warpuv / Resnet-18_mindspore

Image classification done with Mindspore technology

Hack Cambridge 2021

This repo contains our team's submission to the Huawei Challenge on Hack Cambridge 2021 Hackathon (Problem statement).

Motive

To utilize Huawei's Mindspore package for performing image classification on edge devices,i.e, devices with minimal computation power.

What we used

• CIFAR-10 dataset
• Resnet 18
• Quantization Aware Training
• Image preprocessing and parallelism achieved by Mindspore package

Files description

• preprocess.py
  Applies image preprocessing operations from mindspore library
  
• resnetv2.py
  Defines implementation of Resnet architecture
  
• train.py
  Driver code for training the model
  
• eval.py
  Evaluation script for evaluating an particular model checkpoint
  
• checkpoint_40eps.ckpt
  Trained QAT-Resnet18 model.
• pyRates_85_HAVE_Bonus.xlsx
  Result Excel Sheet

Model Details

• Resnet-18 architecture model was trained using Quantization-Aware-Training(QAT) method.
• Adam optimizer and SoftmaxCrossEntropyWithLogits loss function were used.
• The model was trained for a total of 40 epochs on a NVIDIA-1070 GPU using mindspore-ai backend.

Recreation

Steps for training the model:

$python train.py

usage: train.py [-h] [-bsize BATCHSIZE] [-repeatNum REPEATNUM] [-dir SAVEDIR]
               [-e EPOCH] [-opt OPTIMIZER] [-lr LEARNINGRATE] [-m MOMENTUM]
               [-wDecay WEIGHTDECAY]

optional arguments:
  -h, --help            show this help message and exit
  -bsize BATCHSIZE, --batchsize BATCHSIZE
                        batch size
  -repeatNum REPEATNUM, --repeatNum REPEATNUM
                        repeat num
  -dir SAVEDIR, --savedir SAVEDIR
                        save directory
  -e EPOCH, --epoch EPOCH
                        no of epochs
  -opt OPTIMIZER, --optimizer OPTIMIZER
                        optimizer
  -lr LEARNINGRATE, --learningRate LEARNINGRATE
                        learning rate
  -m MOMENTUM, --momentum MOMENTUM
                        momentum
  -wDecay WEIGHTDECAY, --weightDecay WEIGHTDECAY
                        optimizer

Steps for evaulating the model:

$python eval.py

usage: eval.py [-h] [-loc CHECK_POINT]

optional arguments:
  -h, --help            show this help message and exit
  -loc CHECK_POINT, --check_point CHECK_POINT
                        Model checkpoint file

Results

Model Architecture	Accuracy
Resnet-18	87.86

Contributions

• Submitted a pull request to Mindspore on implemented Resnet-18 Architecture

Future Scope

• Implementing Adversarial Training to make the model more robust to real-time data.