# whole_train_data_idx = np.concatenate(train_data_idx)
# whole_num_train = num_train
for rd in range(1, num_round):
#print('Round {}'.format(rd))
##### Start Training Epochs
for epoch in range(0, args.Epoch):
#for epoch in range(0,1):
if args.ExpSum:
fid = open(summary_filepath, 'a')
else:
fid = None
printout('\n\nstart {:d}-th round {:d}-th epoch at {}\n'.format(
rd, epoch, time.ctime()),
write_flag=args.ExpSum,
fid=fid)
#### Shuffle Training Data --- close
data_sort = Loader.Shuffle_TrainSet()
#### Train One Epoch
train_avg_loss, train_avg_acc = TrainOp.TrainOneEpoch(
Loader, file_idx_list, data_idx_list, pts_idx_list)
printout('\nTrainingSet Avg Loss {:.4f} Avg Acc {:.2f}%'.format(
train_avg_loss, 100 * train_avg_acc),
write_flag=args.ExpSum,
fid=fid)
#### Evaluate One Epoch
##### Initialize Training Operations
TrainOp = util.ShapeNet_IncompleteSup()
TrainOp.SetLearningRate(LearningRate=args.LearningRate,
BatchSize=args.batchsize)
##### Define Network
TrainOp.DGCNN_SemiSup(batch_size=args.batchsize, point_num=3000)
##### Restore Checkpoint
#best_filepath = os.path.join(CHECKPOINT_PATH, 'Checkpoint_round-{}'.format('5'))
best_filepath = os.path.join(CHECKPOINT_PATH,
'Checkpoint_epoch-{}'.format('best'))
TrainOp.RestoreCheckPoint(best_filepath)
##### Start Testing
printout('\n\nstart Inference at {}\n'.format(time.ctime()))
#### Evaluate
avg_loss, avg_acc, perdata_miou, pershape_miou = TrainOp.Test(Loader, Eval)
print(
'\nAvg Loss {:.4f} Avg Acc {:.3f}% Avg PerData IoU {:.3f}% Avg PerCat IoU {:.3f}%'
.format(avg_loss, 100 * avg_acc, 100 * np.mean(perdata_miou),
100 * np.mean(pershape_miou)),
end='')
string = '\nEval PerShape IoU:'
for iou in pershape_miou:
string += ' {:.2f}%'.format(100 * iou)
print(string)
pts_idx_list = []
for b_i in range(tmp['pts_idx_list'].shape[1]):
pts_idx_list.append(tmp['pts_idx_list'][0, b_i][0])
else:
pts_idx_list = tmp['pts_idx_list']
##### Start Training Epochs
for epoch in range(0, args.Epoch):
if args.ExpRslt:
fid = open(summary_filepath, 'a')
else:
fid = None
printout('\n\nstart training {:d}-th epoch at {}\n'.format(
epoch, time.ctime()),
write_flag=args.ExpRslt,
fid=fid)
#### Shuffle Training Data
Loader.Shuffle_TrainSet()
#### Train One Epoch
if args.Style == 'Full':
train_avg_loss, train_avg_acc = TrainOp.TrainOneEpoch_Full(
Loader, pts_idx_list, args.batchsize)
elif args.Style == 'Plain':
train_avg_loss, train_avg_acc = TrainOp.TrainOneEpoch(
Loader, pts_idx_list, args.batchsize)
printout('\nTrainingSet Avg Loss {:.4f} Avg Acc {:.2f}%'.format(
train_avg_loss, 100 * train_avg_acc),