def main():
global args
args = (parser.parse_args())
use_cuda = cuda_model.ifUseCuda(args.gpu_id, args.multiGpu)
script_name_stem = dir_utils.get_stem(__file__)
save_directory = dir_utils.get_dir(
os.path.join(
project_root, 'ckpts',
'{:s}-{:s}-{:s}-split-{:d}-claweight-{:s}-{:.1f}-assgin{:.2f}-alpha{:.4f}-dim{:d}-dropout{:.4f}-seqlen{:d}-samplerate-{:d}-{:s}-{:s}'
.format(script_name_stem, args.dataset, args.eval_metrics,
args.split, str(args.set_cls_weight), args.cls_pos_weight,
args.hassign_thres, args.alpha, args.hidden_dim,
args.dropout, args.seq_len, args.sample_rate,
loss_type[args.EMD], match_type[args.hmatch])))
log_file = os.path.join(save_directory,
'log-{:s}.txt'.format(dir_utils.get_date_str()))
logger = log_utils.get_logger(log_file)
log_utils.print_config(vars(args), logger)
model = PointerNetwork(input_dim=args.input_dim,
embedding_dim=args.embedding_dim,
hidden_dim=args.hidden_dim,
max_decoding_len=args.net_outputs,
dropout=args.dropout,
n_enc_layers=2,
output_classes=2)
hassign_thres = args.hassign_thres
logger.info("Number of Params\t{:d}".format(
sum([p.data.nelement() for p in model.parameters()])))
logger.info('Saving logs to {:s}'.format(log_file))
if args.resume is not None:
ckpt_idx = 48
ckpt_filename = args.resume.format(ckpt_idx)
assert os.path.isfile(
ckpt_filename), 'Error: no checkpoint directory found!'
checkpoint = torch.load(ckpt_filename,
map_location=lambda storage, loc: storage)
model.load_state_dict(checkpoint['state_dict'], strict=False)
train_iou = checkpoint['IoU']
args.start_epoch = checkpoint['epoch']
logger.info("=> loading checkpoint '{}', current iou: {:.04f}".format(
ckpt_filename, train_iou))
model = cuda_model.convertModel2Cuda(model,
gpu_id=args.gpu_id,
multiGpu=args.multiGpu)
# get train/val split
if args.dataset == 'SumMe':
train_val_test_perms = np.arange(25)
elif args.dataset == 'TVSum':
train_val_test_perms = np.arange(50)
# fixed permutation
random.Random(0).shuffle(train_val_test_perms)
train_val_test_perms = train_val_test_perms.reshape([5, -1])
train_val_perms = np.delete(train_val_test_perms, args.split,
0).reshape([-1])
train_perms = train_val_perms[:17]
val_perms = train_val_perms[17:]
test_perms = train_val_test_perms[args.split]
logger.info(" training split: " + str(train_perms))
logger.info(" val split: " + str(val_perms))
logger.info(" test split: " + str(test_perms))
if args.location == 'home':
data_path = os.path.join(os.path.expanduser('~'), 'datasets')
else:
data_path = os.path.join('/nfs/%s/boyu/SDN' % (args.location),
'datasets')
train_dataset = vsSumLoader3_c3dd.cDataset(dataset_name=args.dataset,
split='train',
seq_length=args.seq_len,
overlap=0.9,
sample_rate=[args.sample_rate],
train_val_perms=train_perms,
data_path=data_path)
val_evaluator = Evaluator.Evaluator(dataset_name=args.dataset,
split='val',
seq_length=args.seq_len,
overlap=0.9,
sample_rate=[args.sample_rate],
sum_budget=0.15,
train_val_perms=val_perms,
eval_metrics=args.eval_metrics,
data_path=data_path)
test_evaluator = Evaluator.Evaluator(dataset_name=args.dataset,
split='test',
seq_length=args.seq_len,
overlap=0.9,
sample_rate=[args.sample_rate],
sum_budget=0.15,
train_val_perms=test_perms,
eval_metrics=args.eval_metrics,
data_path=data_path)
train_dataloader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
# val_dataloader = DataLoader(val_dataset,
# batch_size=args.batch_size,
# shuffle=False,
# num_workers=4)
model_optim = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=float(args.lr))
optim_scheduler = optim.lr_scheduler.ReduceLROnPlateau(model_optim,
'min',
patience=10)
alpha = args.alpha
# cls_weights = torch.FloatTensor([0.2, 1.0]).cuda()
if args.set_cls_weight:
cls_weights = torch.FloatTensor([
1. * train_dataset.n_positive_train_samples /
train_dataset.n_total_train_samples, args.cls_pos_weight
]).cuda()
else:
cls_weights = torch.FloatTensor([0.5, 0.5]).cuda()
logger.info(" total: {:d}, total pos: {:d}".format(
train_dataset.n_total_train_samples,
train_dataset.n_positive_train_samples))
logger.info(" classify weight: " + str(cls_weights[0]) +
str(cls_weights[1]))
for epoch in range(args.start_epoch, args.nof_epoch + args.start_epoch):
total_losses = AverageMeter()
loc_losses = AverageMeter()
cls_losses = AverageMeter()
Accuracy = AverageMeter()
IOU = AverageMeter()
ordered_IOU = AverageMeter()
model.train()
pbar = progressbar.ProgressBar(max_value=len(train_dataloader))
for i_batch, sample_batch in enumerate(train_dataloader):
pbar.update(i_batch)
feature_batch = Variable(sample_batch[0])
start_indices = Variable(sample_batch[1])
end_indices = Variable(sample_batch[2])
gt_valids = Variable(sample_batch[3])
# seq_labels = Variable(sample_batch[3])
if use_cuda:
feature_batch = feature_batch.cuda()
start_indices = start_indices.cuda()
end_indices = end_indices.cuda()
gt_positions = torch.stack([start_indices, end_indices], dim=-1)
head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(
feature_batch)
pred_positions = torch.stack([head_positions, tail_positions],
dim=-1)
if args.hmatch:
assigned_scores, assigned_locations, total_valid, total_iou = h_match.Assign_Batch_v2(
gt_positions,
pred_positions,
gt_valids,
thres=hassign_thres)
else:
assigned_scores, assigned_locations = f_match.Assign_Batch(
gt_positions,
pred_positions,
gt_valids,
thres=hassign_thres)
_, _, total_valid, total_iou = h_match.Assign_Batch_v2(
gt_positions,
