def create():
print("==========\nArgs:{}\n==========".format(args))
print("Goal: randomly split data for {} times, {:.1%} for training and the rest for testing".format(args.num_splits, args.train_percent))
print("Loading dataset from {}".format(args.dataset))
dataset = h5py.File(args.dataset, 'r')
keys = dataset.keys()
num_videos = len(keys)
num_train = int(math.ceil(num_videos * args.train_percent))
num_test = num_videos - num_train
print("Split breakdown: # total videos {}. # train videos {}. # test videos {}".format(num_videos, num_train, num_test))
splits = []
for split_idx in range(args.num_splits):
train_keys, test_keys = split_random(keys, num_videos, num_train)
splits.append({
'train_keys': train_keys,
'test_keys': test_keys,
})
if args.save_name is None:
save_name = dir_utils.get_stem(args.dataset)
else:
save_name = args.save_name
save_dir = dir_utils.get_dir(args.save_dir)
saveto = osp.join(save_dir, save_name + '.json')
write_json(splits, saveto)
print("Splits saved to {}".format(saveto))
dataset.close()
def parse_c3dfiles(path):
"""
Parse directories holding extracted frames from standard benchmarks
"""
print('parse frames under folder {}'.format(path))
feature_files = glob.glob(os.path.join(path, '*.mat'))
# check RGB
frame_dict = {}
for i, s_feature_file in enumerate(feature_files):
s_feature = scipy.io.loadmat(s_feature_file)['relu6']
n_frames = s_feature.shape[0]
k = dir_utils.get_stem(s_feature_file)
frame_dict[k] = (s_feature_file, n_frames)
print('frame folder analysis done')
return frame_dict
def config_model(char_to_id, tag_to_id, args=None):
config = OrderedDict()
config["model_type"] = dir_utils.get_stem(__file__)
config["num_chars"] = len(char_to_id)
config["char_dim"] = args.char_dim
config["num_tags"] = len(tag_to_id)
config["seg_dim"] = args.seg_dim
config["lstm_dim"] = args.hidden_dim
config["batch_size"] = args.batch_size
config["emb_file"] = 'data/vec.txt'
# config["clip"] = FLAGS.clip
config["dropout"] = args.dropout # Update: 0 means not drop out at all
config["tag_schema"] = args.tag_schema
config['train_file'] = 'data/example.train'
config['dev_file'] = 'data/example.dev'
config['test_file'] = 'data/example.test'
config['map_file'] = 'map.pkl'
return config
def main():
PCA_info = pkl.load(open('pca_c3dd_fc7_val_annot_thumos14.pkl', 'rb'))
x_mean = PCA_info['x_mean']
U = PCA_info['U']
num_red_dim = 500
feature_directory = '/home/zwei/datasets/THUMOS14/features/c3dd-fc7'
feature_files = glob.glob(os.path.join(feature_directory, '*.mat'))
target_directory = dir_utils.get_dir(
'/home/zwei/datasets/THUMOS14/features/c3dd-fc7-red500')
print time.ctime(), 'start'
pbar = progressbar.ProgressBar(max_value=len(feature_files))
for feat_idx, s_feature_file in enumerate(feature_files):
s_file_stem = dir_utils.get_stem(s_feature_file)
output_file = os.path.join(target_directory,
'{:s}.npy'.format(s_file_stem))
pbar.update(feat_idx)
s_feature = scipy.io.loadmat(s_feature_file)['fc7']
s_feature_red = np.dot(s_feature - x_mean, U[:, :num_red_dim])
np.save(output_file, s_feature_red)
def parse_c3ddfiles(path):
"""
Parse directories holding extracted frames from standard benchmarks
"""
print('parse frames under folder {}'.format(path))
feature_files = glob.glob(os.path.join(path, '*.npy'))
frame_directories = glob.glob(
os.path.join('/home/zwei/datasets/THUMOS14/frame', '*/'))
frame_directory = '/home/zwei/datasets/THUMOS14/frame'
# check RGB
frame_dict = {}
for i, s_feature_file in enumerate(feature_files):
s_feature = np.load(s_feature_file)
n_frames = s_feature.shape[0]
k = dir_utils.get_stem(s_feature_file)
s_frames = len(glob.glob(os.path.join(frame_directory, k, '*.jpg')))
assert s_frames == n_frames, 'BUG'
frame_dict[k] = (s_feature_file, n_frames)
print('frame folder analysis done')
return frame_dict
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)
import scipy.io as sio
import sys, os
from PyUtils import dir_utils
import numpy as np
import glob
import progressbar
feature_directory = '/home/zwei/datasets/THUMOS14/features/c3d_feat_dense'
target_directory = dir_utils.get_dir(
'/home/zwei/datasets/THUMOS14/features/c3d-dense')
feature_lists = glob.glob(os.path.join(feature_directory, '*.mat'))
# video_name = 'video_validation_0000940.mat'
pbar = progressbar.ProgressBar(max_value=len(feature_lists))
for file_id, s_feature_path in enumerate(feature_lists):
pbar.update(file_id)
# s_feature_name = os.path.basename(s_feature_path)
s_feature_stem = dir_utils.get_stem(s_feature_path)
# feature_path = os.path.join(, video_name)
s_full_feature = sio.loadmat(s_feature_path)['relu6']
np.save(os.path.join(target_directory, '{:s}.npy'.format(s_feature_stem)),
s_full_feature)
print("DEBUG")
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, 'L2', 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, output_classes=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_locations, total_valid, total_iou = h_match.Assign_Batch_v2(gt_positions, pred_positions, gt_valids, thres=args.hassign_thres)
else:
#FIXME: do it later!
