def load_part_model(self, action_model_path=None, rnn_path=None):
# load action net
if action_model_path != None:
act_data = torch.load(action_model_path)
# act_data = torch.load('./action_net_model.pwf')
## to remove module
new_state_dict = OrderedDict()
for k, v in act_data.items():
# if k.find('module') != -1 :
name = k[7:] # remove `module.`
new_state_dict[name] = v
act_net = ACT_net(self.classes, self.sample_duration)
act_net.create_architecture()
act_net.load_state_dict(new_state_dict)
self.act_net = act_net
else:
self.act_net = ACT_net(self.classes, self.sample_duration)
self.act_net.create_architecture()
# load lstm
if rnn_path != None:
act_rnn = Act_RNN(self.p_feat_size, int(self.p_feat_size / 2),
self.n_classes)
act_rnn_data = torch.load(rnn_path)
act_rnn.load(act_rnn_data)
self.act_rnn = act_rnn
else:
self.act_rnn = Act_RNN(self.p_feat_size, int(self.p_feat_size),
self.n_classes)
spatial_transform=spatial_transform,
temporal_transform=temporal_transform,
bboxes_file=boxes_file,
split_txt_path=split_txt_path,
mode='train',
classes_idx=cls2idx)
train_data_loader = torch.utils.data.DataLoader(train_data,
batch_size=batch_size,
shuffle=True,
num_workers=2,
pin_memory=True)
# Init action_net
act_model = ACT_net(actions, sample_duration, device=device)
act_model.create_architecture(model_path=model_path)
if torch.cuda.device_count() > 1:
print('Using {} GPUs!'.format(torch.cuda.device_count()))
act_model = nn.DataParallel(act_model)
act_model.to(device)
lr = 0.1
lr_decay_step = 10
lr_decay_gamma = 0.1
params = []
for key, value in dict(act_model.named_parameters()).items():
if value.requires_grad:
### get videos id
spatial_transform = Compose([
Scale(sample_size), # [Resize(sample_size),
ToTensor(),
Normalize(mean, [1, 1, 1])
])
temporal_transform = LoopPadding(sample_duration)
n_classes = len(actions)
# Init action_net
model = ACT_net(actions, sample_duration)
model.create_architecture()
model = nn.DataParallel(model)
model.to(device)
model_data = torch.load('./action_net_model_both_without_avg.pwf')
# model.load_state_dict(model_data)
model.eval()
data = Video_Dataset_small_clip(dataset_frames,
frames_dur=sample_duration,
spatial_transform=spatial_transform,
temporal_transform=temporal_transform,
bboxes_file=boxes_file,
split_txt_path=split_txt_path,
mode='train',
# torch.save(act_model.module.act_rpn.state_dict(), "region_net_.pwf")
# torch.save(act_model.state_dict(), "all_model.pwf")
# #####################################################
# # Part 1-2 - train TPN - only reg #
# #####################################################
print(' --------------------------------------------------')
print('| Part 1-2 - train TPN - only reg |')
print(' --------------------------------------------------')
### Only for second part training
# Init action_net
act_model = ACT_net(actions, sample_duration)
act_model.create_architecture()
if torch.cuda.device_count() > 1:
print('Using {} GPUs!'.format(torch.cuda.device_count()))
act_model = nn.DataParallel(act_model)
act_model.to(device)
n_devs = torch.cuda.device_count()
# model_data = torch.load('./action_net_model_part1_1.pwf')
model_data = torch.load('./region_net_.pwf')
# act_model.load_state_dict(model_data)
act_model.module.act_rpn.load_state_dict(model_data)
actions = dataset_cfg.dataset.classes
cls2idx = {actions[i]: i for i in range(0, len(actions))}
### get videos id
vid2idx, vid_names = get_vid_dict(dataset_frames)
spatial_transform = Compose([
Scale(sample_size), # [Resize(sample_size),
ToTensor(),
Normalize(mean, std)
])
temporal_transform = LoopPadding(sample_duration)
# Init action_net
model = ACT_net(actions, sample_duration)
model.create_architecture(model_path)
model = nn.DataParallel(model)
model.to(device)
data = Video_Dataset_small_clip(dataset_frames,
frames_dur=sample_duration,
spatial_transform=spatial_transform,
temporal_transform=temporal_transform,
bboxes_file=boxes_file,
split_txt_path=split_txt_path,
mode='train',
classes_idx=cls2idx)
# data_loader = torch.utils.data.DataLoader(data, batch_size=batch_size,
# shuffle=False, num_workers=n_threads, pin_memory=True)
class Model(nn.Module):
"""
action localizatio network which contains:
-ACT_net : a network for proposing action tubes for 16 frames
-TCN net : a dilation network which classifies the input tubes
"""
def __init__(self, actions, sample_duration, sample_size):
super(Model, self).__init__()
self.classes = actions
self.n_classes = len(actions)
# self.act_net = ACT_net(actions,sample_duration)
## general options
self.sample_duration = sample_duration
self.sample_size = sample_size
self.step = int(self.sample_duration / 2)
self.p_feat_size = 64 # 128 # 256 # 512
# For connection
self.max_num_tubes = conf.MAX_NUMBER_TUBES
self.connection_thresh = conf.CONNECTION_THRESH
self.update_thresh_step = conf.UPDATE_THRESH
self.calc = Calculator(self.max_num_tubes, self.update_thresh_step,
self.connection_thresh)
def forward(self, n_devs, dataset_folder, vid_names, clips, vid_id, boxes,
mode, cls2idx, num_actions, num_frames, h_, w_):
'''
TODO describe procedure
'''
# print('boxes.shape :',boxes.shape)
## define a dataloader for the whole video
# print('----------Inside----------')
# print('num_frames :',num_frames)
# print('clips.shape :',clips.shape)
clips = clips.squeeze(0)
clips = clips[:num_frames]
print('num_frames :', num_frames)
print('clips.shape :', clips.shape)
if self.training:
boxes = boxes.squeeze(0).permute(1, 0, 2).cpu()
boxes = boxes[:num_frames, :num_actions]
batch_size = 2 #
# batch_size = 16 #
num_images = 1
rois_per_image = int(cfg.TRAIN.BATCH_SIZE /
num_images) if self.training else 150
data = single_video(dataset_folder,
h_,
w_,
vid_names,
vid_id,
frames_dur=self.sample_duration,
sample_size=self.sample_size,
classes_idx=cls2idx,
n_frames=num_frames)
data_loader = torch.utils.data.DataLoader(
data,
batch_size=batch_size,
pin_memory=False, # num_workers=num_workers, pin_memory=True,
# shuffle=False, num_workers=8)
shuffle=False)
n_clips = data.__len__()
features = torch.zeros(n_clips, rois_per_image, self.p_feat_size,
self.sample_duration)
p_tubes = torch.zeros(n_clips, rois_per_image, self.sample_duration *
4) # all the proposed tube-rois
actioness_score = torch.zeros(n_clips, rois_per_image)
overlaps_scores = torch.zeros(n_clips, rois_per_image, rois_per_image)
f_tubes = []
if self.training:
f_gt_tubes = torch.zeros(n_clips, num_actions,
self.sample_duration * 4) # gt_tubes
tubes_labels = torch.zeros(n_clips, rois_per_image) # tubes rois
loops = int(np.ceil(n_clips / batch_size))
labels = torch.zeros(num_actions)
for i in range(num_actions):
idx = boxes[:, i, 4].nonzero().view(-1)
labels[i] = boxes[i, idx[0], 4]
for step, dt in enumerate(data_loader):
# if step == 1:
# break
print('\tstep :', step)
frame_indices, im_info, start_fr = dt
clips_ = clips[frame_indices].cuda()
if self.training:
boxes_ = boxes[frame_indices].cuda()
box_ = boxes_.permute(0, 2, 1,
3).float().contiguous()[:, :, :, :-1]
else:
box_ = None
im_info = im_info.cuda()
start_fr = start_fr.cuda()
with torch.no_grad():
tubes, pooled_feat, \
rpn_loss_cls, rpn_loss_bbox, \
_,_, rois_label, \
sgl_rois_bbox_pred, sgl_rois_bbox_loss = self.act_net(clips_.permute(0,2,1,3,4),
im_info,
None,
box_,
start_fr)
pooled_feat = pooled_feat.view(-1, rois_per_image,
self.p_feat_size,
self.sample_duration)
indexes_ = (torch.arange(0, tubes.size(0)) *
int(self.sample_duration / 2) +
start_fr[0].cpu()).unsqueeze(1)
indexes_ = indexes_.expand(tubes.size(0),
tubes.size(1)).type_as(tubes)
idx_s = step * batch_size
idx_e = step * batch_size + batch_size
features[idx_s:idx_e] = pooled_feat
p_tubes[idx_s:idx_e, ] = tubes[:, :, 1:-1]
actioness_score[idx_s:idx_e] = tubes[:, :, -1]
if self.training:
box = boxes_.permute(0, 2, 1, 3).contiguous()[:, :, :, :-2]
box = box.contiguous().view(box.size(0), box.size(1), -1)
f_gt_tubes[idx_s:idx_e] = box
tubes_labels[idx_s:idx_e] = rois_label.squeeze(-1).type_as(
tubes_labels)
########################################################
# Calculate overlaps and connections #
########################################################
overlaps_scores = torch.zeros(n_clips, rois_per_image,
rois_per_image).type_as(overlaps_scores)
for i in range(n_clips - 1):
overlaps_scores[i] = tube_overlaps(
p_tubes[i, :, int(self.sample_duration * 4 / 2):],
p_tubes[i + 1, :, :int(self.sample_duration * 4 / 2)])
if n_clips > 1:
final_scores, final_poss = self.calc(
overlaps_scores.cuda(), actioness_score.cuda(),
torch.Tensor([n_clips]), torch.Tensor([rois_per_image]))
else:
offset = torch.arange(rois_per_image).float()
final_poss = torch.stack([torch.zeros((rois_per_image)), offset],
dim=1).unsqueeze(1).long()
## Now connect the tubes
final_tubes = torch.zeros(final_poss.size(0), num_frames, 4)
f_tubes = []
for i in range(final_poss.size(0)):
tub = []
for j in range(final_poss.size(1)):
curr_ = final_poss[i, j]
start_fr = curr_[0] * int(self.sample_duration / 2)
end_fr = min((curr_[0] * int(self.sample_duration / 2) +
self.sample_duration).type_as(num_frames),
num_frames).type_as(start_fr)
if curr_[0] == -1:
break
curr_frames = p_tubes[curr_[0], curr_[1]]
tub.append((curr_[0].item(), curr_[1].item()))
## TODO change with avg
final_tubes[i, start_fr:end_fr] = torch.max(
curr_frames.view(-1, 4).contiguous()[:(end_fr -
start_fr).long()],
final_tubes[i, start_fr:end_fr].type_as(curr_frames))
f_tubes.append(tub)
###################################################
# Choose gth Tubes for RCNN\TCN #
###################################################
if self.training:
# # get gt tubes and feats
## calculate overlaps
boxes_ = boxes.permute(1, 0, 2).contiguous()
boxes_ = boxes_[:, :, :4].contiguous().view(num_actions, -1)
overlaps = tube_overlaps(final_tubes.view(-1, num_frames * 4),
boxes_.type_as(final_tubes))
max_overlaps, _ = torch.max(overlaps, 1)
max_overlaps = max_overlaps.clamp_(min=0)
## TODO change numbers
bg_tubes_indices = max_overlaps.lt(0.3).nonzero()
bg_tubes_indices_picked = (torch.rand(5) *
bg_tubes_indices.size(0)).long()
bg_tubes_list = [
f_tubes[i] for i in bg_tubes_indices[bg_tubes_indices_picked]
]
bg_labels = torch.zeros(len(bg_tubes_list))
gt_tubes_list = [[] for i in range(num_actions)]
for i in range(n_clips):
overlaps = tube_overlaps(p_tubes[i], f_gt_tubes[i])
max_overlaps, argmax_overlaps = torch.max(overlaps, 0)
for j in range(num_actions):
if max_overlaps[j] == 1.0:
gt_tubes_list[j].append((i, j))
## concate fb, bg tubes
f_tubes = gt_tubes_list + bg_tubes_list
target_lbl = torch.cat([labels, bg_labels], dim=0)
##############################################
if len(f_tubes) == 0:
print('------------------')
print(' empty tube ')
return torch.Tensor([]).cuda(), torch.Tensor([]).cuda(), None
max_seq = reduce(lambda x, y: y if len(y) > len(x) else x, f_tubes)
max_length = len(max_seq)
## calculate input rois
## f_feats.shape : [#f_tubes, max_length, 512]
final_video_tubes = torch.zeros(len(f_tubes), 6).cuda()
prob_out = torch.zeros(len(f_tubes), self.n_classes).cuda()
for i in range(len(f_tubes)):
seq = f_tubes[i]
tmp_tube = torch.Tensor(len(seq), 6)
feats = torch.Tensor(len(seq), self.p_feat_size)
for j in range(len(seq)):
feats[j] = features[seq[j][0], seq[j][1]].mean(1)
tmp_tube[j] = p_tubes[seq[j]][1:7]
prob_out[i] = self.act_rnn(feats.cuda())
if prob_out[i, 0] != prob_out[i, 0]:
print('tmp_tube :',tmp_tube, ' prob_out :', prob_out ,' feats :',feats.cpu().numpy(), ' numpy(), feats.shape :,', feats.shape ,' target_lbl :',target_lbl, \
' \ntmp_tube :',tmp_tube, )
exit(-1)
# ##########################################
# # Time for Linear Loss #
# ##########################################
cls_loss = torch.Tensor([0]).cuda()
final_tubes = final_tubes.type_as(final_poss)
# # classification probability
if self.training:
cls_loss = F.cross_entropy(prob_out.cpu(),
target_lbl.long()).cuda()
if self.training:
return final_tubes, prob_out, cls_loss,
else:
return final_tubes, prob_out, None
def deactivate_action_net_grad(self):
for p in self.act_net.parameters():
p.requires_grad = False
# self.act_net.eval()
# for key, value in dict(self.named_parameters()).items():
# print(key, value.requires_grad)
def load_part_model(self, action_model_path=None, rnn_path=None):
# load action net
if action_model_path != None:
act_data = torch.load(action_model_path)
# act_data = torch.load('./action_net_model.pwf')
## to remove module
new_state_dict = OrderedDict()
for k, v in act_data.items():
# if k.find('module') != -1 :
name = k[7:] # remove `module.`
new_state_dict[name] = v
act_net = ACT_net(self.classes, self.sample_duration)
act_net.create_architecture()
act_net.load_state_dict(new_state_dict)
self.act_net = act_net
else:
self.act_net = ACT_net(self.classes, self.sample_duration)
self.act_net.create_architecture()
# load lstm
if rnn_path != None:
act_rnn = Act_RNN(self.p_feat_size, int(self.p_feat_size / 2),
self.n_classes)
act_rnn_data = torch.load(rnn_path)
act_rnn.load(act_rnn_data)
self.act_rnn = act_rnn
else:
self.act_rnn = Act_RNN(self.p_feat_size, int(self.p_feat_size),
self.n_classes)
def load_part_model(self,
resnet_path=None,
action_model_path=None,
rnn_path=None):
# load action net
if action_model_path != None:
act_data = torch.load(action_model_path)
# act_data = torch.load('./action_net_model.pwf')
## to remove module
new_state_dict = OrderedDict()
for k, v in act_data.items():
# if k.find('module') != -1 :
name = k[7:] # remove `module.`
new_state_dict[name] = v
act_net = ACT_net(self.classes, self.sample_duration)
if resnet_path is not None:
act_net.create_architecture(model_path=resnet_path)
else:
act_net.create_architecture()
act_net.load_state_dict(new_state_dict)
self.act_net = act_net
else:
self.act_net = ACT_net(self.classes, self.sample_duration)
if resnet_path is not None:
self.act_net.create_architecture(model_path=resnet_path)
else:
self.act_net.create_architecture()
# load lstm
if rnn_path != None:
# act_rnn = Act_RNN(self.p_feat_size,int(self.p_feat_size/2),self.n_classes)
# act_rnn_data = torch.load(rnn_path)
# act_rnn.load_state_dict(act_rnn_data)
act_rnn = nn.Sequential(
# nn.Linear(64*self.sample_duration, 256),
# nn.ReLU(True),
# nn.Dropout(0.8),
# nn.Linear(256,self.n_classes),
nn.Linear(64 * self.sample_duration, self.n_classes),
# nn.ReLU(True),
# nn.Dropout(0.8),
# nn.Linear(256,self.n_classes),
)
act_rnn_data = torch.load(rnn_path)
act_rnn.load_state_dict(act_rnn_data)
self.act_rnn = act_rnn
else:
# self.act_rnn =Act_RNN(self.p_feat_size,int(self.p_feat_size/2),self.n_classes)
self.act_rnn = nn.Sequential(
# nn.Linear(64*self.sample_duration, 256),
# nn.ReLU(True),
# nn.Dropout(0.8),
# nn.Linear(256,self.n_classes),
nn.Linear(64 * self.sample_duration, self.n_classes),
# nn.ReLU(True),
# nn.Dropout(0.8),
# nn.Linear(256,self.n_classes),
)
for m in self.act_rnn.modules():
if m == nn.Linear:
m.weight.data.normal_().fmod_(2).mul_(stddev).add_(
mean) # not a perfect approximation
class Model(nn.Module):
"""
action localizatio network which contains:
-ACT_net : a network for proposing action tubes for 16 frames
-TCN net : a dilation network which classifies the input tubes
"""
def __init__(self, actions, sample_duration, sample_size):
super(Model, self).__init__()
self.classes = actions
self.n_classes = len(actions)
# self.act_net = ACT_net(actions,sample_duration)
## general options
self.sample_duration = sample_duration
self.sample_size = sample_size
self.step = int(self.sample_duration / 2)
self.p_feat_size = 64 # 128 # 256 # 512
# For connection
self.max_num_tubes = conf.MAX_NUMBER_TUBES
self.connection_thresh = conf.CONNECTION_THRESH
self.update_thresh = conf.UPDATE_THRESH
self.calc = Calculator(self.max_num_tubes, self.update_thresh,
self.connection_thresh)
def forward(self, n_devs, dataset_folder, vid_names, clips, vid_id, boxes,
mode, cls2idx, num_actions, num_frames, h_, w_):
'''
TODO describe procedure
'''
# print('boxes.shape :',boxes.shape)
## define a dataloader for the whole video
# print('----------Inside----------')
# print('num_frames :',num_frames)
# print('clips.shape :',clips.shape)
clips = clips.squeeze(0)
ret_n_frames = clips.size(0)
clips = clips[:num_frames]
# print('num_frames :',num_frames)
# print('clips.shape :',clips.shape)
# exit(-1)
if self.training:
boxes = boxes.squeeze(0).permute(1, 0, 2).cpu()
boxes = boxes[:num_frames, :num_actions].clamp_(min=0)
batch_size = 4 #
# batch_size = 2 #
# batch_size = 16 #
num_images = 1
rois_per_image = int(conf.TRAIN.BATCH_SIZE /
num_images) if self.training else 150
data = single_video(dataset_folder,
h_,
w_,
vid_names,
vid_id,
frames_dur=self.sample_duration,
sample_size=self.sample_size,
classes_idx=cls2idx,
n_frames=num_frames)
data_loader = torch.utils.data.DataLoader(
data,
batch_size=batch_size,
pin_memory=False, # num_workers=num_workers, pin_memory=True,
# shuffle=False, num_workers=8)
shuffle=False)
n_clips = data.__len__()
features = torch.zeros(n_clips, rois_per_image, self.p_feat_size,
self.sample_duration).type_as(clips)
p_tubes = torch.zeros(n_clips, rois_per_image, self.sample_duration *
4).type_as(clips) # all the proposed tube-rois
actioness_score = torch.zeros(n_clips, rois_per_image).type_as(clips)
overlaps_scores = torch.zeros(n_clips, rois_per_image,
rois_per_image).type_as(clips)
f_tubes = []
# #
# overlaps_scores = torch.zeros(n_clips, rois_per_image, rois_per_image).type_as(overlaps_scores)
if self.training:
f_gt_tubes = torch.zeros(n_clips, num_actions,
self.sample_duration * 4) # gt_tubes
tubes_labels = torch.zeros(n_clips, rois_per_image) # tubes rois
loops = int(np.ceil(n_clips / batch_size))
labels = torch.zeros(num_actions)
for i in range(num_actions):
idx = boxes[:, i, 4].nonzero().view(-1)
labels[i] = boxes[idx[0], i, 4]
## Init connect thresh
self.calc.thresh = self.connection_thresh
for step, dt in enumerate(data_loader):
frame_indices, im_info, start_fr = dt
clips_ = clips[frame_indices].cuda()
if self.training:
boxes_ = boxes[frame_indices].cuda()
box_ = boxes_.permute(0, 2, 1,
3).float().contiguous()[:, :, :, :-1]
else:
box_ = None
im_info = im_info.cuda()
start_fr = start_fr.cuda()
with torch.no_grad():
tubes, pooled_feat, \
rpn_loss_cls, rpn_loss_bbox, \
_,_, rois_label, \
sgl_rois_bbox_pred, sgl_rois_bbox_loss = self.act_net(clips_.permute(0,2,1,3,4),
im_info,
None,
box_,
start_fr)
pooled_feat = pooled_feat.mean(-1).mean(-1)
pooled_feat = pooled_feat.view(-1, rois_per_image,
self.p_feat_size,
self.sample_duration)
# regression
n_tubes = len(tubes)
if not self.training:
tubes = tubes.view(-1, self.sample_duration * 4 + 2)
tubes[:,1:-1] = tube_transform_inv(tubes[:,1:-1],\
sgl_rois_bbox_pred.view(-1,self.sample_duration*4),(1.0,1.0,1.0,1.0))
tubes = tubes.view(n_tubes, rois_per_image,
self.sample_duration * 4 + 2)
tubes[:, :, 1:-1] = clip_boxes(tubes[:, :, 1:-1], im_info,
tubes.size(0))
indexes_ = (torch.arange(0, tubes.size(0)) *
int(self.sample_duration / 2) +
start_fr[0].cpu()).unsqueeze(1)
indexes_ = indexes_.expand(tubes.size(0),
tubes.size(1)).type_as(tubes)
idx_s = step * batch_size
idx_e = min(step * batch_size + batch_size, n_clips)
features[idx_s:idx_e] = pooled_feat
p_tubes[idx_s:idx_e, ] = tubes[:, :, 1:-1]
actioness_score[idx_s:idx_e] = tubes[:, :, -1]
if self.training:
box = boxes_.permute(0, 2, 1, 3).contiguous()[:, :, :, :-2]
box = box.contiguous().view(box.size(0), box.size(1), -1)
f_gt_tubes[idx_s:idx_e] = box
# connection algo
for i in range(idx_s, idx_e):
if i == 0:
# Init tensors for connecting
offset = torch.arange(0, rois_per_image).int().cuda()
ones_t = torch.ones(rois_per_image).int().cuda()
zeros_t = torch.zeros(rois_per_image, n_clips,
2).int().cuda() - 1
pos = torch.zeros(rois_per_image, n_clips,
2).int().cuda() - 1 # initial pos
pos[:, 0, 0] = 0
pos[:, 0, 1] = offset.contiguous(
) # contains the current tubes to be connected
pos_indices = torch.zeros(rois_per_image).int().cuda(
) # contains the pos of the last element of the previous tensor
