Skip to content

Latest commit

 

History

History
144 lines (100 loc) · 5.09 KB

README.md

File metadata and controls

144 lines (100 loc) · 5.09 KB

Step-by-Step

This document is used to list steps of reproducing MXNet ResNet18_v1/ResNet50_v1/Squeezenet1.0/MobileNet1.0/MobileNetv2_1.0/Inceptionv3 tuning zoo result.

Prerequisite

1. Installation

pip install -r requirements.txt

2. Prepare Dataset

You can use prepare_dataset.sh to download dataset for this example. like below:

bash ./prepare_dataset.sh --dataset_location=./data

This will download validation dataset val_256_q90.rec, it will put to the directory ./data/

3. Prepare Pre-trained model

You can use prepare_model.py to download model for this example. like below:

python prepare_model.py --model_name mobilenet1.0 --model_path ./model

This will download the pre-trained model mobilenet1.0, and then put to the directory ./model/. For more details, see below:

python prepare_model.py -h

usage: prepare_model.py [-h]
                        [--model_name {resnet18_v1,resnet50_v1,squeezenet1.0,mobilenet1.0,mobilenetv2_1.0,inceptionv3}]
                        [--model_path MODEL_PATH] [--image_shape IMAGE_SHAPE]

Prepare pre-trained model for MXNet ImageNet Classifier

optional arguments:
  -h, --help            show this help message and exit
  --model_name {resnet18_v1,resnet50_v1,squeezenet1.0,mobilenet1.0,mobilenetv2_1.0,inceptionv3}
                        model to download, default is resnet18_v1
  --model_path MODEL_PATH
                        directory to put models, default is ./model
  --image_shape IMAGE_SHAPE
                        model input shape, default is 3,224,224

Run

ResNet18_v1

bash run_tuning.sh --topology=resnet18_v1 --dataset_location=./data/val_256_q90.rec --input_model=/PATH/TO/MODEL --output_model=./ilit_resnet18

ResNet50_v1

bash run_tuning.sh --topology=resnet50_v1 --dataset_location=./data/val_256_q90.rec --input_model=/PATH/TO/MODEL --output_model=./ilit_resnet50_v1

SqueezeNet1

bash run_tuning.sh --topology=squeezenet1.0 --dataset_location=./data/val_256_q90.rec --input_model=/PATH/TO/MODEL --output_model=./ilit_squeezenet

MobileNet1.0

bash run_tuning.sh --topology=mobilenet1.0 --dataset_location=./data/val_256_q90.rec --input_model=/PATH/TO/MODEL --output_model=./ilit_mobilenet1.0

MobileNetv2_1.0

bash run_tuning.sh --topology=mobilenetv2_1.0 --dataset_location=./data/val_256_q90.rec --input_model=/PATH/TO/MODEL --output_model=./ilit_mobilenetv2_1.0

Inception_v3

bash run_tuning.sh --topology=inceptionv3 --dataset_location=./data/val_256_q90.rec --input_model=/PATH/TO/MODEL --output_model=./ilit_inception_v3

Benchmark

Use resnet18 as an example:

# accuracy mode, run the whole test dataset and get accuracy
bash run_benchmark.sh --topology=resnet18_v1 --dataset_location=./data/val_256_q90.rec --input_model=./model --batch_size=32 --mode=accuracy

# benchmark mode, specify iteration number and batch_size in option, get throughput and latency
bash run_benchmark.sh --topology=resnet18_v1 --dataset_location=./data/val_256_q90.rec --input_model=./model --batch_size=32 --iters=100 --mode=benchmark

Examples of enabling Intel® Low Precision Optimization Tool auto tuning on MXNet ResNet50

This is a tutorial of how to enable a MXNet classification model with Intel® Low Precision Optimization Tool.

User Code Analysis

Intel® Low Precision Optimization Tool supports two usages:

  1. User specifies fp32 "model", calibration dataset "q_dataloader", evaluation dataset "eval_dataloader" and metric in tuning.metric field of model-specific yaml config file.

  2. User specifies fp32 "model", calibration dataset "q_dataloader" and a custom "eval_func" which encapsulates the evaluation dataset and metric by itself.

As ResNet50_v1/Squeezenet1.0/MobileNet1.0/MobileNetv2_1.0/Inceptionv3 series are typical classification models, use Top-K as metric which is built-in supported by Intel® Low Precision Optimization Tool. So here we integrate MXNet ResNet with Intel® Low Precision Optimization Tool by the first use case for simplicity.

Write Yaml config file

In examples directory, there is a template.yaml. We could remove most of items and only keep mandotory item for tuning.

#conf.yaml

framework:
  - name: mxnet

tuning:
    metric:
      - topk: 1
    accuracy_criterion:
      - relative: 0.01
    timeout: 0
    random_seed: 9527

Here we choose topk built-in metric and set accuracy target as tolerating 0.01 relative accuracy loss of baseline. The default tuning strategy is basic strategy. The timeout 0 means early stop as well as a tuning config meet accuracy target.

code update

After prepare step is done, we just need update imagenet_inference.py like below.

    import ilit
    fp32_model = load_model(symbol_file, param_file, logger)
    cnn_tuner = Tuner("./cnn.yaml")
    cnn_tuner.tune(fp32_model, q_dataloader=data, eval_dataloader=data)

The tune() function will return a best quantized model during timeout constrain.