def fit(self, samples_train, samples_val):
epoch_start, epoch_end = 0, 200
best_val_mAP = 0
best_stats = None
# Logs stats for each epoch and saves them as .csv at the end
epoch_logger = utils.EpochLogger(self.name +
'-split_{}'.format(self.split))
net = DataParallel(self.net).cuda()
optimizer = self.optimizer
lr_scheduler = utils.CyclicLR(optimizer, 5, {
0: (1e-4, 1e-6),
100: (0.5e-4, 1e-6),
160: (1e-5, 1e-6),
})
if os.path.exists(os.path.join(self.path, 'checkpoint')):
epoch_start, epoch_logger, best_val_mAP = self.load_checkpoint(
net, optimizer)
print(
'Loading from checkpoint. Continuing from epoch {}...'.format(
epoch_start))
print(epoch_logger)
# Training
for e in range(epoch_start, epoch_end):
lr_scheduler.step(e)
stats_train = self.train(net, samples_train, optimizer, e)
stats_val = self.validate(net, samples_val, e)
stats = {**stats_train, **stats_val}
epoch_logger.add_epoch(stats)
current_mAP = stats_val['val_mAP']
if current_mAP > best_val_mAP:
best_val_mAP = current_mAP
best_stats = stats
self.save()
self.save_checkpoint(net, optimizer, e, {'epoch': epoch_logger},
best_val_mAP)
# Post training
epoch_logger.save()
return best_stats
def fit(self, samples_train, samples_val):
net = DataParallel(self.net)
epochs = 200
optimizer = SGD(net.parameters(),
lr=1e-2,
weight_decay=1e-4,
momentum=0.9,
nesterov=True)
lr_scheduler = utils.PolyLR(optimizer,
51,
0.9,
steps={
0: 1e-2,
50: 0.5 * 1e-2,
100: 0.5 * 0.5 * 1e-2,
150: 0.5 * 0.5 * 0.5 * 1e-2,
})
best_val_mAP = 0
best_stats = None
# Logs stats for each epoch and saves them as .csv at the end
epoch_logger = utils.EpochLogger(self.name +
'-split_{}'.format(self.split))
# Training
for e in range(epochs):
lr_scheduler.step(e)
stats_train = self.train(net, samples_train, optimizer, e)
stats_val = self.validate(net, samples_val, e)
stats = {**stats_train, **stats_val}
epoch_logger.add_epoch(stats)
current_mAP = stats_val['val_mAP']
if current_mAP > best_val_mAP:
best_val_mAP = current_mAP
best_stats = stats
self.save()
# Post training
epoch_logger.save()
return best_stats
def fit(self, samples_train, samples_val):
net = DataParallel(self.net)
optimizer = NDAdam(net.parameters(), lr=1e-4, weight_decay=1e-4)
lr_scheduler = utils.CyclicLR(optimizer, 5, {
0: (1e-4, 1e-6),
100: (0.5e-4, 1e-6),
160: (1e-5, 1e-6),
})
epochs = 200
best_val_mAP = 0
best_stats = None
# Logs stats for each epoch and saves them as .csv at the end
epoch_logger = utils.EpochLogger(self.name +
'-split_{}'.format(self.split))
# Training
for e in range(epochs):
lr_scheduler.step(e)
stats_train = self.train(net, samples_train, optimizer, e)
stats_val = self.validate(net, samples_val, e)
stats = {**stats_train, **stats_val}
epoch_logger.add_epoch(stats)
current_mAP = stats_val['val_mAP']
if current_mAP > best_val_mAP:
best_val_mAP = current_mAP
best_stats = stats
self.save()
# Post training
epoch_logger.save()
return best_stats
