ResNet50

This folder contains a ResNet model for Image Classification using CIFAR-10 dataset. ResNet50 is convolutional neural network 50 layers deep.

Requirments

In order to use MindSpore, Windows-x64 is required. Python 3.7.5 must be installed, you can find how to do so here. For this project we have created a virtual environment local to this repository. This code has been implemented for --divice_target=CPU.

Folder Structure

Place your test and training files within the CIFAR-10 folder, the path has been specified within the code therefore it is important to place the target files in the right location.

CIFAR-10
   ├─cifar-10-batches-bin
   │
   └─cifar-10-verify-bin

Download the dataset from CIFAR-10.

If you are using linux simply use the command:

wget https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz

and

tar -zvxf cifar-10-binary.tar.gz

Once downloaded save your file such that your CIFAR-10 folder will look like the following

CIFAR-10
   ├─cifar-10-batches-bin
   │            │  
   │            │─batches.meta
   │            │─ data_batch_1.bin
   │            │─ data_batch_2.bin
   │            │─ data_batch_3.bin
   │            │─ data_batch_4.bin
   │            │─ data_batch_5.bin
   │
   │
   └─cifar-10-verify-bin
             │  
             │─batches.meta
             │─ data_batch_1.bin
             │─ data_batch_2.bin
             │─ data_batch_3.bin
             │─ data_batch_4.bin
             │─ data_batch_5.bin
             │─ test_batch.bin
             
    
   
   

Create a Virtual Env for Python 3.7 with virtualenv

• Install pyenv

• install pipenv using pip install pipenv or pip install -user pipenv

• navigate to the project directory, in this case ./mindSpore and use the following command:

pipenv install --python 3.7

now you may follow the instruction to activate your virtual environment.

Install MindSpore

To install MindSpore, follow the installation guide here. Select your machine specifications to obtain the correct command.

Usage

Training

python trainCIFAR_10.py --device_target=CPU  --pre_trained=[CHECKPOINT_PATH](optional)

The checkpoint given at 16:00 is MindSpore_Resnet50_cifar10.ckpt which is also the default for the evaluation script The configuration requirments are the following:

    "class_num": 10,
    "batch_size": 32,
    "loss_scale": 1024,
    "momentum": 0.9,
    "weight_decay": 1e-4,
    "epoch_size": 90,
    "pretrain_epoch_size": 0,
    "save_checkpoint": True,
    "save_checkpoint_epochs": 5,
    "keep_checkpoint_max": 10,
    "save_checkpoint_path": "./",
    "warmup_epochs": 5,
    "lr_decay_mode": "poly",
    "lr_init": 0.01,
    "lr_end": 0.00001,
    "lr_max": 0.1

Evaluation

python evalCIFAR_10.py --device_target=CPU  --checkpoint_path=[CHECKPOINT_PATH]

Accuracy 24/01/2021 morning

During the hackathon we have trained the model up to 2 epoch without starting from a checkpoint and the accuracy was ~ 50% (fixed typo)

============= Accuracy:{'Accuracy': 78% } ==============

For the result at submission tim 5pm please look at the resultfolder.

Checkpoint

A list of checkpoints can be found below:

• train_resnet_cifar10_1-1_15.ckpt with loss 0.38

• train_resnet_cifar10_1-1_25.ckpt with loss ~ 0.28

other checkpoints can be found here

Note

We initialy chose batch_num=128 and keep_checkpoint_max=38 and then used more suitable parameters. We suggest to use the configuration mentioned in the documentation (i.e. batch_num=5 and keep_checkpoint_max=10)