README.md

Efficient-CV

A repo to evaluate different low-latency inference methods using PyTorch

• data: Cifar-100
• techniques:
  • Quantization
  • Pruning
  • Knowledge Distillation

1. Train baseline model

• Install dependency

pip install -r requirements.txt

• Train baseline model (resnet18)

python efficient_cv/train.py --do_train --do_eval --n_gpu 1 --optim sgd --num_epochs 100 --batch_size 128 --scheduler multistep --lr 0.1 --model_name resnet18

2. Train distillation student model

• Direct logits training (resnet10) (paper):

python efficient_cv/train.py --do_train --n_gpu 1 --optim sgd --num_epochs 200 --batch_size 128 --scheduler multistep --lr 0.1 --model_name resnet18 --student_model_name resnet10 --test_model_path <trained-model-path-from-step-2> --distill_method logits --distill_weight 0.1

• Soft target training (resnet10) (paper):

python efficient_cv/train.py --do_train --n_gpu 1 --optim sgd --num_epochs 200 --batch_size 128 --scheduler multistep --lr 0.1 --model_name resnet18 --student_model_name resnet10 --test_model_path <trained-model-path-from-step-2> --distill_method soft --distill_weight 0.1 --temperature 4

3. Evalution

• Evaluate model with low-latency techniques

python efficient_cv/train.py --do_eval --test_model_path <trained-model-path-from-step-2> --n_gpu 1 --benchmarks baseline,quantization,fp16,pruning,distillation --model_name resnet18 --student_model_name resnet10 --student_test_model_path <trained-model-path-from-step-3>

Results

---

Distillation method refers to https://github.com/AberHu/Knowledge-Distillation-Zoo

Cifar100 baseline method refers to https://github.com/weiaicunzai/pytorch-cifar100