colab_p100.md

TensorFlow 2 Detection Model Zoo Benchmarks (Colab Tesla P100)

Environment

• Colab Pro
• HW
  • CPU: Intel(R) Xeon(R) CPU @ 2.30GHz 4 cores
  • GPU: NVIDIA Tesla P100
  • MEM: 25GB

Dataset

• COCO2017

How to benchmarks

Source

• NobuoTsukamoto/tensorrt

Notebook

• Colab Notebook

Parameters

• Input batch size: 1
• TF-TRT
  • Model
    • Native FP32
    • TF-TRT FP32
    • TF-TRT FP16
  • Converter Params
    • max_batch_size: 1
    • max_workspace_size_bytes: 2147483648
    • maximum_cached_engines: 1
    • minimum_segment_size: 3, 20, 50

Results

• All results: Colab_TF2.4.1_P100.csv
• Note: CenterNet MobileNetV2 (detection and keypoints) will result in an error inference by the saved model.

Model	Input	mAP	latency(ms)
CenterNet HourGlass104	512x512	38.47	100.49
CenterNet HourGlass104 Keypoints	512x512	36.73	108.91
CenterNet HourGlass104	1024x1024	39.76	277.89
CenterNet HourGlass104 Keypoints	1024x1024	38.48	299.93
CenterNet Resnet50 V1 FPN	512x512	27.59	37.05
CenterNet Resnet50 V1 FPN Keypoints	512x512	25.76	42.54
CenterNet Resnet101 V1 FPN	512x512	31.28	47.16
CenterNet Resnet50 V2	512x512	26.23	36.8
CenterNet Resnet50 V2 Keypoints	512x512	24.49	42.36
CenterNet MobileNetV2 FPN	512x512	0	0
CenterNet MobileNetV2 FPN Keypoints	512x512	0	0
EfficientDet D0	512x512	32.82	146.18
EfficientDet D1	640x640	36.63	210.53
EfficientDet D2	768x768	39.84	273.42
EfficientDet D3	896x896	42.48	356.11
EfficientDet D4	1024x1024	45.65	507.74
EfficientDet D5	1280x1280	46.14	825.46
EfficientDet D6	1280x1280	46.6	982.43
EfficientDet D7	1536x1536	47.75	1187.02
SSD MobileNet v2	320x320	20.02	47.42
SSD MobileNet V1 FPN	640x640	28.67	147.5
SSD MobileNet V2 FPNLite	320x320	22.1	71.34
SSD MobileNet V2 FPNLite	640x640	27.69	133.55
SSD ResNet50 V1 FPN RetinaNet50	640x640	33.7	154.92
SSD ResNet50 V1 FPN RetinaNet50	1024x1024	37.58	310.2
SSD ResNet101 V1 FPN RetinaNet101	640x640	35.03	162.86
SSD ResNet101 V1 FPN RetinaNet101	1024x1024	38.76	325.68
SSD ResNet152 V1 FPN RetinaNet152	640x640	34.85	177.29
SSD ResNet152 V1 FPN RetinaNet152	1024x1024	38.9	357.52
Faster R-CNN ResNet50 V1	640x640	27.94	94.27
Faster R-CNN ResNet50 V1	1024x1024	30.66	112.47
Faster R-CNN ResNet50 V1	800x1333	26.66	118.31
Faster R-CNN ResNet101 V1	640x640	30.2	103.87
Faster R-CNN ResNet101 V1	1024x1024	36.56	131.79
Faster R-CNN ResNet101 V1	800x1333	31.45	153.31
Faster R-CNN ResNet152 V1	640x640	31.16	115.56
Faster R-CNN ResNet152 V1	1024x1024	36.86	154.98
Faster R-CNN ResNet152 V1	800x1333	32.75	189.54
Faster R-CNN Inception ResNet V2	640x640	36.89	369.53
Faster R-CNN Inception ResNet V2	1024x1024	36.02	425.06
Mask R-CNN Inception ResNet V2	1024x1024	37.94	413.84

Latency mean (ms)