MNIST Classification:

Goal:

DagsHub started to support 400 datasets from AWS Registry, covering geology, biology, tabular data, audio, computer vision, NLP, and Urban datasets. The supported datasets can be found in DagsHub Dataset.The goal for this project is to train a computer vision model using the computer vision dataset provided by DagsHub Dataset.

Dataset:

fast-ai-imageclas-dataset

Requirement Package:

• torch
• torchvision
• mlflow
• pandas ​

1. Config:

• param.yaml: configuration for trainval / logging / prepare ​

2. Data Preprocess:

python ./src/prepare.py

• Using Direct Data Accessing functionality provided by DagsHub to get access to the mnist dataset in the fast-ai-imageclas-dataset
• The script will output train.csv, test.csv and labels.txt ​

3. Train / Val Model:

python ./src/main.py

• Classification model support ResNet18, ResNet34, ResNet50, ResNet101
• The training process is accelerated by the m1-pro chip. (Change the gpu_type in params.yaml)
• Save final model to ./model folder
• Default Training Parameters:
  • gpu_type: mps
  • batchsize: 64
  • model_type: ResNet18
  • loss_type: CrossEntropyLoss
  • epoch: 5
  • opt: SGD
  • opt_params:
    • lr: 0.01
    • momentum: 0.9
    • weight_decay: 0.0005
  • lr_sched: StepLR
  • lr_sched_params:
    • step_size: 2
    • gamma: 0.1
  • img_resize: [32, 32]
  • img_crop: [28, 28]
  • img_rflip_p: 0.5
  • img_mean: [0, 0, 0]
  • img_std: [1, 1, 1]

Baseline Performance:

categories	Accuracy (%)
class: 0	98.9 %
class: 1	98.8 %
class: 2	94.9 %
class: 3	99.3 %
class: 4	97.0 %
class: 5	96.5 %
class: 6	98.9 %
class: 7	96.2 %
class: 8	99.6 %
class: 9	98.4 %
Total	97.9 %