analysis result
Analysis result
Analysis result
Same dataset, same model ,same library
Frozen 100 base layer 15Epoch
- Model use 51 **mobileNetV2 **base layer and classification layer to training. Total 154 base layer.
- Traning mathon traning the classification, 5 epoch traning the 54 base layer with classification . fine-tune
- Dataset : Fruit360 in kaggle
| Chip type | Colab TPU | Colab GPU |
|---|---|---|
| model classification training Spend time | 644 s | |
| Model 54 base layer with head training spend time | 331 s | |
| Train Acceleration | 0.9002 = 90.02% | |
| Train loss | 0.3169 = 31.69% | |
| Valiation acceleration | 0.9368 = 93.68% | |
| Valiation loss | 0.4223 = 42.23% | |
| Test acceleration | 93.75% |
For Detect Result
10 images averaged, every images only one fruit
Type
Computer use tensorflow
Rpi use tensorflow-lite
| Condition | Computer | Computer+Coral | Rpi | Rpi + Coral |
|---|---|---|---|---|
| Speed | - | 2.5ms | - | 3.8ms |
| Accuracy | 86.52% | 72.93% | 84.23% | 71.52% |
GPU Curve
- only train classification head acc and loss
- 10 Epoch
- Cross Entropy = Loss
- Continue training 5 Epoch for 54 base layer of MobileNetV2 with classification head
