pallet_detection #19
pallet_detection #19
Conversation
|
|
||
| def objective(trial): | ||
| # # Define the hyperparameter search space | ||
| res = trial.suggest_categorical("resolution", [384]) # square resolutions |
There was a problem hiding this comment.
would use smaller resolution, 256 should be plenty (lower resolution -> smaller model -> less overfit)
There was a problem hiding this comment.
I noticed that I didn't make any changes for optimisation, so I left them here. These parameters are not used during training manually, and will be adjusted before the start.
| heatmap_size=(48, 64), | ||
| sigma=2, | ||
| unbiased=True) | ||
| type='MSRAHeatmap', input_size=(448, 448), heatmap_size=(112, 112), sigma=2) |
There was a problem hiding this comment.
Also, I will set the change that is adjusted before training to a lower value by default (224x224)
|
|
||
| # pipelines | ||
| train_pipeline = [ | ||
| dict(type='LoadImage'), | ||
| dict(type='GetBBoxCenterScale'), | ||
| dict(type='RandomFlip', direction='horizontal'), |
There was a problem hiding this comment.
Perhaps I misunderstood you, but on the call we decided to remove it, I can return it.
There was a problem hiding this comment.
my question was: can we return it? so, does it generate correct annotations?
Probably the anwer is no, right? because then the bottom left would be bottom right
There was a problem hiding this comment.
then, to get more data, it might be useful to do a custom version of it, that flips the keypoints and also left/right names
No |
|
@jplapp |
|
I think using RepeatDataset would be helpful. Could you also commit the changes now used for training? |
No description provided.