Notes:
- provided two frozen graphs files for effcientnet-b0, one with (1,224, 224, 3) as input shape one with no shape
- input name: input_tensor; output name: softmax_tensor
- python inference.py --model efficientnet-b0.pb --data_dir /path_to_imagenet_val_tfrecord --batch_size 1 --intra 1 --inter 16
- added efficientnet-S.pb which is from efficientnet-edgetpu-S, please modify the inference.py to do the test. Change MEAN_RGB and STDDEV_RGB to [127.0,127.0,127.0] and [128.0,128.0,128.0], and notice this model has 1001 classes instead of 1000
2080TI, TensorFlow (FP32)
efficientnet-b0 top1 accuracy: 76.62 top5 accuracy: 93.03 latency_mean(ms): 7.6744
efficientnet-S.pb: top1 accuracy: 77.31 top5 accuracy: 93.66 latency_mean(ms): 4.4133
TensorRT FP32
efficientnet-b0 top1 accuracy: 76.62 top5 accuracy: 93.03 latency_mean(ms): 3.4078
effcientnet-edgetpu-S top1 accuracy: 77.31 top5 accuracy: 93.66 latency_mean(ms): 2.2195
FP16
efficientnet-b0 top1 accuracy: 76.63 top5 accuracy: 93.03 latency_mean(ms): 3.2828
effcientnet-edgetpu-S top1 accuracy: 77.32 top5 accuracy: 93.66 latency_mean(ms): 1.5076
INT8
Renset50-INT8 top5 accuracy: 93.11 latency_mean(ms): 1.4580