pred_positions,
gt_valids,
thres=hassign_thres)
if total_valid > 0:
IOU.update(total_iou / total_valid, total_valid)
assigned_scores = Variable(torch.LongTensor(assigned_scores),
requires_grad=False)
assigned_locations = Variable(torch.LongTensor(assigned_locations),
requires_grad=False)
if use_cuda:
assigned_scores = assigned_scores.cuda()
assigned_locations = assigned_locations.cuda()
cls_scores = cls_scores.contiguous().view(-1,
cls_scores.size()[-1])
assigned_scores = assigned_scores.contiguous().view(-1)
cls_loss = F.cross_entropy(cls_scores,
assigned_scores,
weight=cls_weights)
if total_valid > 0:
assigned_head_positions = assigned_locations[:, :, 0]
assigned_head_positions = assigned_head_positions.contiguous(
).view(-1)
#
assigned_tail_positions = assigned_locations[:, :, 1]
assigned_tail_positions = assigned_tail_positions.contiguous(
).view(-1)
head_pointer_probs = head_pointer_probs.contiguous().view(
-1,
head_pointer_probs.size()[-1])
tail_pointer_probs = tail_pointer_probs.contiguous().view(
-1,
tail_pointer_probs.size()[-1])
assigned_head_positions = torch.masked_select(
assigned_head_positions, assigned_scores.byte())
assigned_tail_positions = torch.masked_select(
assigned_tail_positions, assigned_scores.byte())
head_pointer_probs = torch.index_select(
head_pointer_probs,
dim=0,
index=assigned_scores.nonzero().squeeze(1))
tail_pointer_probs = torch.index_select(
tail_pointer_probs,
dim=0,
index=assigned_scores.nonzero().squeeze(1))
if args.EMD:
assigned_head_positions = to_one_hot(
assigned_head_positions, args.seq_len)
assigned_tail_positions = to_one_hot(
assigned_tail_positions, args.seq_len)
prediction_head_loss = EMD_L2(head_pointer_probs,
assigned_head_positions,
needSoftMax=True)
prediction_tail_loss = EMD_L2(tail_pointer_probs,
assigned_tail_positions,
needSoftMax=True)
else:
prediction_head_loss = F.cross_entropy(
head_pointer_probs, assigned_head_positions)
prediction_tail_loss = F.cross_entropy(
tail_pointer_probs, assigned_tail_positions)
loc_losses.update(
prediction_head_loss.data.item() +
prediction_tail_loss.data.item(), feature_batch.size(0))
total_loss = alpha * (prediction_head_loss +
prediction_tail_loss) + cls_loss
else:
total_loss = cls_loss
model_optim.zero_grad()
total_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
model_optim.step()
cls_losses.update(cls_loss.data.item(), feature_batch.size(0))
total_losses.update(total_loss.item(), feature_batch.size(0))
logger.info(
"Train -- Epoch :{:06d}, LR: {:.6f},\tloss={:.4f}, \t c-loss:{:.4f}, \tloc-loss:{:.4f}\tcls-Accuracy:{:.4f}\tloc-Avg-IOU:{:.4f}\t topIOU:{:.4f}"
.format(epoch, model_optim.param_groups[0]['lr'], total_losses.avg,
cls_losses.avg, loc_losses.avg, Accuracy.avg, IOU.avg,
ordered_IOU.avg))
optim_scheduler.step(total_losses.avg)
model.eval()
# IOU = AverageMeter()
# pbar = progressbar.ProgressBar(max_value=len(val_evaluator))
# for i_batch, sample_batch in enumerate(val_dataloader):
# pbar.update(i_batch)
# feature_batch = Variable(sample_batch[0])
# start_indices = Variable(sample_batch[1])
# end_indices = Variable(sample_batch[2])
# gt_valids = Variable(sample_batch[3])
# # valid_indices = Variable(sample_batch[3])
# if use_cuda:
# feature_batch = feature_batch.cuda()
# start_indices = start_indices.cuda()
# end_indices = end_indices.cuda()
# gt_positions = torch.stack([start_indices, end_indices], dim=-1)
# head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(
# feature_batch)#Update: compared to the previous version, we now update the matching rules
# pred_positions = torch.stack([head_positions, tail_positions], dim=-1)
# pred_scores = cls_scores[:, :, -1]
# #TODO: should NOT change here for evaluation!
# assigned_scores, assigned_locations, total_valid, total_iou = h_match.Assign_Batch_v2(gt_positions, pred_positions, gt_valids, thres=hassign_thres)
# if total_valid>0:
# IOU.update(total_iou / total_valid, total_valid)
val_F1s = val_evaluator.Evaluate(model)
test_F1s = test_evaluator.Evaluate(model)
logger.info("Val -- Epoch :{:06d}, LR: {:.6f},\tF1s:{:.4f}".format(
epoch, model_optim.param_groups[0]['lr'], val_F1s))
logger.info("Test -- Epoch :{:06d},\tF1s:{:.4f}".format(
epoch, test_F1s))
if epoch % 1 == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'loss': total_losses.avg,
'cls_loss': cls_losses.avg,
'loc_loss': loc_losses.avg,
'IoU': IOU.avg,
'val_F1s': val_F1s,
'test_F1s': test_F1s
}, (epoch + 1),
file_direcotry=save_directory)
def main():
global args
args = (parser.parse_args())
use_cuda = cuda_model.ifUseCuda(args.gpu_id, args.multiGpu)
# Pretty print the run args
pp.pprint(vars(args))
model = PointerNetwork(input_dim=args.input_dim, embedding_dim=args.embedding_dim,
hidden_dim=args.hidden_dim, max_decoding_len=args.net_outputs, dropout=args.dropout, n_enc_layers=2)
hassign_thres = args.hassign_thres
print("Number of Params\t{:d}".format(sum([p.data.nelement() for p in model.parameters()])))
script_name_stem = dir_utils.get_stem(__file__)
save_directory = os.path.join(project_root, 'ckpts', '{:s}-assgin{:.2f}-alpha{:.4f}-dim{:d}-dropout{:.4f}-seqlen{:d}-ckpt'.
format(script_name_stem, hassign_thres, args.alpha, args.hidden_dim, args.dropout, args.seq_len))
print("Save ckpt to {:s}".format(save_directory))
if args.resume is not None:
ckpt_idx = 3
ckpt_filename = args.resume.format(ckpt_idx)
assert os.path.isfile(ckpt_filename), 'Error: no checkpoint directory found!'
