[Fix] Fix bugs in RandomConcatDataset, SeqCropLikeStark and TridentSampling #438
Conversation
mmtrack/datasets/dataset_wrappers.py
Outdated
| self.CLASSES = datasets[0].CLASSES | ||
| if hasattr(datasets[0], 'flag'): | ||
| flags = [] | ||
| for i in range(0, len(datasets)): | ||
| flags.append(datasets[i].flag) | ||
| self.flag = np.concatenate(flags) |
There was a problem hiding this comment.
the isinstance(dataset, BaseSOTDataset) in build_dataloader need it. The RandomSampleConcatDataset is not an instance of BaseSOTDataset. It will go into other case which needs flag.
I have tried an alternative.
| extra_template_index = self.random_sample_inds( | ||
| video_visibility, | ||
| num_samples=1, | ||
| frame_range=[min_ind, max_ind], | ||
| allow_invisible=True)[0] | ||
| allow_invisible=False)[0] |
There was a problem hiding this comment.
the bboxes of invisible object are invalid, and it will cause interruptions in the pipeline.
Codecov Report
@@ Coverage Diff @@
## master #438 +/- ##
==========================================
+ Coverage 71.03% 71.20% +0.16%
==========================================
Files 119 119
Lines 6747 6748 +1
Branches 1295 1294 -1
==========================================
+ Hits 4793 4805 +12
+ Misses 1558 1548 -10
+ Partials 396 395 -1
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
| @@ -56,10 +57,13 @@ def build_dataloader(dataset, | |||
| DataLoader: A PyTorch dataloader. | |||
| """ | |||
| rank, world_size = get_dist_info() | |||
| is_sotdataset = isinstance(dataset, BaseSOTDataset) or ( | |||
| isinstance(dataset, ConcatDataset) | |||
| and isinstance(dataset.datasets[0], BaseSOTDataset)) | |||
There was a problem hiding this comment.
add comments for isinstance(dataset, ConcatDataset) and isinstance(dataset.datasets[0], BaseSOTDataset)
| @@ -32,6 +32,8 @@ def __init__(self, dataset_cfgs, dataset_sampling_weights=None): | |||
| ] | |||
|
|
|||
| datasets = [build_dataset(cfg) for cfg in dataset_cfgs] | |||
| self.CLASSES = datasets[0].CLASSES | |||
There was a problem hiding this comment.
add comments that why we need CLASSES
No description provided.