assigned_scores, assigned_locations, total_valid, total_iou = f_match.Assign_Batch_v2(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_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 = Simple_L2(head_pointer_probs, assigned_head_positions, needSoftMax=True)
prediction_tail_loss = Simple_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 = 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(), 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():
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 = '{:s}-assgin{:.2f}-alpha{:.4f}-dim{:d}-dropout{:.4f}-ckpt'.format(
script_name_stem, hassign_thres, args.alpha, args.hidden_dim,
args.dropout)
print("Save ckpt to {:s}".format(save_directory))
if args.resume is not None:
ckpt_idx = 7
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 = THUMOST14(seq_length=args.seq_len, overlap=0.9, sample_rate=[4], dataset_split='train',rdDrop=True,rdOffset=True)
val_dataset = THUMOST14(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(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])
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) > 0:
print "Output at {:d}".format(i_batch)
# n_valid = valid_indices.data[0, 0]
# view_idx = valid_indices.nonzero()[0][0].item()
# n_valid = valid_indices[view_idx, 0].item()
print "GT:"
print(assigned_locations[0])
print("Pred")
print(pred_positions[0])
_, head_sort = head_pointer_probs[0, 0, :].sort()
_, tail_sort = tail_pointer_probs[0, 0, :].sort()
print("END of {:d}".format(i_batch))
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}-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 = BaseLSTMNetwork(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 = MNIST(dataset_split='train')
val_dataset =MNIST(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])
labels = Variable(sample_batch[1])
if use_cuda:
feature_batch = feature_batch.cuda()
labels = labels.cuda()
# end_indices = end_indices.cuda()
pred_labels = model(feature_batch)
labels = labels.contiguous().view(-1)
pred_labels = pred_labels.contiguous().view(-1, pred_labels.size()[-1])
pred_probs = F.softmax(pred_labels, dim=1)[:, 1]
pred_probs[pred_probs>0.5] = 1
pred_probs[pred_probs<=0.5] = -1
n_positives = torch.sum(labels).item()
iou = torch.sum(pred_probs==labels.float()).item()*1. / n_positives
IOU.update(iou, 1.)
total_loss = F.cross_entropy(pred_labels, labels)
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])
labels = Variable(sample_batch[1])
if use_cuda:
feature_batch = feature_batch.cuda()
labels = labels.cuda()
labels = labels.contiguous().view(-1)
pred_labels = model(feature_batch)
pred_labels = pred_labels.contiguous().view(-1, pred_labels.size()[-1])
pred_probs = F.softmax(pred_labels, dim=1)[:, 1]
n_positives = torch.sum(labels).item()
pred_probs[pred_probs > 0.5] = 1
pred_probs[pred_probs <= 0.5] = -1
iou = torch.sum(pred_probs == labels.float()).item() * 1. / n_positives
IOU.update(iou, 1.)