checkpoint = torch.load(ckpt_filename, map_location=lambda storage, loc: storage)
model.load_state_dict(checkpoint['state_dict'], strict=False)
train_iou = checkpoint['IoU']
args.start_epoch = checkpoint['epoch']
print("=> loading checkpoint '{}', current iou: {:.04f}".format(ckpt_filename, train_iou))
model = cuda_model.convertModel2Cuda(model, gpu_id=args.gpu_id, multiGpu=args.multiGpu)
train_dataset = cDataset(seq_length=args.seq_len, overlap=0.9, sample_rate=[4], dataset_split='train', rdDrop=True, rdOffset=True)
val_dataset = cDataset(seq_length=args.seq_len, overlap=0.9, sample_rate=[4], dataset_split='val', rdDrop=False, rdOffset=False)
train_dataloader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
val_dataloader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=4)
model_optim = optim.Adam(filter(lambda p:p.requires_grad, model.parameters()), lr=float(args.lr))
optim_scheduler = optim.lr_scheduler.ReduceLROnPlateau(model_optim, 'min', patience=20)
alpha=args.alpha
cls_weights = torch.FloatTensor([0.05, 1.0]).cuda()
for epoch in range(args.start_epoch, args.nof_epoch+args.start_epoch):
total_losses = AverageMeter()
loc_losses = AverageMeter()
cls_losses = AverageMeter()
Accuracy = AverageMeter()
IOU = AverageMeter()
ordered_IOU = AverageMeter()
model.train()
pbar = progressbar.ProgressBar(max_value=len(train_dataloader))
for i_batch, sample_batch in enumerate(train_dataloader):
pbar.update(i_batch)
feature_batch = Variable(sample_batch[0])
start_indices = Variable(sample_batch[1])
end_indices = Variable(sample_batch[2])
gt_valids = Variable(sample_batch[3])
if use_cuda:
feature_batch = feature_batch.cuda()
start_indices = start_indices.cuda()
end_indices = end_indices.cuda()
gt_positions = torch.stack([start_indices, end_indices], dim=-1)
head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(feature_batch)
pred_positions = torch.stack([head_positions, tail_positions], dim=-1)
assigned_scores, assigned_locations = h_assign.Assign_Batch(gt_positions, pred_positions, gt_valids, thres=hassign_thres)
# if np.sum(assigned_scores) > 1:
# print("DEBUG")
# correct_predictions = np.sum(assigned_scores[:,:args.n_outputs])
# cls_rate = correct_predictions*1./np.sum(assigned_scores)
if np.sum(assigned_scores)>=1:
iou_rate, effective_positives = Metrics.get_avg_iou2(np.reshape(pred_positions.data.cpu().numpy(), (-1, 2)),
np.reshape(assigned_locations, (-1, 2)), np.reshape(assigned_scores,
assigned_scores.shape[
0] *
assigned_scores.shape[
1]))
IOU.update(iou_rate/(effective_positives), effective_positives)
# ordered_IOU.update(ordered_iou_rate/(args.batch_size*args.n_outputs),args.batch_size*args.n_outputs)
# n_effective_batches += 1
assigned_scores = Variable(torch.LongTensor(assigned_scores),requires_grad=False)
assigned_locations = Variable(torch.LongTensor(assigned_locations), requires_grad=False)
if use_cuda:
assigned_scores = assigned_scores.cuda()
assigned_locations = assigned_locations.cuda()
cls_scores = cls_scores.contiguous().view(-1, cls_scores.size()[-1])
assigned_scores = assigned_scores.contiguous().view(-1)
cls_loss = F.cross_entropy(cls_scores, assigned_scores, weight=cls_weights)
if torch.sum(assigned_scores)>0:
# print("HAHA")
assigned_head_positions = assigned_locations[:,:,0]
assigned_head_positions = assigned_head_positions.contiguous().view(-1)
#
assigned_tail_positions = assigned_locations[:,:,1]
assigned_tail_positions = assigned_tail_positions.contiguous().view(-1)
head_pointer_probs = head_pointer_probs.contiguous().view(-1, head_pointer_probs.size()[-1])
tail_pointer_probs = tail_pointer_probs.contiguous().view(-1, tail_pointer_probs.size()[-1])
# mask here: if there is non in assigned scores, no need to compute ...
assigned_head_positions = torch.masked_select(assigned_head_positions, assigned_scores.byte())
assigned_tail_positions = torch.masked_select(assigned_tail_positions, assigned_scores.byte())
head_pointer_probs = torch.index_select(head_pointer_probs, dim=0, index=assigned_scores.nonzero().squeeze(1))
tail_pointer_probs = torch.index_select(tail_pointer_probs, dim=0, index=assigned_scores.nonzero().squeeze(1))
assigned_head_positions = to_one_hot(assigned_head_positions, args.seq_len)
assigned_tail_positions = to_one_hot(assigned_tail_positions, args.seq_len)
prediction_head_loss = EMD_L2(head_pointer_probs, assigned_head_positions, needSoftMax=True)
prediction_tail_loss = EMD_L2(tail_pointer_probs, assigned_tail_positions, needSoftMax=True)
loc_losses.update(prediction_head_loss.data.item() + prediction_tail_loss.data.item(),
feature_batch.size(0))
total_loss = alpha * (prediction_head_loss + prediction_tail_loss) + cls_loss
else:
total_loss = cls_loss
model_optim.zero_grad()
total_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
model_optim.step()
cls_losses.update(cls_loss.data.item(), feature_batch.size(0))
total_losses.update(total_loss.item(), feature_batch.size(0))
print(
"Train -- Epoch :{:06d}, LR: {:.6f},\tloss={:.4f}, \t c-loss:{:.4f}, \tloc-loss:{:.4f}\tcls-Accuracy:{:.4f}\tloc-Avg-IOU:{:.4f}\t topIOU:{:.4f}".format(
epoch,
model_optim.param_groups[0]['lr'], total_losses.avg, cls_losses.avg, loc_losses.avg, Accuracy.avg, IOU.avg, ordered_IOU.avg))
optim_scheduler.step(total_losses.avg)
model.eval()
total_losses = AverageMeter()
loc_losses = AverageMeter()
cls_losses = AverageMeter()
Accuracy = AverageMeter()
IOU = AverageMeter()
ordered_IOU = AverageMeter()
pbar = progressbar.ProgressBar(max_value=len(val_dataloader))
for i_batch, sample_batch in enumerate(val_dataloader):
pbar.update(i_batch)
feature_batch = Variable(sample_batch[0])
start_indices = Variable(sample_batch[1])
end_indices = Variable(sample_batch[2])
gt_valids = Variable(sample_batch[3])
# valid_indices = Variable(sample_batch[3])
if use_cuda:
feature_batch = feature_batch.cuda()
start_indices = start_indices.cuda()
end_indices = end_indices.cuda()
gt_positions = torch.stack([start_indices, end_indices], dim=-1)
head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(
feature_batch)
pred_positions = torch.stack([head_positions, tail_positions], dim=-1)
assigned_scores, assigned_locations = h_assign.Assign_Batch(gt_positions, pred_positions, gt_valids, thres=hassign_thres)
# if np.sum(assigned_scores) > 1:
# print("DEBUG")
# correct_predictions = np.sum(assigned_scores[:,:args.n_outputs])
# cls_rate = correct_predictions*1./np.sum(assigned_scores)
if np.sum(assigned_scores) >= 1:
iou_rate, effective_positives = Metrics.get_avg_iou2(
np.reshape(pred_positions.data.cpu().numpy(), (-1, 2)),
np.reshape(assigned_locations, (-1, 2)), np.reshape(assigned_scores,
assigned_scores.shape[
0] *
assigned_scores.shape[
1]))
IOU.update(iou_rate / (effective_positives), effective_positives)
assigned_scores = Variable(torch.LongTensor(assigned_scores), requires_grad=False)
assigned_locations = Variable(torch.LongTensor(assigned_locations), requires_grad=False)
if use_cuda:
assigned_scores = assigned_scores.cuda()
assigned_locations = assigned_locations.cuda()
cls_scores = cls_scores.contiguous().view(-1, cls_scores.size()[-1])
assigned_scores = assigned_scores.contiguous().view(-1)
cls_loss = F.cross_entropy(cls_scores, assigned_scores, weight=cls_weights)
if torch.sum(assigned_scores)>0:
# print("HAHA")
assigned_head_positions = assigned_locations[:,:,0]
assigned_head_positions = assigned_head_positions.contiguous().view(-1)
#
assigned_tail_positions = assigned_locations[:,:,1]
assigned_tail_positions = assigned_tail_positions.contiguous().view(-1)
head_pointer_probs = head_pointer_probs.contiguous().view(-1, head_pointer_probs.size()[-1])
tail_pointer_probs = tail_pointer_probs.contiguous().view(-1, tail_pointer_probs.size()[-1])
