
# if epoch >= 1:
# LEARNING_RATE = 0.01
# if epoch >= 30:
# LEARNING_RATE = 0.005
# if epoch >= 60:
# LEARNING_RATE = 0.001
# if epoch >= 90: