Add SSN model. #55
Add SSN model. #55
Conversation
Codecov Report
@@ Coverage Diff @@
## master #55 +/- ##
==========================================
+ Coverage 84.21% 85.08% +0.86%
==========================================
Files 73 75 +2
Lines 4175 4418 +243
Branches 638 674 +36
==========================================
+ Hits 3516 3759 +243
+ Misses 554 549 -5
- Partials 105 110 +5
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
JackyTown
force-pushed
the
tjq/ssn-model
branch
from
32e4aa2 to
15d853b
Compare
JackyTown
force-pushed
the
tjq/ssn-model
branch
from
15d853b to
d3821a2
Compare
mmaction/models/heads/ssn_head.py
Outdated
| dim=1) / num_multiplier | ||
| if scale_factors is not None: | ||
| part_feat = ( | ||
| part_feat * scale_factors.view(num_samples, 1)) |
There was a problem hiding this comment.
I tried it, but failed, there was an shape error.
| assert x.size(1) == self.feat_dim | ||
| num_ticks = proposal_ticks.size(0) | ||
|
|
||
| out_activity_scores = torch.zeros((num_ticks, self.activity_score_len), |
There was a problem hiding this comment.
zeros_like, so that both the dtype and device are correct
There was a problem hiding this comment.
torch.zeors_like() seems not suitable for this case.
mmaction/models/heads/ssn_head.py
Outdated
| scale_factor = 1.0 | ||
|
|
||
| sum_parts = sum(stage_cfg) | ||
| tick_left = float(ticks[stage_idx]) |
There was a problem hiding this comment.
is it safe to remove float?
There was a problem hiding this comment.
Maybe it is better to add float to tick_right. If tick_right is torch.Tensor(3, dtype=np.int32), then torch.arange(tick_left, tick_right+1e-5, num_parts) will not work.
There was a problem hiding this comment.
tick_right+1e-5 -> torch.Tensor(3, dtype=np.int32)
mmaction/models/heads/ssn_head.py
Outdated
| raw_scale_score = raw_score.mean( | ||
| dim=0) * scale_factor | ||
| out_scores[ | ||
| index, :] += raw_scale_score.detach().cpu() |
There was a problem hiding this comment.
is it possible to move computations to gpu?
| self.regressor_fc.out_features, | ||
| stpp_feat_multiplier, self.activity_fc.in_features).transpose( | ||
| 0, 1).contiguous().view(-1, self.activity_fc.in_features) | ||
| reg_bias = self.regressor_fc.bias.data.view(1, -1).expand( |
There was a problem hiding this comment.
is expand().contiguous() equivalent to repeat() ?
There was a problem hiding this comment.
I tried it, but failed, there was an shape error.
JackyTown
force-pushed
the
tjq/ssn-model
branch
from
3cdfd5c to
93f260e
Compare
No description provided.