Running and benchmarking diffrent neural network models using tensorflow lite and esp-nn.
Repo is founding on https://github.com/espressif/esp-tflite-micro/tree/master/examples/person_detection.
TFLite, int8 quantized, Inference_time, ms | Google Colab CPU | S3 opt ratio | Amlogic S905x3 | S3 opt ratio | Allwinner H3 | S3 opt ratio | Raspberry pi 4 | S3 opt ratio | ESP32-S3 esp-nn | ESP32-S3 non opt |
---|---|---|---|---|---|---|---|---|---|---|
int8_NoActivation_reshape_mbnetv3_small_1.0_96x96.tflite | 3,7 | 31,62162162 | 5,4 | 21,66666667 | 16,5 | 7,090909091 | 4 | 29,25 | 117 | 547 |
int8_NoActivation_reshape_mbnetv3_minim_0.35_96x96.tflite | 1,75 | 30,85714286 | 2,35 | 22,9787234 | 7,17 | 7,531380753 | 1,75 | 30,85714286 | 54 | 188 |
int8_NoAct_mobilenetV2_0.35_96x96.tflite | 7,9 | 24,43037975 | 8,4 | 22,97619048 | 27,85 | 6,929982047 | 6,8 | 28,38235294 | 193 | 813 |
int8_NoAct_mobilenetV2_0.05_96x96.tflite | 5,3 | 24,33962264 | 5,8 | 22,24137931 | 19,07 | 6,764551652 | 4,7 | 27,44680851 | 129 | 458 |
int8_NoAct_mobilenetV1_0.35_96x96.tflite | 6,1 | 44,75409836 | 9 | 30,33333333 | 24 | 11,375 | 6,3 | 43,33333333 | 273 | 763 |
int8_NoAct_mobilenetV1_0.05_96x96.tflite | 0,9 | 35,55555556 | 1,84 | 17,39130435 | 4,695 | 6,815761448 | 1,41 | 22,69503546 | 32 | 48 |
int8_NoActivation_reshape_mbnetv3_minim_1.0_224x224.tflite | 23 | 25,73913043 | #DIV/0! | #DIV/0! | #DIV/0! | 592 | 3151 | |||
CoreMark 1.0 | 19043,99162 | 31,6182199 | 7926,785328 | 13,16062549 | 3707,365299 | 6,155237494 | 7947,402644 | 13,19485586 | 602,310683 | |
CoreMark * inference_time | ||||||||||
int8_NoActivation_reshape_mbnetv3_small_1.0_96x96.tflite | 70462,769 | 42804,64077 | 61171,52743 | 31789,61058 | 70470,34991 | |||||
int8_NoActivation_reshape_mbnetv3_minim_0.35_96x96.tflite | 33326,98534 | 18627,94552 | 26581,80919 | 13907,95463 | 32524,77688 | |||||
int8_NoAct_mobilenetV2_0.35_96x96.tflite | 150447,5338 | 66584,99676 | 103250,1236 | 54042,33798 | 116245,9618 | |||||
int8_NoAct_mobilenetV2_0.05_96x96.tflite | 100933,1556 | 45975,3549 | 70699,45625 | 37352,79243 | 77698,07811 | |||||
int8_NoAct_mobilenetV1_0.35_96x96.tflite | 116168,3489 | 71341,06795 | 88976,76718 | 50068,63666 | 164430,8165 | |||||
int8_NoAct_mobilenetV1_0.05_96x96.tflite | 17139,59246 | 14585,285 | 17406,08008 | 11205,83773 | 19273,94186 | |||||
Keras, f32, Inference_time, ms | Google Colab CPU | S3 opt ratio | Amlogic S905x3 | S3 opt ratio | Allwinner H3 | S3 opt ratio | Raspberry pi 4 | S3 opt ratio | ESP32-S3 esp-nn | ESP32-S3 non opt |
int8_NoActivation_reshape_mbnetv3_minim_1.0_224x224.tflite | 100 | 5,92 | 592 | 3151 | ||||||
int8_NoActivation_reshape_mbnetv3_minim_0.35_224x224.tflite | 92 | 3,402173913 | 313 | 1182 | ||||||
int8_NoActivation_reshape_mbnetv3_small_1.0_96x96.tflite | 110,37 | 1,060070671 | 430,4 | 0,2718401487 | #DIV/0! | 117 | 547 | |||
int8_NoActivation_reshape_mbnetv3_minim_0.35_96x96.tflite | 98,7 | 0,547112462 | 367,4 | 0,1469787697 | #DIV/0! | 54 | 188 | |||
int8_NoAct_mobilenetV2_0.35_96x96.tflite | #DIV/0! | #DIV/0! | #DIV/0! | 193 | 813 | |||||
int8_NoAct_mobilenetV2_0.05_96x96.tflite | #DIV/0! | #DIV/0! | #DIV/0! | 129 | 458 | |||||
int8_NoAct_mobilenetV1_0.35_96x96.tflite | #DIV/0! | #DIV/0! | #DIV/0! | 273 | 763 | |||||
int8_NoAct_mobilenetV1_0.05_96x96.tflite | #DIV/0! | #DIV/0! | #DIV/0! | 32 | 48 | |||||
ESP32-S3 | f32 TFLite esp-nn | f32 TFLite non opt | i8 TFLite esp-nn | i8 TFLite non opt | f32 TVM | |||||
int8_NoActivation_reshape_mbnetv3_small_1.0_96x96.tflite | 1201 | 1261 | 117 | 547 | 602 |