# mask here: if there is non in assigned scores, no need to compute ...
assigned_head_positions = torch.masked_select(assigned_head_positions, assigned_scores.byte())
assigned_tail_positions = torch.masked_select(assigned_tail_positions, assigned_scores.byte())
head_pointer_probs = torch.index_select(head_pointer_probs, dim=0, index=assigned_scores.nonzero().squeeze(1))
tail_pointer_probs = torch.index_select(tail_pointer_probs, dim=0, index=assigned_scores.nonzero().squeeze(1))
assigned_head_positions = to_one_hot(assigned_head_positions, args.seq_len)
assigned_tail_positions = to_one_hot(assigned_tail_positions, args.seq_len)
prediction_head_loss = EMD_L2(head_pointer_probs, assigned_head_positions, needSoftMax=True)
prediction_tail_loss = EMD_L2(tail_pointer_probs, assigned_tail_positions, needSoftMax=True)
loc_losses.update(prediction_head_loss.data.item() + prediction_tail_loss.data.item(),
feature_batch.size(0))
total_loss = alpha * (prediction_head_loss + prediction_tail_loss) + cls_loss
else:
total_loss = cls_loss
cls_losses.update(cls_loss.data.item(), feature_batch.size(0))
total_losses.update(total_loss.item(), feature_batch.size(0))
print(
"Val -- Epoch :{:06d}, LR: {:.6f},\tloss={:.4f}, \t c-loss:{:.4f}, \tloc-loss:{:.4f}\tcls-Accuracy:{:.4f}\tloc-Avg-IOU:{:.4f}\t topIOU:{:.4f}".format(
epoch,
model_optim.param_groups[0]['lr'], total_losses.avg, cls_losses.avg, loc_losses.avg, Accuracy.avg,
IOU.avg, ordered_IOU.avg))
if epoch % 1 == 0:
save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'loss':total_losses.avg,
'cls_loss': cls_losses.avg,
'loc_loss': loc_losses.avg,
'IoU': IOU.avg}, (epoch+1), file_direcotry=save_directory)
def main():
global args
args = (parser.parse_args())
use_cuda = cuda_model.ifUseCuda(args.gpu_id, args.multiGpu)
script_name_stem = dir_utils.get_stem(__file__)
if args.resume is None:
save_directory = dir_utils.get_dir(
os.path.join(
project_root, 'ckpts', '{:s}'.format(args.dataset),
'{:s}-{:s}-assgin{:.2f}-alpha{:.4f}-dim{:d}-dropout{:.4f}-seqlen{:d}-{:s}-{:s}'
.format(script_name_stem, args.sufix, args.hassign_thres,
args.alpha, args.hidden_dim, args.dropout,
args.seq_len, loss_type[args.EMD],
match_type[args.hmatch])))
else:
save_directory = args.resume
log_file = os.path.join(save_directory,
'log-{:s}.txt'.format(dir_utils.get_date_str()))
logger = log_utils.get_logger(log_file)
log_utils.print_config(vars(args), logger)
model = PointerNetwork(input_dim=args.input_dim,
embedding_dim=args.embedding_dim,
hidden_dim=args.hidden_dim,
max_decoding_len=args.net_outputs,
dropout=args.dropout,
n_enc_layers=2)
logger.info("Number of Params\t{:d}".format(
sum([p.data.nelement() for p in model.parameters()])))
logger.info('Saving logs to {:s}'.format(log_file))
if args.resume is not None:
ckpt_idx = args.fileid
ckpt_filename = os.path.join(
args.resume, 'checkpoint_{:04d}.pth.tar'.format(ckpt_idx))
assert os.path.isfile(
ckpt_filename), 'Error: no checkpoint directory found!'
checkpoint = torch.load(ckpt_filename,
map_location=lambda storage, loc: storage)
model.load_state_dict(checkpoint['state_dict'], strict=False)
train_iou = checkpoint['IoU']
args.start_epoch = checkpoint['epoch']
logger.info("=> loading checkpoint '{}', current iou: {:.04f}".format(
ckpt_filename, train_iou))
model = cuda_model.convertModel2Cuda(model,
gpu_id=args.gpu_id,
multiGpu=args.multiGpu)
train_dataset = cDataset(dataset_split='train',
seq_length=args.seq_len,
sample_rate=[4],
rdOffset=True,
rdDrop=True)
val_dataset = cDataset(dataset_split='val',
seq_length=args.seq_len,
sample_rate=[4],
rdDrop=False,
rdOffset=False)
train_dataloader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
val_dataloader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=4)
model_optim = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=float(args.lr))
optim_scheduler = optim.lr_scheduler.ReduceLROnPlateau(model_optim,
'min',
patience=10)
# cls_weights = torch.FloatTensor([0.05, 1.0]).cuda()
cls_weights = None
widgets = [
'Test: ', ' -- [ ',
progressbar.Counter(), '|',
str(len(train_dataloader)), ' ] ',
progressbar.Bar(), ' cls loss: ',
progressbar.FormatLabel(''), ' loc loss: ',
progressbar.FormatLabel(''), ' IoU : ',
progressbar.FormatLabel(''), ' (',
progressbar.ETA(), ' ) '
]
# bar = progressbar.ProgressBar(max_value=step_per_epoch, widgets=widgets)
# bar.start()
for epoch in range(args.start_epoch, args.nof_epoch + args.start_epoch):
total_losses = AverageMeter()
loc_losses = AverageMeter()
cls_losses = AverageMeter()
matched_IOU = AverageMeter()
true_IOU = AverageMeter()
model.train()
pbar = progressbar.ProgressBar(max_value=len(train_dataloader),
widgets=widgets)
pbar.start()
for i_batch, sample_batch in enumerate(train_dataloader):
# pbar.update(i_batch)
feature_batch = Variable(sample_batch[0])
start_indices = Variable(sample_batch[1])
end_indices = Variable(sample_batch[2])
gt_valids = Variable(sample_batch[3])
gt_overlaps = Variable(sample_batch[4])
# seq_labels = Variable(sample_batch[3])
if use_cuda:
feature_batch = feature_batch.cuda()
start_indices = start_indices.cuda()
end_indices = end_indices.cuda()
gt_positions = torch.stack([start_indices, end_indices], dim=-1)
head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(
feature_batch)
pred_positions = torch.stack([head_positions, tail_positions],
dim=-1)
pred_scores = F.sigmoid(cls_scores)
if args.hmatch:
assigned_scores, assigned_overlaps, assigned_locations, total_valid, total_iou = h_match.Assign_Batch_v2_regression(
gt_positions,
pred_positions,
gt_valids,
gt_overlaps,
thres=args.hassign_thres)
else:
#FIXME: do it later!