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():
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}-decoderatio-{:.2f}-alpha{:.4f}-dim{:d}-dropout{:.4f}'
.format(script_name_stem, args.dataset, args.eval_metrics,
args.split, args.decode_ratio, args.alpha, args.hidden_dim,
args.dropout)))
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)
# get train/val split
if args.dataset == 'SumMe':
train_val_perms = np.arange(25)
elif args.dataset == 'TVSum':
train_val_perms = np.arange(50)
# fixed permutation
random.Random(0).shuffle(train_val_perms)
train_val_perms = train_val_perms.reshape([5, -1])
train_perms = np.delete(train_val_perms, args.split, 0).reshape([-1])
val_perms = train_val_perms[args.split]
logger.info(" training split: " + str(train_perms))
logger.info(" val split: " + str(val_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 = vsTVSum_Loader3_c3dd_segment.cDataset(
dataset_name=args.dataset,
split='train',
decode_ratio=args.decode_ratio,
train_val_perms=train_perms,
data_path=data_path)
max_input_len = train_dataset.max_input_len
maximum_outputs = int(args.decode_ratio * max_input_len)
val_dataset = vsTVSum_Loader3_c3dd_segment.cDataset(
dataset_name=args.dataset,
split='val',
decode_ratio=args.decode_ratio,
train_val_perms=val_perms,
data_path=data_path)
train_evaluator = Evaluator.Evaluator(dataset_name=args.dataset,
split='tr',
max_input_len=max_input_len,
maximum_outputs=maximum_outputs,
sum_budget=0.15,
train_val_perms=train_perms,
eval_metrics=args.eval_metrics,
data_path=data_path)
val_evaluator = Evaluator.Evaluator(dataset_name=args.dataset,
split='val',
max_input_len=max_input_len,
maximum_outputs=maximum_outputs,
sum_budget=0.15,
train_val_perms=val_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 = PointerNetwork(input_dim=args.input_dim,
embedding_dim=args.embedding_dim,
hidden_dim=args.hidden_dim,
max_decoding_len=maximum_outputs,
dropout=args.dropout,
n_enc_layers=2,
output_classes=1)
# 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)
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()
pointer_losses = AverageMeter()
rgs_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])
pointer_indices = Variable(sample_batch[1])
pointer_scores = Variable(sample_batch[2])
gt_valids = Variable(sample_batch[3])
# seq_labels = Variable(sample_batch[3])
if use_cuda:
feature_batch = feature_batch.cuda()
pointer_indices = pointer_indices.cuda()
pointer_scores = pointer_scores.cuda()
gt_positions = pointer_indices
gt_scores = pointer_scores
pointer_probs, pointer_positions, cls_scores, _ = model(
feature_batch)
pred_positions = pointer_positions
cls_scores = cls_scores.contiguous().squeeze(2)
# print(pointer_probs.size())
# print(gt_positions.size())
pointer_loss = F.cross_entropy(pointer_probs.permute(0, 2, 1),
gt_positions)
rgs_loss = F.mse_loss(cls_scores, gt_scores)
total_loss = alpha * pointer_loss + rgs_loss
model_optim.zero_grad()
total_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.)
model_optim.step()
pointer_losses.update(pointer_loss.data.item(),
feature_batch.size(0))
rgs_losses.update(rgs_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 pointer-loss:{:.4f}, \tregress-loss:{:.4f}\tcls-Accuracy:{:.4f}"
.format(epoch, model_optim.param_groups[0]['lr'], total_losses.avg,
pointer_losses.avg, rgs_losses.avg, Accuracy.avg))
optim_scheduler.step(total_losses.avg)
model.eval()
pointer_losses = AverageMeter()
rgs_losses = 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])
pointer_indices = Variable(sample_batch[1])
pointer_scores = Variable(sample_batch[2])
gt_valids = Variable(sample_batch[3])
# valid_indices = Variable(sample_batch[3])
if use_cuda:
feature_batch = feature_batch.cuda()
pointer_indices = pointer_indices.cuda()
pointer_scores = pointer_scores.cuda()
gt_positions = pointer_indices
gt_scores = pointer_scores
pointer_probs, pointer_positions, cls_scores, _ = model(
feature_batch)
pred_positions = pointer_positions
cls_scores = cls_scores.contiguous().squeeze(2)
pointer_loss = F.cross_entropy(pointer_probs.permute(0, 2, 1),
gt_positions)
rgs_loss = F.mse_loss(cls_scores, gt_scores)
pointer_losses.update(pointer_loss.data.item(),
feature_batch.size(0))
rgs_losses.update(rgs_loss.data.item(), feature_batch.size(0))
train_F1s = train_evaluator.Evaluate(model)
val_F1s = val_evaluator.Evaluate(model)
logger.info("Train -- Epoch :{:06d},\tF1s:{:.4f}".format(
epoch, train_F1s))
logger.info(
"Val -- Epoch :{:06d},\t pointer-loss:{:.4f}, \tregress-loss:{:.4f}, \tF1s{:.4f}"
.format(epoch, pointer_losses.avg, rgs_losses.avg, val_F1s))
if epoch % 1 == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'loss': total_losses.avg,
'pointer_loss': pointer_losses.avg,
'rgs_loss': rgs_losses.avg,
'val_F1s': val_F1s
}, (epoch + 1),
file_direcotry=save_directory)