actioness_scr = actioness_score[0].float().cuda(
) # actioness sum of active tubes
overlaps_scr = torch.zeros(rois_per_image).float().cuda(
) # overlaps sum of active tubes
final_scores = torch.Tensor().float().cuda(
) # final scores
final_poss = torch.Tensor().int().cuda() # final tubes
continue
overlaps_ = tube_overlaps(
p_tubes[i - 1, :,
int(self.sample_duration * 4 / 2):],
p_tubes[i, :, :int(self.sample_duration * 4 /
2)]).type_as(p_tubes)
pos, pos_indices, \
f_scores, actioness_scr, \
overlaps_scr = self.calc(torch.Tensor([n_clips]),torch.Tensor([rois_per_image]),torch.Tensor([pos.size(0)]),
pos, pos_indices, actioness_scr, overlaps_scr,
overlaps_, actioness_score[i], torch.Tensor([i]))
if pos.size(0) > self.update_thresh:
final_scores, final_poss, pos , pos_indices, \
actioness_scr, overlaps_scr, f_scores = self.calc.update_scores(final_scores,final_poss, f_scores, pos, pos_indices, actioness_scr, overlaps_scr)
if f_scores.dim() == 0:
f_scores = f_scores.unsqueeze(0)
pos = pos.unsqueeze(0)
pos_indices = pos_indices.unsqueeze(0)
actioness_scr = actioness_scr.unsqueeze(0)
overlaps_scr = overlaps_scr.unsqueeze(0)
if final_scores.dim() == 0:
final_scores = final_scores.unsqueeze(0)
final_poss = final_poss.unsqueeze(0)
try:
final_scores = torch.cat((final_scores, f_scores))
except:
print('final_scores :', final_scores)
print('final_scores.shape :', final_scores.shape)
print('final_scores.dim() :', final_scores.dim())
print('f_scores :', f_scores)
print('f_scores.shape :', f_scores.shape)
print('f_scores.dim() :', f_scores.dim())
exit(-1)
try:
final_poss = torch.cat((final_poss, pos))
except:
print('final_poss :', final_poss)
print('final_poss.shape :', final_poss.shape)
print('final_poss.dim() :', final_poss.dim())
print('pos :', pos)
print('pos.shape :', pos.shape)
print('pos.dim() :', pos.dim())
exit(-1)
# add new tubes
pos = torch.cat((pos, zeros_t))
pos[-rois_per_image:, 0, 0] = ones_t * i
pos[-rois_per_image:, 0, 1] = offset
pos_indices = torch.cat(
(pos_indices, torch.zeros(
(rois_per_image)).type_as(pos_indices)))
actioness_scr = torch.cat((actioness_scr, actioness_score[i]))
overlaps_scr = torch.cat(
(overlaps_scr, torch.zeros(
(rois_per_image)).type_as(overlaps_scr)))
## add only last layers
## TODO check again
indices = actioness_score[-1].ge(self.calc.thresh).nonzero().view(-1)
if indices.nelement() > 0:
zeros_t[:, 0, 0] = idx_e - 1
zeros_t[:, 0, 1] = offset
final_poss = torch.cat([final_poss, zeros_t[indices]])
if pos.size(0) > self.update_thresh:
print('Updating thresh...', final_scores.shape, final_poss.shape,
pos.shape, f_scores.shape, pos_indices.shape)
final_scores, final_poss, pos , pos_indices, \
actioness_scr, overlaps_scr, f_scores = self.calc.update_scores(final_scores,final_poss, f_scores, pos, pos_indices, actioness_scr, overlaps_scr)
print('Updating thresh...', final_scores.shape, final_poss.shape,
pos.shape, f_scores.shape, pos_indices.shape)
final_tubes = torch.zeros(final_poss.size(0), num_frames, 4)
f_tubes = []
for i in range(final_poss.size(0)):
tub = []
for j in range(final_poss.size(1)):
curr_ = final_poss[i, j]
start_fr = curr_[0] * int(self.sample_duration / 2)
end_fr = min((curr_[0] * int(self.sample_duration / 2) +
self.sample_duration).type_as(num_frames),
num_frames).type_as(start_fr)
if curr_[0] == -1:
break
curr_frames = p_tubes[curr_[0], curr_[1]]
tub.append((curr_[0].item(), curr_[1].item()))
## TODO change with avg
final_tubes[i, start_fr:end_fr] = torch.max(
curr_frames.view(-1, 4).contiguous()[:(end_fr -
start_fr).long()],
final_tubes[i, start_fr:end_fr].type_as(curr_frames))
f_tubes.append(tub)
###################################################
# Choose gth Tubes for RCNN\TCN #
###################################################
if self.training:
# # get gt tubes and feats
## calculate overlaps
boxes_ = boxes.permute(1, 0, 2).contiguous()
boxes_ = boxes_[:, :, :4].contiguous().view(num_actions, -1)
if final_tubes.nelement() == 0:
print('problem final_tubes ...')