assigned_scores, assigned_locations = f_match.Assign_Batch(
gt_positions,
pred_positions,
gt_valids,
thres=args.hassign_thres)
_, _, total_valid, total_iou = h_match.Assign_Batch_v2(
gt_positions,
pred_positions,
gt_valids,
thres=args.hassign_thres)
true_valid, true_iou = h_match.totalMatch_Batch(
gt_positions, pred_positions, gt_valids)
assert true_valid == total_valid, 'WRONG'
if total_valid > 0:
matched_IOU.update(total_iou / total_valid, total_valid)
true_IOU.update(true_iou / total_valid, total_valid)
assigned_scores = Variable(torch.LongTensor(assigned_scores),
requires_grad=False)
assigned_overlaps = Variable(torch.FloatTensor(assigned_overlaps),
requires_grad=False)
assigned_locations = Variable(torch.LongTensor(assigned_locations),
requires_grad=False)
if use_cuda:
assigned_scores = assigned_scores.cuda()
assigned_locations = assigned_locations.cuda()
assigned_overlaps = assigned_overlaps.cuda()
pred_scores = pred_scores.contiguous().view(-1)
assigned_scores = assigned_scores.contiguous().view(-1)
assigned_overlaps = assigned_overlaps.contiguous().view(-1)
cls_loss = ClsLocLoss2_OneClsRegression(pred_scores,
assigned_scores,
assigned_overlaps)
# cls_loss = F.cross_entropy(cls_scores, assigned_scores, weight=cls_weights)
if total_valid > 0:
assigned_head_positions = assigned_locations[:, :, 0]
assigned_head_positions = assigned_head_positions.contiguous(
).view(-1)
#
assigned_tail_positions = assigned_locations[:, :, 1]
assigned_tail_positions = assigned_tail_positions.contiguous(
).view(-1)
head_pointer_probs = head_pointer_probs.contiguous().view(
-1,
head_pointer_probs.size()[-1])
tail_pointer_probs = tail_pointer_probs.contiguous().view(
-1,
tail_pointer_probs.size()[-1])
assigned_head_positions = torch.masked_select(
assigned_head_positions, assigned_scores.byte())
assigned_tail_positions = torch.masked_select(
assigned_tail_positions, assigned_scores.byte())
head_pointer_probs = torch.index_select(
head_pointer_probs,
dim=0,
index=assigned_scores.nonzero().squeeze(1))
tail_pointer_probs = torch.index_select(
tail_pointer_probs,
dim=0,
index=assigned_scores.nonzero().squeeze(1))
if args.EMD:
assigned_head_positions = to_one_hot(
assigned_head_positions, args.seq_len)
assigned_tail_positions = to_one_hot(
assigned_tail_positions, args.seq_len)
prediction_head_loss = EMD_L2(head_pointer_probs,
assigned_head_positions,
needSoftMax=True)
prediction_tail_loss = EMD_L2(tail_pointer_probs,
assigned_tail_positions,
needSoftMax=True)
else:
prediction_head_loss = F.cross_entropy(
head_pointer_probs, assigned_head_positions)
prediction_tail_loss = F.cross_entropy(
tail_pointer_probs, assigned_tail_positions)
loc_losses.update(prediction_head_loss.data.item() +
prediction_tail_loss.data.item(),
total_valid) #FIXME
total_loss = (prediction_head_loss +
prediction_tail_loss) + cls_loss
else:
total_loss = cls_loss
model_optim.zero_grad()
total_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
model_optim.step()
cls_losses.update(cls_loss.data.item(), feature_batch.size(0))
total_losses.update(total_loss.item(), feature_batch.size(0))
widgets[-8] = progressbar.FormatLabel('{:04.4f}'.format(
cls_losses.avg))
widgets[-6] = progressbar.FormatLabel('{:04.4f}'.format(
loc_losses.avg))
widgets[-4] = progressbar.FormatLabel('{:01.4f}'.format(
matched_IOU.avg))
pbar.update(i_batch)
logger.info(
"Train -- Epoch :{:06d}, LR: {:.6f},\tloss={:.4f}, \t c-loss:{:.4f}, \tloc-loss:{:.4f}\tAvg-matched_IOU:{:.4f}\t Avg-true-IOU:{:.4f}"
.format(epoch, model_optim.param_groups[0]['lr'], total_losses.avg,
cls_losses.avg, loc_losses.avg, matched_IOU.avg,
true_IOU.avg))
train_iou = matched_IOU.avg
optim_scheduler.step(total_losses.avg)
model.eval()
matched_IOU = AverageMeter()
pbar = progressbar.ProgressBar(max_value=len(val_dataloader))
for i_batch, sample_batch in enumerate(val_dataloader):
pbar.update(i_batch)
feature_batch = Variable(sample_batch[0])
start_indices = Variable(sample_batch[1])
end_indices = Variable(sample_batch[2])
gt_valids = Variable(sample_batch[3])
# valid_indices = Variable(sample_batch[3])
if use_cuda:
feature_batch = feature_batch.cuda()
start_indices = start_indices.cuda()
end_indices = end_indices.cuda()
gt_positions = torch.stack([start_indices, end_indices], dim=-1)
head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(
feature_batch)
pred_positions = torch.stack([head_positions, tail_positions],
dim=-1)
# assigned_scores, assigned_locations, total_valid, total_iou = h_match.Assign_Batch_eval(gt_positions, pred_positions, gt_valids, thres=args.hassign_thres) #FIXME
matched_valid, matched_iou = h_match.totalMatch_Batch(
gt_positions, pred_positions, gt_valids)
if matched_valid > 0:
matched_IOU.update(matched_iou / matched_valid, matched_valid)
logger.info(
"Val -- Epoch :{:06d}, LR: {:.6f},\tloc-Avg-matched_IOU:{:.4f}".
format(
epoch,
model_optim.param_groups[0]['lr'],
matched_IOU.avg,
))
if epoch % 1 == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'loss': total_losses.avg,
'cls_loss': cls_losses.avg,
'loc_loss': loc_losses.avg,
'train-IOU': train_iou,
'IoU': matched_IOU.avg
}, (epoch + 1),
file_direcotry=save_directory)
def main():
global args
args = (parser.parse_args())
use_cuda = cuda_model.ifUseCuda(args.gpu_id, args.multiGpu)
script_name_stem = dir_utils.get_stem(__file__)
save_directory = dir_utils.get_dir(
os.path.join(
project_root, 'ckpts',
'Delete-{:s}-assgin{:.2f}-alpha{:.4f}-dim{:d}-dropout{:.4f}-seqlen{:d}-{:s}-{:s}'
.format(script_name_stem, args.hassign_thres, args.alpha,
args.hidden_dim, args.dropout, args.seq_len,
loss_type[args.EMD], match_type[args.hmatch])))
log_file = os.path.join(save_directory,
'log-{:s}.txt'.format(dir_utils.get_date_str()))
logger = chinese_utils.get_logger(log_file)
chinese_utils.print_config(vars(args), logger)
model = PointerNetwork(input_dim=args.input_dim,
embedding_dim=args.embedding_dim,
hidden_dim=args.hidden_dim,
max_decoding_len=args.net_outputs,
dropout=args.dropout,
n_enc_layers=2)
hassign_thres = args.hassign_thres
logger.info("Number of Params\t{:d}".format(
sum([p.data.nelement() for p in model.parameters()])))
logger.info('Saving logs to {:s}'.format(log_file))
if args.resume is not None:
ckpt_idx = 48
ckpt_filename = args.resume.format(ckpt_idx)
assert os.path.isfile(
ckpt_filename), 'Error: no checkpoint directory found!'
