MobileNetV2 Validation & Retraining with Reduced-precision Weights

Dependencies

• Pytorch (ver.1.1.0)
• torchvision
• tqdm (for progress bar)
• You can check other dependencies on top of *.py files

MobileNetV2(Model + Pretrained weights & biases)

• Paper: "Inverted Residuals and Linear Bottlenecks"

• Model Code & Imagenet Dataset

• Model Architecture: See the model_info folder's text files or run model_info.py

  Top 1 & Top 5

  • Pretrained model: './data/mobilenet_v2-b0353104.pth'
  • Top 1 Accuracy: 71.88 %
  • Top 5 Accuracy: 90.29 %

Usage

1. model_info.py

• Model Information(Architecture, Parameters, etc.)
• If you want to get other pretrained model's parameter values, you have to change the weight path(-w).
• All default outputs are in the model_info folder.

$ python3 model_info.py -h
usage: model_info.py [-h] [-w PATH] [-i #]

Model Information

optional arguments:
  -h, --help                    show this help message and exit
  -w PATH, --weights PATH       Pretrained parameters PATH | Default: ./data/mobilenet_v2-b0353104.pth                      
  -i #, --info_num #            Info you want | 0: summary 1: arcitecture 2: paramters value | Default: 0      

2. inference.py

• Inference form an image.
• Run on CPU and GPU.

$ python3 inference.py -h
usage: inference.py [-h] [-w PATH] [-i PATH]

Inference only one image

optional arguments:
  -h, --help                    show this help message and exit
  -w PATH, --weights PATH       Pretrained parameters PATH | Default: ./data/mobilenet_v2-b0353104.pth                      
  -i PATH, --input PATH         Input image PATH | Default: ./data/dog.jpg                

3. validation.py

• Validate a trained model with ImageNet validation set.
• Run on CPU and GPU.

$ python3 validation.py -h
usage: validation.py [-h] [-b #] [-p PATH] [-w PATH]

Float Shift Validation

optional arguments:
  -h, --help                   show this help message and exit
  -b #, --batch #              Batch Size | Default: 400
  -p PATH, --path PATH         Imagenet Dataset PATH | Default: /Data/ImageNet/ILSVRC2012/                
  -w PATH, --weights PATH      Pretrained parameters PATH | Default: ./data/mobilenet_v2-b0353104.pth

4. train.py

• Train from (pre)trained model using ImageNet training set(--pretrained True) or, train from scratch(--pretrained False).
• After every epoch, validation is performed.
• Run on only GPU.

$ python3 train.py -h
usage: train.py [-h] [-b #] [-e #] [-l #] [-d T/F] [--pretrained T/F] [-s T/F] [-p PATH] [-w PATH]

Training MobileNetV2

optional arguments:
  -h, --help                  show this help message and exit
  -b #, --batch #             Batch Size | Default: 400
  -e #, --epoch #             Epoches | Default: 1
  -l #, --lr #                Learning Rate | Default: 0.045
  -d T/F, --decay T/F         Learning Rate Decay | Default: True
  --pretrained T/F            Train from pretrained model | Default: True
  -s T/F, --save T/F          Save all models after every epoch(True) | Save best model(False)
  -p PATH, --path PATH        Imagenet Dataset PATH | Default: /Data/ImageNet/ILSVRC2012/
  -w PATH, --weights PATH     Pretrained parameters PATH | Default: ./data/mobilenet_v2-b0353104.pth

5. mask_frac.py

• Mask parameter values' fractional bits.
• Output will be saved in data folder.
• Takes about 2 minutes.

ex) 32FP(sign: 1bit, exp: 8bits, frac: 23bits)

- Before
--------------------------------------------
|s|-exponent-|----------fractional----------
--------------------------------------------

- After(Remain n_digits bits in fractional)
--------------------------------------------
|s|-exponent-|--frac--|--------zeros--------
--------------------------------------------

$ python3 train.py -h
usage: mask_frac.py [-h] [-n #] [-w PATH]

Mask Weights(FP32) Fractional

optional arguments:
  -h, --help                  show this help message and exit
  -n #, --n-digits #          Fractional bits | Default: 5
  -w PATH, --weights PATH     Pretrained parameters PATH | Default: ./data/mobilenet_v2-b0353104.pth