print('boxes :', boxes.cpu().numpy())
print('boxes_ :', boxes_)
print('boxes_.shape :', boxes_.shape)
print('final_tubes :', final_tubes)
print('self.calc.thresh:', self.calc.thresh)
print('final_scores :', final_scores.shape)
print('final_pos.shape :', final_poss.shape)
if final_tubes.nelement() > 0:
overlaps = tube_overlaps(final_tubes.view(-1, num_frames * 4),
boxes_.type_as(final_tubes))
max_overlaps, _ = torch.max(overlaps, 1)
max_overlaps = max_overlaps.clamp_(min=0)
## TODO change numbers
bg_tubes_indices = max_overlaps.lt(0.3).nonzero()
if bg_tubes_indices.nelement() > 0:
bg_tubes_indices_picked = (
torch.rand(9) * bg_tubes_indices.size(0)).long()
bg_tubes_list = [
f_tubes[i]
for i in bg_tubes_indices[bg_tubes_indices_picked]
]
bg_labels = torch.zeros(len(bg_tubes_list))
else:
bg_tubes_list = []
bg_labels = torch.Tensor([])
else:
bg_tubes_list = []
bg_labels = torch.Tensor([])
gt_tubes_list = [[] for i in range(num_actions)]
# print('n_clips :',n_clips)
for i in range(n_clips):
# print('i :',i)
# print('p_tubes.shape :',p_tubes.shape)
# print('f_gt_tubes.shape :',f_gt_tubes.shape)
# print('p_tubes.shape :',p_tubes[i])
# print('f_gt_tubes.shape :',f_gt_tubes[i])
overlaps = tube_overlaps(p_tubes[i],
f_gt_tubes[i].type_as(p_tubes))
# print('overlaps :',overlaps)
max_overlaps, argmax_overlaps = torch.max(overlaps, 0)
for j in range(num_actions):
if max_overlaps[j] == 1.0:
gt_tubes_list[j].append((i, j))
gt_tubes_list = [i for i in gt_tubes_list if i != []]
if len(gt_tubes_list) != num_actions:
print('len(gt_tubes_list :', len(gt_tubes_list))
print('num_actions :', num_actions)
print('boxes.cpu().numpy() :', boxes.cpu().numpy())
# print('gt_tubes_list :',gt_tubes_list)
## concate fb, bg tubes
if gt_tubes_list == [[]]:
print('overlaps :', overlaps)
print('max_overlaps :', max_overlaps)
print('p_tubes :', p_tubes)
print('f_gt_tubes :', f_gt_tubes)
exit(-1)
if bg_tubes_list != []:
f_tubes = gt_tubes_list + bg_tubes_list
target_lbl = torch.cat([labels, bg_labels], dim=0)
else:
f_tubes = gt_tubes_list
target_lbl = labels
# print('num_frames :',num_frames)
# print('gt_tubes_list :',gt_tubes_list, ' labels :',labels)
# print('f_tubes :',f_tubes, ' target_lbl :',target_lbl)
##############################################
if len(f_tubes) == 0:
print('------------------')
print(' empty tube ')
print(' vid_id :', vid_id)
print('self.calc.thresh :', self.calc.thresh)
return torch.Tensor([]).cuda(), torch.Tensor([]).cuda(), None
max_seq = reduce(lambda x, y: y if len(y) > len(x) else x, f_tubes)
max_length = len(max_seq)
## calculate input rois
## f_feats.shape : [#f_tubes, max_length, 512]
# final_video_tubes = torch.zeros(len(f_tubes),6).cuda()
prob_out = torch.zeros(len(f_tubes), self.n_classes).cuda()
# final_feats = []
f_feats = torch.zeros(len(f_tubes), n_clips, 64,
self.sample_duration).type_as(features) - 1
f_feats_len = torch.zeros(len(f_tubes)).type_as(features) - 1
for i in range(len(f_tubes)):
seq = f_tubes[i]
# tmp_tube = torch.Tensor(len(seq),6)
# feats = torch.Tensor(len(seq),self.p_feat_size)
feats = torch.Tensor(len(seq), self.p_feat_size,
self.sample_duration)
for j in range(len(seq)):
# feats[j] = features[seq[j][0],seq[j][1]].mean(1)
feats[j] = features[seq[j][0], seq[j][1]]
# tmp_tube[j] = p_tubes[seq[j]][1:7]
f_feats_len[i] = len(seq)
f_feats[i, :len(seq)] = feats
prob_out[i] = self.act_rnn(
feats.mean(0).view(1, -1).contiguous().cuda())
# # feats = torch.mean(feats, dim=0)
# if mode == 'extract':
# final_feats.append(feats)
# try:
# prob_out[i] = self.act_rnn(feats.view(-1).cuda())
# except Exception as e:
# print('feats.shape :',feats.shape)
# print('seq :',seq)
# for i in range(len(f_tubes)):
# print('seq[i] :',f_tubes[i])
# print('e :',e)
# exit(-1)
# if prob_out[i,0] != prob_out[i,0]:
# print(' prob_out :', prob_out ,' feats :',feats.cpu().numpy(), ' numpy(), feats.shape :,', feats.shape ,' target_lbl :',target_lbl, \
# ' \ntmp_tube :',tmp_tube, )
# exit(-1)
if mode == 'extract':
