def main():
global args
args = (parser.parse_args())
ckpt_idx = args.fileid
proposal_save_file = 'Dev/NetModules/ActionLocalizationDevs/PropEval/baselines_results/inception-s4-EMD-gru-aug-{:04d}_thumos14_test.csv'.format(
ckpt_idx)
feature_directory = os.path.join(user_home_directory,
'datasets/THUMOS14/features/BNInception')
ground_truth_file = os.path.join(
user_home_directory,
'/home/zwei/Dev/NetModules/ActionLocalizationDevs/action_det_prep/thumos14_tag_test_proposal_list.csv'
)
ground_truth = pd.read_csv(ground_truth_file, sep=' ')
target_video_frms = ground_truth[['video-name',
'video-frames']].drop_duplicates().values
frm_nums = {}
for s_target_videofrms in target_video_frms:
frm_nums[s_target_videofrms[0]] = s_target_videofrms[1]
target_file_names = ground_truth['video-name'].unique()
feature_file_ext = 'npy'
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)
print("Number of Params\t{:d}".format(
sum([p.data.nelement() for p in model.parameters()])))
model = cuda_model.convertModel2Cuda(model,
gpu_id=args.gpu_id,
multiGpu=args.multiGpu)
model.eval()
if args.eval is not None:
# if os.path.isfile(args.resume):
ckpt_filename = os.path.join(
args.eval, '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=True)
train_iou = checkpoint['IoU']
print("=> loading checkpoint '{}', current iou: {:.04f}".format(
ckpt_filename, train_iou))
predict_results = {}
overlap = 0.6
seq_length = 90
sample_rate = [1, 2, 4]
for s_sample_rate in sample_rate:
for video_idx, s_target_filename in enumerate(target_file_names):
if not os.path.exists(
os.path.join(
feature_directory, '{:s}.{:s}'.format(
s_target_filename, feature_file_ext))):
print('{:s} Not found'.format(s_target_filename))
continue
s_feature_path = os.path.join(
feature_directory, '{:s}.{:s}'.format(s_target_filename,
feature_file_ext))
singlevideo_data = SingleVideoLoader(feature_path=s_feature_path,
seq_length=seq_length,
overlap=overlap,
sample_rate=[s_sample_rate])
n_video_len = singlevideo_data.n_features
n_video_clips = len(singlevideo_data.video_clips)
singlevideo_dataset = DataLoader(singlevideo_data,
batch_size=args.batch_size,
shuffle=False,
num_workers=4)
predict_proposals = []
for batch_idx, data in enumerate(singlevideo_dataset):
clip_feature = Variable(data[0], requires_grad=False)
clip_start_positions = Variable(data[1], requires_grad=False)
clip_end_positions = Variable(data[2], requires_grad=False)
if use_cuda:
clip_feature = clip_feature.cuda()
clip_start_positions = clip_start_positions.cuda()
clip_end_positions = clip_end_positions.cuda()
clip_start_positions = clip_start_positions.repeat(
1, args.net_outputs)
clip_end_positions = clip_end_positions.repeat(
1, args.net_outputs)
head_pointer_probs, head_positions, tail_pointer_probs, tail_positions, cls_scores, _ = model(
clip_feature)
cls_scores = F.softmax(cls_scores, dim=2)
head_positions, tail_positions = helper.reorder(
head_positions, tail_positions)
head_positions = (head_positions * s_sample_rate +
clip_start_positions)
tail_positions = (tail_positions * s_sample_rate +
clip_start_positions)
cls_scores = cls_scores[:, :, 1].contiguous().view(-1)
head_positions = head_positions.contiguous().view(-1)
tail_positions = tail_positions.contiguous().view(-1)
outputs = torch.stack([
head_positions.float(),
tail_positions.float(), cls_scores
],
dim=-1)
outputs = outputs.data.cpu().numpy()
for output_idx, s_output in enumerate(outputs):
if s_output[0] == s_output[1]:
s_output[0] -= s_sample_rate / 2
s_output[1] += s_sample_rate / 2
s_output[0] = max(0, s_output[0])
s_output[1] = min(n_video_len, s_output[1])
outputs[output_idx] = s_output
predict_proposals.append(outputs)
predict_proposals = np.concatenate(predict_proposals, axis=0)
predict_proposals, _ = PropUtils.non_maxima_supression(
predict_proposals, overlap=0.999)
# sorted_idx = np.argsort(predict_proposals[:,-1])[::-1]
# predict_proposals = predict_proposals[sorted_idx]
if s_target_filename in predict_results.keys():
predict_results[s_target_filename] = np.concatenate(
(predict_results[s_target_filename], predict_proposals),
axis=0)
else:
predict_results[s_target_filename] = predict_proposals
n_proposals = len(predict_proposals)
print(