checkpoint = torch.load(ckpt_filename,
map_location=lambda storage, loc: storage)
model.load_state_dict(checkpoint['state_dict'], strict=False)
train_iou = checkpoint['IoU']
args.start_epoch = checkpoint['epoch']
logger.info("=> loading checkpoint '{}', current iou: {:.04f}".format(
ckpt_filename, train_iou))
model = cuda_model.convertModel2Cuda(model,
gpu_id=args.gpu_id,
multiGpu=args.multiGpu)
train_dataset = cDataset(dataset_split='train')
val_dataset = cDataset(dataset_split='val')
train_dataloader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
val_dataloader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=4)
model_optim = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=float(args.lr))
optim_scheduler = optim.lr_scheduler.ReduceLROnPlateau(model_optim,
'min',
patience=10)
alpha = args.alpha
# cls_weights = torch.FloatTensor([0.05, 1.0]).cuda()
for epoch in range(args.start_epoch, args.nof_epoch + args.start_epoch):
total_losses = AverageMeter()
loc_losses = AverageMeter()
cls_losses = AverageMeter()
Accuracy = AverageMeter()
IOU = AverageMeter()
ordered_IOU = AverageMeter()
model.train()
pbar = progressbar.ProgressBar(max_value=len(train_dataloader))
for i_batch, sample_batch in enumerate(train_dataloader):
pbar.update(i_batch)
feature_batch = Variable(sample_batch[0])
start_indices = Variable(sample_batch[1])
end_indices = Variable(sample_batch[2])
gt_valids = Variable(sample_batch[3])
# seq_labels = Variable(sample_batch[4])
if use_cuda:
feature_batch = feature_batch.cuda()
start_indices = start_indices.cuda()
end_indices = end_indices.cuda()
gt_positions = torch.stack([start_indices, end_indices], dim=-1)
head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(
feature_batch)
pred_positions = torch.stack([head_positions, tail_positions],
dim=-1)
if args.hmatch:
assigned_scores, assigned_locations, total_valid, total_iou = h_match.Assign_Batch_v2(
gt_positions,
pred_positions,
gt_valids,
thres=hassign_thres)
IOU.update(total_iou / total_valid, total_valid)
else:
assigned_scores, assigned_locations = f_match.Assign_Batch(
gt_positions,
pred_positions,
gt_valids,
thres=hassign_thres)
_, _, total_valid, total_iou = h_match.Assign_Batch_v2(
gt_positions,
pred_positions,
gt_valids,
thres=hassign_thres)
IOU.update(total_iou / total_valid, total_valid)
assigned_scores = Variable(torch.LongTensor(assigned_scores),
requires_grad=False)
assigned_locations = Variable(torch.LongTensor(assigned_locations),
requires_grad=False)
if use_cuda:
assigned_scores = assigned_scores.cuda()
assigned_locations = assigned_locations.cuda()
cls_scores = cls_scores.contiguous().view(-1,
cls_scores.size()[-1])
assigned_scores = assigned_scores.contiguous().view(-1)
cls_loss = F.cross_entropy(cls_scores, assigned_scores)
if total_valid > 0:
assigned_head_positions = assigned_locations[:, :, 0]
assigned_head_positions = assigned_head_positions.contiguous(
).view(-1)
#
assigned_tail_positions = assigned_locations[:, :, 1]
assigned_tail_positions = assigned_tail_positions.contiguous(
).view(-1)
head_pointer_probs = head_pointer_probs.contiguous().view(
-1,
head_pointer_probs.size()[-1])
tail_pointer_probs = tail_pointer_probs.contiguous().view(
-1,
tail_pointer_probs.size()[-1])
assigned_head_positions = torch.masked_select(
assigned_head_positions, assigned_scores.byte())
assigned_tail_positions = torch.masked_select(
assigned_tail_positions, assigned_scores.byte())
head_pointer_probs = torch.index_select(
head_pointer_probs,
dim=0,
index=assigned_scores.nonzero().squeeze(1))
tail_pointer_probs = torch.index_select(
tail_pointer_probs,
dim=0,
index=assigned_scores.nonzero().squeeze(1))
if args.EMD:
assigned_head_positions = to_one_hot(
assigned_head_positions, args.seq_len)
assigned_tail_positions = to_one_hot(
assigned_tail_positions, args.seq_len)
prediction_head_loss = EMD_L2(head_pointer_probs,
assigned_head_positions,
needSoftMax=True)
prediction_tail_loss = EMD_L2(tail_pointer_probs,
assigned_tail_positions,
needSoftMax=True)
else:
prediction_head_loss = F.cross_entropy(
head_pointer_probs, assigned_head_positions)
prediction_tail_loss = F.cross_entropy(
tail_pointer_probs, assigned_tail_positions)
loc_losses.update(
prediction_head_loss.data.item() +
prediction_tail_loss.data.item(), feature_batch.size(0))
total_loss = alpha * (prediction_head_loss +
prediction_tail_loss) + cls_loss
else:
total_loss = cls_loss
model_optim.zero_grad()
total_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
model_optim.step()
cls_losses.update(cls_loss.data.item(), feature_batch.size(0))
total_losses.update(total_loss.item(), feature_batch.size(0))
logger.info(
"Train -- Epoch :{:06d}, LR: {:.6f},\tloss={:.4f}, \t c-loss:{:.4f}, \tloc-loss:{:.4f}\tcls-Accuracy:{:.4f}\tloc-Avg-IOU:{:.4f}\t topIOU:{:.4f}"
.format(epoch, model_optim.param_groups[0]['lr'], total_losses.avg,
cls_losses.avg, loc_losses.avg, Accuracy.avg, IOU.avg,
ordered_IOU.avg))
optim_scheduler.step(total_losses.avg)
model.eval()
IOU = AverageMeter()
pbar = progressbar.ProgressBar(max_value=len(val_dataloader))
for i_batch, sample_batch in enumerate(val_dataloader):
pbar.update(i_batch)
feature_batch = Variable(sample_batch[0])
start_indices = Variable(sample_batch[1])
end_indices = Variable(sample_batch[2])
gt_valids = Variable(sample_batch[3])
# valid_indices = Variable(sample_batch[4])
if use_cuda:
feature_batch = feature_batch.cuda()
start_indices = start_indices.cuda()
end_indices = end_indices.cuda()
gt_positions = torch.stack([start_indices, end_indices], dim=-1)
head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(
feature_batch
) #Update: compared to the previous version, we now update the matching rules
pred_positions = torch.stack([head_positions, tail_positions],
dim=-1)
#TODO: should NOT change here for evaluation!
