Keep track of experiments

[nathanhubens]

Here are the results on some experiments I made:

EfficientNet-B1

Results

ID	Description	epochs	data augmentation	Loss function	Training Loss	Valid (Img)	Valid (Video)/Balanced Score	Accuracy
0	/	5	0	CE	0.0865	0.3131	0.6011	0.8668
1	/	5	1	CE	0.1075	0.3237	0.6120	0.8559
2	/	5	1	Label Smoothing CE	0.2692	0.4314	0.5366	0.8471
3	Oversampling	5	1	Label Smoothing CE	0.2599	0.4144	0.5392	0.8621
4	/	15	1	CE	0.0546	0.2898	0.5967	0.8896
5	OverSampling	15	1	CE	0.0742	0.3082	0.4982	0.8751
6	OverSampling	15	1	Label Smoothing CE	0.2325	0.3971	0.5429	0.8875
7	OverSampling	25	1	CE	0.1222	0.3163	0.317/0.319	0.8704
8	OverSampling	15	1	Focal Loss	0.1549	0.2954	0.34/0.37	0.8564
9	/	15	1	Focal Loss	0.0902	0.2853	0.30/0.33	0.8725
10	OverSampling	15	2	CE	0.0832	0.3413	0.316/0.325	0.8668
11	OverSampling	15	3	CE	0.1039	0.3398	0.5547	0.8512
12	Resnet18	15	1	CE	0.1789	0.3438	0.4213/0.5	0.7743
13	Undersampling	25	1	CE	0.0350	0.3643	0.31/0.32	0.8891
14	3DCNN Undersampling	13 (early stopping)	1	CE	0.128607	0.303398	0.4010/0.3842	0.875000

Data Augmentation:

[0]: .transform(get_transforms(max_warp=0), size=size,resize_method=ResizeMethod.PAD, padding_mode="zeros")
[1]

.transform(get_transforms(max_warp=0, xtra_tfms=tfms), size=size, resize_method=ResizeMethod.PAD, padding_mode="zeros")

tfms += [alb_tfm2fastai(A.MotionBlur(blur_limit=9, p=.35), order=11)]
tfms += [alb_tfm2fastai(A.GaussNoise(var_limit=(15.0, 75.0), p=.35), order=12)]
tfms += [alb_tfm2fastai(A.JpegCompression(quality_lower=30, p=.35), order=13)]
tfms += [downscale(scale=0.5, p=.35)]

[2]

.transform(get_transforms(max_warp=0, xtra_tfms=tfms), size=size, resize_method=ResizeMethod.PAD, padding_mode="zeros")

tfms += [alb_tfm2fastai(A.MotionBlur(blur_limit=11, p=.35), order=11)]
tfms += [alb_tfm2fastai(A.GaussNoise(var_limit=(15.0, 85.0), p=.35), order=12)]
tfms += [alb_tfm2fastai(A.JpegCompression(quality_lower=20, p=.35), order=13)]
tfms += [downscale(scale=0.4, p=.35)]

�[3]

tfms = []
tfms += [alb_tfm2fastai(A.MotionBlur(p=.3), order=11)]
tfms += [alb_tfm2fastai(A.GaussNoise(p=.3), order=12)]
tfms += [alb_tfm2fastai(A.JpegCompression(quality_lower=40, p=.3), order=13)]
tfms += [downscale(scale=0.5, p=.3)]

Full RetinaDetection:

Exp7. -> 0.481

Ensemble

• Exp7 + Exp10 -> 0.481

[JoshVarty]

Since we're going to run a lot more of these than submissions, I'm going to separate mine out into its own comment.

EfficientNet-B1

Results

ID	Description	epochs	data augmentation	Loss function	Training Loss	Valid (Img)	Valid (Video)	Accuracy
J.0	/	10	0	CE	0.0391	0.3716	0.5849	0.8746
J.1	RetinaFace fallback	10	0	CE	0.0391	0.3716	0.5205	0.8746

Data Augmentation:

[0]: get_transforms(max_warp=0, max_zoom=1)

[qAp]

EfficientNet-B0

Results

ID	Description	epochs	data aug	Loss func	Train Loss	Valid (Img)	Valid (Video)	Accuracy
0	oselect V3 class V3 infer V1 eval V2	25	0	L.S.C.E	0.4368	0.48	0.65	0.79
1	oselect V3 class V3 infer/eval V1	25	0	L.S.C.E	0.4368	0.48	0.45	0.79
2	oselect V3 class V3 infer/eval V2	25	0	L.S.C.E	0.4368	0.48	0.44	0.79

data aug:

[0]:

tfms = []
tfms += [alb_tfm2fastai(A.MotionBlur(blur_limit=9, p=.25), order=11)]
tfms += [alb_tfm2fastai(A.GaussNoise(var_limit=(15.0, 75.0), p=.25), order=12)]
tfms += [alb_tfm2fastai(A.JpegCompression(p=.25, quality_lower=30), order=13)]
tfms += [downscale(scale=0.5, p=.25)]
.transform(get_transforms(max_warp=0, xtra_tfms=tfms)