# now we use mean so we can have a tensor containing all features
# final_tubes = final_tubes.cuda()
target_lbl = target_lbl.cuda()
# max_length = torch.Tensor([max_length]).cuda()
return f_feats, target_lbl, f_feats_len
# ##########################################
# # Time for Linear Loss #
# ##########################################
cls_loss = torch.Tensor([0]).cuda()
final_tubes = final_tubes.type_as(final_poss)
# # classification probability
if self.training:
cls_loss = F.cross_entropy(prob_out.cpu(),
target_lbl.long()).cuda()
if self.training:
return None, prob_out, cls_loss,
else:
# init padding tubes because of multi-GPU system
if final_tubes.size(0) > conf.UPDATE_THRESH:
_, indices = torch.sort(final_scores)
final_tubes = final_tubes[
indices[:conf.UPDATE_THRESH]].contiguous()
prob_out = prob_out[indices[:conf.UPDATE_THRESH]].contiguous()
max_prob_out, _ = torch.max(prob_out, 1)
f_tubes = torch.cat([
final_tubes.view(-1, num_frames * 4),
max_prob_out.view(-1, 1).type_as(final_tubes)
],
dim=1)
keep = torch.Tensor(py_cpu_nms_tubes(f_tubes.float(), 0.5)).long()
final_tubes = final_tubes[keep]
prob_out = prob_out[keep]
ret_tubes = torch.zeros(1, conf.UPDATE_THRESH, ret_n_frames,
4).type_as(final_tubes).float() - 1
ret_prob_out = torch.zeros(
1, conf.UPDATE_THRESH,
self.n_classes).type_as(final_tubes).float() - 1
ret_tubes[0, :final_tubes.size(0), :num_frames] = final_tubes
ret_prob_out[0, :final_tubes.size(0)] = prob_out
return ret_tubes, ret_prob_out, torch.Tensor([final_tubes.size(0)
]).cuda()
# return final_tubes, prob_out, None
def deactivate_action_net_grad(self):
for p in self.act_net.parameters():
p.requires_grad = False
# self.act_net.eval()
# for key, value in dict(self.named_parameters()).items():
# print(key, value.requires_grad)
def load_part_model(self,
resnet_path=None,
action_model_path=None,
rnn_path=None):
# load action net
if action_model_path != None:
act_data = torch.load(action_model_path)
# act_data = torch.load('./action_net_model.pwf')
## to remove module
new_state_dict = OrderedDict()
for k, v in act_data.items():
# if k.find('module') != -1 :
name = k[7:] # remove `module.`
new_state_dict[name] = v
act_net = ACT_net(self.classes, self.sample_duration)
if resnet_path is not None:
act_net.create_architecture(model_path=resnet_path)
else:
act_net.create_architecture()
act_net.load_state_dict(new_state_dict)
self.act_net = act_net
else:
self.act_net = ACT_net(self.classes, self.sample_duration)
if resnet_path is not None:
self.act_net.create_architecture(model_path=resnet_path)
else:
self.act_net.create_architecture()
# load lstm
if rnn_path != None:
# act_rnn = Act_RNN(self.p_feat_size,int(self.p_feat_size/2),self.n_classes)
# act_rnn_data = torch.load(rnn_path)
# act_rnn.load_state_dict(act_rnn_data)
act_rnn = nn.Sequential(
# nn.Linear(64*self.sample_duration, 256),
# nn.ReLU(True),
# nn.Dropout(0.8),
# nn.Linear(256,self.n_classes),
nn.Linear(64 * self.sample_duration, self.n_classes),
# nn.ReLU(True),
# nn.Dropout(0.8),
# nn.Linear(256,self.n_classes),
)
act_rnn_data = torch.load(rnn_path)
act_rnn.load_state_dict(act_rnn_data)
self.act_rnn = act_rnn
else:
# self.act_rnn =Act_RNN(self.p_feat_size,int(self.p_feat_size/2),self.n_classes)
self.act_rnn = nn.Sequential(
# nn.Linear(64*self.sample_duration, 256),
# nn.ReLU(True),
# nn.Dropout(0.8),
# nn.Linear(256,self.n_classes),
nn.Linear(64 * self.sample_duration, self.n_classes),
# nn.ReLU(True),
# nn.Dropout(0.8),
# nn.Linear(256,self.n_classes),
)
for m in self.act_rnn.modules():
if m == nn.Linear:
m.weight.data.normal_().fmod_(2).mul_(stddev).add_(
mean) # not a perfect approximation