assigned_scores, assigned_locations, total_valid, total_iou = h_match.Assign_Batch_eval(
gt_positions, pred_positions, gt_valids, thres=hassign_thres)
IOU.update(total_iou / total_valid, total_valid)
logger.info(
"Val -- Epoch :{:06d}, LR: {:.6f},\tloc-Avg-IOU:{:.4f}".format(
epoch,
model_optim.param_groups[0]['lr'],
IOU.avg,
))
if epoch % 1 == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'loss': total_losses.avg,
'cls_loss': cls_losses.avg,
'loc_loss': loc_losses.avg,
'IoU': IOU.avg
}, (epoch + 1),
file_direcotry=save_directory)
def main():
# load data sets
global args
args = parser.parse_args()
pp.pprint(vars(args))
running_name = 'X'
use_cuda = cuda_model.ifUseCuda(args.gpu_id, args.multiGpu)
# use_cuda = False
train_file = 'data/example.train'
dev_file = 'data/example.dev'
test_file = 'data/example.test'
embedding_file = 'data/vec.txt'
map_file = 'map.pkl'
config_file = 'config_file_pytorch'
tag_file = 'tag.pkl'
embedding_easy_file = 'data/easy_embedding.npy'
train_sentences = load_sentences(train_file)
dev_sentences = load_sentences(dev_file)
test_sentences = load_sentences(test_file)
# train_sentences = dev_sentences
update_tag_scheme(train_sentences, args.tag_schema)
update_tag_scheme(test_sentences, args.tag_schema)
update_tag_scheme(dev_sentences, args.tag_schema)
if not os.path.isfile(tag_file):
_, tag_to_id, id_to_tag = tag_mapping(train_sentences)
with open(tag_file, "wb") as f:
pickle.dump([tag_to_id, id_to_tag], f)
else:
with open(tag_file, 'rb') as t:
tag_to_id, id_to_tag = pickle.load(t)
if not os.path.isfile(map_file):
# create dictionary for word
dico_chars_train = char_mapping(train_sentences)[0]
dico_chars, char_to_id, id_to_char = augment_with_pretrained(
dico_chars_train.copy(), embedding_file,
list(
itertools.chain.from_iterable([[w[0] for w in s]
for s in test_sentences])))
# _, tag_to_id, id_to_tag = tag_mapping(train_sentences)
with open(map_file, "wb") as f:
pickle.dump([char_to_id, id_to_char], f)
else:
with open(map_file, "rb") as f:
char_to_id, id_to_char = pickle.load(f)
# prepare data, get a collection of list containing index
train_data = prepare_dataset(train_sentences, char_to_id, tag_to_id)
dev_data = prepare_dataset(dev_sentences, char_to_id, tag_to_id)
test_data = prepare_dataset(test_sentences, char_to_id, tag_to_id)
print("%i / %i / %i sentences in train / dev / test." %
(len(train_data), len(dev_data), len(test_data)))
train_manager = BatchManager(train_data, args.batch_size)
dev_manager = BatchManager(dev_data, 50)
test_manager = BatchManager(test_data, 50)
# make path for store log and model if not exist
# make_path(FLAGS)
if os.path.isfile(config_file):
config = load_config(config_file)
else:
config = config_model(char_to_id, tag_to_id, args)
save_config(config, config_file)
# make_path(running_name)
save_places = dir_utils.save_places(running_name)
# log_path = os.path.join("log", FLAGS.log_file)
logger = get_logger(
os.path.join(save_places.log_save_dir,
'{:s}.txt'.format(dir_utils.get_date_str())))
print_config(config, logger)
logger.info("start training")
# loss = []
#Update: create model and embedding!
model = NERModel.CNERPointer(char_dim=args.char_dim,
seg_dim=args.seg_dim,
hidden_dim=args.hidden_dim,
max_length=15,
embedding_path=embedding_file,
id_to_word=id_to_char,
easy_load=embedding_easy_file)
print("Number of Params\t{:d}".format(
sum([p.data.nelement() for p in model.parameters()])))
#Update: this won't work!
# model = cuda_model.convertModel2Cuda(model, gpu_id=args.gpu_id, multiGpu=args.multiGpu)
if use_cuda:
model = model.cuda()
model_optim = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=float(args.lr))
optim_scheduler = optim.lr_scheduler.ReduceLROnPlateau(model_optim,
'min',
patience=10)
for epoch in range(args.start_epoch, args.nof_epoch + args.start_epoch):
total_losses = AverageMeter()
loc_losses = AverageMeter()
cls_losses = AverageMeter()
Accuracy = AverageMeter()
IOU = AverageMeter()
ordered_IOU = AverageMeter()
model.train()
pbar = progressbar.ProgressBar(max_value=train_manager.len_data)
for batch_idx, batch in enumerate(
train_manager.iter_batch(shuffle=True)):
pbar.update(batch_idx)
word_vectors = torch.LongTensor(batch[1])
seg_vectors = torch.LongTensor(batch[2])
batch_size = word_vectors.shape[0]
input_length = word_vectors.shape[1]
word_input = Variable(word_vectors)
seg_input = Variable(seg_vectors)
if use_cuda:
word_input = word_input.cuda()
seg_input = seg_input.cuda()
tagging_BIOUS = batch[3]
segments, max_len = convertBIOU2SegmentsBatch(
tagging_BIOUS, id_to_tag)
gt_positions, gt_valids = createPytorchLabels(segments, max_len)
head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(
word_input, seg_input, max_len)
pred_positions = torch.stack([head_positions, tail_positions],
dim=-1)
assigned_scores, assigned_locations = h_assign.Assign_Batch(
gt_positions,
pred_positions,
gt_valids,
thres=args.hassign_thres)
if np.sum(assigned_scores) >= 1:
iou_rate, effective_positives = Metrics.get_avg_iou2(
np.reshape(pred_positions.data.cpu().numpy(), (-1, 2)),
np.reshape(assigned_locations, (-1, 2)),
np.reshape(
assigned_scores,
assigned_scores.shape[0] * assigned_scores.shape[1]))
IOU.update(iou_rate / (effective_positives),
effective_positives)
# ordered_IOU.update(ordered_iou_rate/(args.batch_size*args.n_outputs),args.batch_size*args.n_outputs)
# n_effective_batches += 1
assigned_scores = Variable(torch.LongTensor(assigned_scores),
requires_grad=False)
assigned_locations = Variable(torch.LongTensor(assigned_locations),
requires_grad=False)
if use_cuda:
assigned_scores = assigned_scores.cuda()
assigned_locations = assigned_locations.cuda()
cls_scores = cls_scores.contiguous().view(-1,
cls_scores.size()[-1])
assigned_scores = assigned_scores.contiguous().view(-1)
cls_loss = F.cross_entropy(cls_scores, assigned_scores)
if torch.sum(assigned_scores) > 0:
# print("HAHA")
assigned_head_positions = assigned_locations[:, :, 0]
assigned_head_positions = assigned_head_positions.contiguous(
).view(-1)
#
assigned_tail_positions = assigned_locations[:, :, 1]
assigned_tail_positions = assigned_tail_positions.contiguous(
).view(-1)
head_pointer_probs = head_pointer_probs.contiguous().view(
-1,
head_pointer_probs.size()[-1])
tail_pointer_probs = tail_pointer_probs.contiguous().view(
-1,
tail_pointer_probs.size()[-1])
