Keep Track of Submissions

[JoshVarty]

In order to keep track of what we've submitted, let's log information about our submissions here.

• ID Number to identify submission/model. If you need to add additional comments, please leave a comment below and include the ID number.
• Description Quick model description.
• Epochs Number of epochs during training
• Loss Loss function used for training
• Valid (Img) The validation loss given by fastai against the validation images
• Valid (Video) The validation loss given by the inference kernel against the validation videos
• Accuracy Accuracy at the end of the training
• LB The public leaderboard score returned by Kaggle

Results

ID	Description	epochs	Loss	Valid (Img)	Valid (Video)	Accuracy	LB
0	EfficientNet-B1	10		0.207	N/A		0.47683
0.a	EfficientNet-B1 w/RetinaFace	10		0.207	N/A		0.47536
1	EfficientNet-B1	15	CE	Not trained on same data	N/A	Not trained on same data	0.43543
2	EfficientNet-B1	15	CE	0.238 *	0.5693	0.901	0.40877
3	EfficientNet-B1	15	Smooth-CE	0.326 (0.205 CE)*	0.4938	0.928	0.41119
			cropped_facesV2
4	EfficientNet-B1	20	CE	0.2500*	0.6158	0.885	0.43069
			cropped_facesV3
5	EfficientNet-B1 (exp7)	25	CE	0.3163	0.4877	0.8704	0.46419
6	EfficientNet-B1 (exp10)	15	CE	0.3413	0.4915	0.8668	0.41556
7	3D ResNet18	10	CE	0.251659	0.38	0.886144	0.51749
8	Same as 2 but on 16frames	/	/	/	/	/	0.55941
9	Same as 2 but on 16frames (+Retina if len(faces)<8	/	/	/	/	/	0.64139

*Data Augmentation consisted of:

• Default get_transforms() parameters
• MotionBlur(blur_limit=9, p=.25)
• GaussNoise(var_limit=(15.0, 75.0), p=.25)
• JpegCompression(p=.25, quality_lower=30)
• downscale(scale=0.5, p=.25)