"[{:d} | {:d}]{:s}\t {:d} Frames\t {:d} Clips\t{:d} Proposals @ rate:{:d}"
.format(video_idx, len(target_file_names), s_target_filename,
n_video_len, n_video_clips, n_proposals,
s_sample_rate))
data_frame = pkl_frame2dataframe(predict_results, frm_nums)
results = pd.DataFrame(
data_frame,
columns=['f-end', 'f-init', 'score', 'video-frames', 'video-name'])
results.to_csv(os.path.join(user_home_directory, proposal_save_file),
sep=' ',
index=False)
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=0.5)
print("Number of Params\t{:d}".format(
sum([p.data.nelement() for p in model.parameters()])))
save_directory = 'gru2heads_proposal_s4-3_cls_AUG_ckpt'
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 = THUMOST14(seq_length=args.seq_len,
overlap=0.9,
sample_rate=[1, 2, 4],
dataset_split='train',
rdDrop=True,
rdOffset=True)
val_dataset = THUMOST14(seq_length=args.seq_len,
overlap=0.9,
sample_rate=[1, 2, 4],
dataset_split='val',
rdDrop=True,
rdOffset=True)
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')
alpha = 0.1
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])
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, valid_indices, thres=0.5)
# 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))
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))
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))
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])
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, valid_indices, thres=0.5)
# 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))
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
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)
optim_scheduler.step(total_losses.avg)
def main():
global args
args = (parser.parse_args())
use_cuda = cuda_model.ifUseCuda(args.gpu_id, args.multiGpu)
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=0.5)
print("Number of Params\t{:d}".format(
sum([p.data.nelement() for p in model.parameters()])))
# save_directory = 'gru2heads_proposal_s4-2_ckpt'
if args.resume is not None:
ckpt_idx = 2
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=True)
args.start_epoch = checkpoint['epoch']
train_iou = checkpoint['IoU']
train_tloss = checkpoint['loss']
train_cls_loss = checkpoint['cls_loss']
train_loc_loss = checkpoint['loc_loss']
print(
"=> loading checkpoint '{}', total loss: {:.04f},\t cls_loss: {:.04f},\t loc_loss: {:.04f},"
" \tcurrent iou: {:.04f}".format(ckpt_filename, train_tloss,
train_cls_loss, train_loc_loss,
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=[1, 2, 4],
dataset_split='val')
train_dataloader = DataLoader(train_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')
alpha = 0.1
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()
#Update here!
model.eval()
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])
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_match.Assign_Batch(
gt_positions, pred_positions, valid_indices, thres=0.5)
if valid_indices.byte().any() > 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(gt_positions[view_idx, :n_valid, :])
print("Pred")
print(pred_positions[view_idx])
_, head_sort = head_pointer_probs[view_idx, 0, :].sort()
_, tail_sort = tail_pointer_probs[view_idx, 0, :].sort()
print("END of {:d}".format(i_batch))
#
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)
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])
# start_indices = start_indices.contiguous().view(-1)
# end_indices = end_indices.contiguous().view(-1)
# with case instances....
prediction_head_loss = F.cross_entropy((head_pointer_probs),
assigned_head_positions,
reduce=False)
prediction_head_loss = torch.mean(prediction_head_loss *
assigned_scores.float())
prediction_tail_loss = F.cross_entropy((tail_pointer_probs),
assigned_tail_positions,
reduce=False)
prediction_tail_loss = torch.mean(prediction_tail_loss *
assigned_scores.float())
total_loss = alpha * (prediction_head_loss +
prediction_tail_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))
loc_losses.update(
prediction_head_loss.data.item() +
prediction_tail_loss.data.item(), feature_batch.size(0))
total_losses.update(total_loss.data.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))
break