Table 2.
Parameter | YoloV5-large | YoloV5-small | Faster-R-CNN | RetinaNet | EfficientDet-D2 |
---|---|---|---|---|---|
Learning rate | 8.94e3 | 1.15e2 | 1.0e4 | 1.0e4 | 1.0e4 |
Optimizer | SGD | SGD | Adam | Adam | Adam |
Batch size | 8 | 16 | 16 | 16 | 16 |
Epochs | 20 | 20 | 25 | 60 | 20 |
SWA start | N/A | N/A | 20 | 45 | 15 |
SWA epochs | N/A | N/A | 5 | 15 | 5 |
Warmup epochs | 2.5 | 2.8 | 5.0 | 5.0 | 5.0 |
Gradient clipping value | N/A | N/A | 3.0 | 3.0 | 3.0 |
For gradient clipping, the gradients’ global norm is clipped to the reported values.