# mask here: if there is non in assigned scores, no need to compute ...
assigned_head_positions = torch.masked_select(
assigned_head_positions, assigned_scores.byte())
assigned_tail_positions = torch.masked_select(
assigned_tail_positions, assigned_scores.byte())
head_pointer_probs = torch.index_select(
head_pointer_probs,
dim=0,
index=assigned_scores.nonzero().squeeze(1))
tail_pointer_probs = torch.index_select(
tail_pointer_probs,
dim=0,
index=assigned_scores.nonzero().squeeze(1))
assigned_head_positions = to_one_hot(assigned_head_positions,
input_length)
assigned_tail_positions = to_one_hot(assigned_tail_positions,
input_length)
prediction_head_loss = EMD_L2(head_pointer_probs,
assigned_head_positions,
needSoftMax=True)
prediction_tail_loss = EMD_L2(tail_pointer_probs,
assigned_tail_positions,
needSoftMax=True)
loc_losses.update(
prediction_head_loss.data.item() +
prediction_tail_loss.data.item(), batch_size)
total_loss = args.alpha * (prediction_head_loss +
prediction_tail_loss) + cls_loss
else:
total_loss = cls_loss
model_optim.zero_grad()
total_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
model_optim.step()
cls_losses.update(cls_loss.data.item(), batch_size)
total_losses.update(total_loss.item(), batch_size)
logger.info(
"Train -- Epoch :{:06d}, LR: {:.6f},\tloss={:.4f}, \t c-loss:{:.4f}, \tloc-loss:{:.4f}\tcls-Accuracy:{:.4f}\tloc-Avg-IOU:{:.4f}\t topIOU:{:.4f}"
.format(epoch, model_optim.param_groups[0]['lr'], total_losses.avg,
cls_losses.avg, loc_losses.avg, Accuracy.avg, IOU.avg,
ordered_IOU.avg))
optim_scheduler.step(total_losses.avg)
total_losses = AverageMeter()
loc_losses = AverageMeter()
cls_losses = AverageMeter()
Accuracy = AverageMeter()
IOU = AverageMeter()
ordered_IOU = AverageMeter()
model.eval()
pbar = progressbar.ProgressBar(max_value=dev_manager.len_data)
for batch_idx, batch in enumerate(
dev_manager.iter_batch(shuffle=True)):
pbar.update(batch_idx)
word_vectors = torch.LongTensor(batch[1])
seg_vectors = torch.LongTensor(batch[2])
batch_size = word_vectors.shape[0]
input_length = word_vectors.shape[1]
word_input = Variable(word_vectors)
seg_input = Variable(seg_vectors)
if use_cuda:
word_input = word_input.cuda()
seg_input = seg_input.cuda()
tagging_BIOUS = batch[3]
segments, max_len = convertBIOU2SegmentsBatch(
tagging_BIOUS, id_to_tag)
head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(
word_input, seg_input, max_len)
pred_positions = torch.stack([head_positions, tail_positions],
dim=-1)
gt_positions, gt_valids = createPytorchLabels(segments, max_len)
assigned_scores, assigned_locations = h_assign.Assign_Batch(
gt_positions,
pred_positions,
gt_valids,
thres=args.hassign_thres)
if np.sum(assigned_scores) >= 1:
iou_rate, effective_positives = Metrics.get_avg_iou2(
np.reshape(pred_positions.data.cpu().numpy(), (-1, 2)),
np.reshape(assigned_locations, (-1, 2)),
np.reshape(
assigned_scores,
assigned_scores.shape[0] * assigned_scores.shape[1]))
IOU.update(iou_rate / (effective_positives),
effective_positives)
# ordered_IOU.update(ordered_iou_rate/(args.batch_size*args.n_outputs),args.batch_size*args.n_outputs)
# n_effective_batches += 1
assigned_scores = Variable(torch.LongTensor(assigned_scores),
requires_grad=False)
assigned_locations = Variable(torch.LongTensor(assigned_locations),
requires_grad=False)
if use_cuda:
assigned_scores = assigned_scores.cuda()
assigned_locations = assigned_locations.cuda()
cls_scores = cls_scores.contiguous().view(-1,
cls_scores.size()[-1])
assigned_scores = assigned_scores.contiguous().view(-1)
cls_loss = F.cross_entropy(cls_scores, assigned_scores)
if torch.sum(assigned_scores) > 0:
# print("HAHA")
assigned_head_positions = assigned_locations[:, :, 0]
assigned_head_positions = assigned_head_positions.contiguous(
).view(-1)
#
assigned_tail_positions = assigned_locations[:, :, 1]
assigned_tail_positions = assigned_tail_positions.contiguous(
).view(-1)
head_pointer_probs = head_pointer_probs.contiguous().view(
-1,
head_pointer_probs.size()[-1])
tail_pointer_probs = tail_pointer_probs.contiguous().view(
-1,
tail_pointer_probs.size()[-1])
# mask here: if there is non in assigned scores, no need to compute ...
assigned_head_positions = torch.masked_select(
assigned_head_positions, assigned_scores.byte())
assigned_tail_positions = torch.masked_select(
assigned_tail_positions, assigned_scores.byte())
head_pointer_probs = torch.index_select(
head_pointer_probs,
dim=0,
index=assigned_scores.nonzero().squeeze(1))
tail_pointer_probs = torch.index_select(
tail_pointer_probs,
dim=0,
index=assigned_scores.nonzero().squeeze(1))
assigned_head_positions = to_one_hot(assigned_head_positions,
input_length)
assigned_tail_positions = to_one_hot(assigned_tail_positions,
input_length)
prediction_head_loss = EMD_L2(head_pointer_probs,
assigned_head_positions,
needSoftMax=True)
prediction_tail_loss = EMD_L2(tail_pointer_probs,
assigned_tail_positions,
needSoftMax=True)
loc_losses.update(
prediction_head_loss.data.item() +
prediction_tail_loss.data.item(), batch_size)
total_loss = args.alpha * (prediction_head_loss +
prediction_tail_loss) + cls_loss
else:
total_loss = cls_loss
# model_optim.zero_grad()
# total_loss.backward()
# torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
# model_optim.step()
cls_losses.update(cls_loss.data.item(), batch_size)
total_losses.update(total_loss.item(), batch_size)
logger.info(
"Val -- Epoch :{:06d}, LR: {:.6f},\tloss={:.4f}, \t c-loss:{:.4f}, \tloc-loss:{:.4f}\tcls-Accuracy:{:.4f}\tloc-Avg-IOU:{:.4f}\t topIOU:{:.4f}"
.format(epoch, model_optim.param_groups[0]['lr'], total_losses.avg,
cls_losses.avg, loc_losses.avg, Accuracy.avg, IOU.avg,
ordered_IOU.avg))
if epoch % 1 == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'loss': total_losses.avg,
'cls_loss': cls_losses.avg,
'loc_loss': loc_losses.avg,
'IoU': IOU.avg
}, (epoch + 1),
file_direcotry=save_places.model_save_dir)