About

This repo contains training pipeline on Tiny ImageNet dataset and REST API for classifier using FastAPI

Project structure

- database "contains all necessary functions for storing requests in sqlite"
- modules "contains model and dataset"
- service.py "contains FastAPI wrapper on model"
- train_solution.py "contains whole pipeline"

How to install

• Create virtualenv with python 3.7
• pip install -r requirements.txt
• place dataset in ./data folder

How to train

• python train_solution.py
• pytorch-lightning is used for training
• In results folder will be created folder with experiments with checkpoints and tb_logs
• to see metrics run tensorboard --log_dir results/name_of_experiment/tb_logs
• after training; csv file with test predictions will be created

Improvement steps

• use pre-trained model (+28% acc)
• resize image while using pre-trained model
• use better backbone (efficient net)
• use model with bigger capacity (b0 ef_net vs b4)
• use normalization params from original ImageNet dataset
• change optimizer to more robust Adam
• lower lr while doing fine tuning of pretrained model
• add scheduler ReduceLROnPlateau(patience=1, min_lr=1e-6, mode='min', factor=0.3)
• add label smoothing stabilize training (+1% acc)

Possible improvements (that i did not tried):

• train for longer time (Can't do that because of lack of computing power)
• make ensemble of models (usually it helps), starting from simple probabilities averaging and take predictions from different efficient nets
• make tricky augmentations for example from ablumentations
• try Focal Loss (in my experinece gives an improvement)
• increase capacity of model for example b7 efficient net (did not tried it because of my small gpu)
• use SOTA model on ImageNet dataset according to paperswithcode for example efficient net b7 noisy student
• make even more experiments with optimizer and schedulers. For example use LR Cyclic scheduler, lr warm up and decay
• use Test Time Augmentations

Result:

• model: pretrained efficeint_netb4 with label smoothing, Adam, ReduceLROnPlateau, EarlyStopping
• final val accuaracy: 0.76
• final csv with test predictions: download

How to run service (FastAPI):

• download model and place it to results_folder; service will use this checkpoint in inference
• uvicorn --host 0.0.0.0 service:app --reload
• POST: 127.0.0.1:8000/predict/
  • curl -X POST "http://127.0.0.1:8000/predict/" -H "accept: application/json" -H "Content-Type: multipart/form-data" -F "file=@path_to_image.jpg;type=image/webp"
• GET: 127.0.0.1:8000/get_all/
  • curl -X GET "http://127.0.0.1:8000/get_all/" -H "accept: application/json"
• Shows nice UI where you can check all endpoints 127.0.0.1:8000/docs