Playing Cards v2.3

DVC Report

Parameters

task	dataset	weights	shape	epochs	batch-size	optimizer.method	optimizer.learning-rate	optimizer.strategy.name	optimizer.strategy.parameters.decay	optimizer.strategy.parameters.warmup-learning-rate	optimizer.strategy.parameters.warmup-epochs	metric	logs	checkpoints	validation.iou	validation.score	loss_weights.w_obj	loss_weights.w_cls	loss_weights.w_loc	display	augmentation.Flip	augmentation.RandomBrightnessContrast	augmentation.ToGray	augmentation.ChannelShuffle	augmentation.RandomFog	augmentation.ShiftScaleRotate	augmentation.Mosaic	encoding_iou	objects-size	use_fake_background	use_power_decoder	focus	skip-validation	skip-initial-validation	freeze-bn	weight-decay	freeze-backbone	anchors	compute-anchors	save-anchors	upsample	NCHW	version	strides
detection	dataset.yaml	coco	[240, 320, 3]	100	128	adam	0.01	warmup	0	0.0001	5	acc	out	out	0.5	0.45	1	1	1	0	0.5	0.2	0.2	0.01	0.01	0.3	0.3	0.3	large	True	True	False	1	False	False	1e-05	False	[[[69.0, 84.0], [83.0, 115.0], [101.0, 67.0]], [[109.0, 138.0], [131.0, 84.0], [147.0, 187.0]]]	True		resize	False	2.2.11	[16, 30]

Training Metrics

mACC 0.50	mACC 0.75	mACC 0.50-0.95	mAP 0.50	mAP 0.75	mAP 0.50-0.95	mAR 0.50	mAR 0.75	mAR 0.50-0.95	Overall ACC	Overall AP	Overall AR	Learning Rate	step
83	77.64	63.11	94.24	90.82	75.39	87.71	84.38	70.09	84.56	94.18	87.68	0.0001	99

Validation Metrics

Path	Overall ACC	Overall AP	Overall AR	Total Class FP	Total FN	Total GT	Total Loc FP	Total TP	mACC 0.50	mACC 0.50-0.95	mACC 0.75	mAP 0.50	mAP 0.50-0.95	mAP 0.75	mAR 0.50	mAR 0.50-0.95	mAR 0.75
best.h5	0.84018	0.90196	0.9064	8	30	406	32	368	0.8246	0.61969	0.76157	0.90082	0.71389	0.85966	0.90894	0.72027	0.86682
best.rtm	0.75	0.9375	0.77586	7	84	406	14	315	0.72997	0.48829	0.58237	0.93934	0.66768	0.81365	0.76862	0.54546	0.66514

Playing Cards 2.2

DVC Report

params.yaml

task	dataset	weights	shape	epochs	batch-size	optimizer.method	optimizer.learning-rate	optimizer.strategy.name	optimizer.strategy.parameters.decay	optimizer.strategy.parameters.warmup-learning-rate	optimizer.strategy.parameters.warmup-epochs	metric	logs	checkpoints	validation.iou	validation.score	loss_weights.w_obj	loss_weights.w_cls	loss_weights.w_loc	display	augmentation.Flip	augmentation.RandomBrightnessContrast	augmentation.ToGray	augmentation.ChannelShuffle	augmentation.RandomFog	augmentation.ShiftScaleRotate	augmentation.Mosaic	encoding_iou	objects-size	use_fake_background	use_power_decoder	focus	skip-validation	skip-initial-validation	freeze-bn	weight-decay	freeze-backbone	anchors	compute-anchors	save-anchors	upsample	NCHW	version	strides
detection	dataset.yaml	coco	[240, 320, 3]	100	100	adam	0.01	warmup	0	0.0001	5	acc	out	out	0.5	0.45	1	1	1	0	0.5	0.2	0.2	0.01	0.01	0.3	0.3	0.3	large	True	True	False	1	False	False	1e-05	False	[[[48.0, 61.0], [67.0, 90.0], [74.0, 48.0]], [[98.0, 121.0], [113.0, 74.0], [130.0, 172.0]]]	True		resize	False	2.2.6	[16, 30]

metrics.json

mACC 0.50	mACC 0.75	mACC 0.50-0.95	mAP 0.50	mAP 0.75	mAP 0.50-0.95	mAR 0.50	mAR 0.75	mAR 0.50-0.95	Overall ACC	Overall AP	Overall AR	Learning Rate	step
80.87	73.18	59.82	93.35	88.14	72.6	85.9	81.08	66.82	82.08	93.3	85.71	0.0001	99

Playing Cards 2.1

Playing Cards 2.1 has been updated to use native training, without Docker, using a Python virtual environment managed by pipenv. Training is performed using the latest ModelPack 2.2.5.

While Docker-based training is still supported the native training does address some limitations with DVC when running inside Docker related to multiple experiment support and Git.

v2.0

DVC Report

params.yaml

task	dataset	weights	shape	epochs	batch-size	optimizer.method	optimizer.learning-rate	optimizer.strategy.name	optimizer.strategy.parameters.decay	optimizer.strategy.parameters.warmup-learning-rate	optimizer.strategy.parameters.warmup-epochs	metric	logs	checkpoints	validation.iou	validation.score	loss_weights.w_obj	loss_weights.w_cls	loss_weights.w_loc	display	augmentation.Flip	augmentation.RandomBrightnessContrast	augmentation.ToGray	augmentation.ChannelShuffle	augmentation.RandomFog	augmentation.ShiftScaleRotate	augmentation.Mosaic	encoding_iou	letterbox	objects-size	use_fake_background	use_power_decoder	focus	anchors	strides
detection	dataset.yaml	coco	[240, 320, 3]	100	100	adam	0.01	warmup	0	0.0001	2	acc	out	out	0.5	0.45	1	1	1	0	0.5	0.2	0.2	0.01	0.01	0.3	0.3	0.3	False	large	True	True	True	[[[68.0, 85.0], [95.0, 118.0], [106.0, 65.0]], [[109.0, 174.0], [155.0, 91.0], [180.0, 183.0]]]	[30, 60]

metrics.json

mACC 0.50	mACC 0.75	mACC 0.50-0.95	mAP 0.50	mAP 0.75	mAP 0.50-0.95	mAR 0.50	mAR 0.75	mAR 0.50-0.95	Overall ACC	Overall AP	Overall AR	Learning Rate	step
0.705074	0.689452	0.595443	0.857923	0.847678	0.750065	0.797941	0.788078	0.697114	0.755245	0.86631	0.79803	0	99

Playing Cards 1.0

Initial release of ModelPack trained on Playing Cards dataset and using the DVC-based training and deployment pipeline.