tools/tensorflow-quantization/examples/mobilenet/README.md
This script presents a QAT end-to-end workflow (TF2-to-ONNX) for MobileNet models in tf.keras.applications.
Requirements • Workflow • Results
Install base requirements and prepare data. Please refer to examples' README.
Similar to ResNet: different model and different input pre-processing (
mobilenet).
Please run the following to quantize, fine-tune, and save the final graph in SavedModel format (checkpoints are also saved).
python run_qat_workflow.py
Step 1 already does the conversion from SavedModel to ONNX automatically. For manual steps, please see step 3 in EfficientNet's README.
Please refer to the examples' README.
Results obtained on NVIDIA's A100 GPU and TensorRT 8.4.10.1.
| Model | TF (%) | TF latency (ms, bs=1) | TRT(%) | TRT latency (ms, bs=1) |
|---|---|---|---|---|
| Baseline | 70.60 | 1.99 | 70.60 | 0.32 |
| PTQ | - | - | 69.31 | 0.16 |
| QAT | 70.51 (ep2) | 50.49 | 70.43 | 0.16 |
Note: no residual connections exist in MobileNet-v1.
| Model | TF (%) | TF latency (ms, bs=1) | TRT(%) | TRT latency (ms, bs=1) |
|---|---|---|---|---|
| Baseline | 71.77 | 3.71 | 71.77 | 0.55 |
| PTQ | - | - | 70.87 | 0.30 |
| QAT | 71.68 (ep1) | 74.27 | 71.62 | 0.30 |
Note: residual connections exist in MobileNet-v2.
piecewise_sgd, lr_schedule=[(1.0, 1), (0.1, 2), (0.01, 7)] (default)bs=64, ep=10, lr=0.001bs=64.Conv->BN->((Add->Clip->Mul), ())->Mul), which is not a kernel fusion in TRT.