BirdClassifier

This repository gathers the code for bird image classification from the in-class CodaLab competition.

Use ResNeXt-101 32x8d with data augmentation and SGD with weight_decay optimizer + cosine annealing learning rate scheduler.

Reproducing Submission

Our model achieve 78.5031% accuracy in testing set.

To reproduce my submission without retrainig, do the following steps:

1. Requirements
2. Inference

Hardware

Ubuntu 18.04.5 LTS

Intel® Core™ i7-3770 CPU @ 3.40GHz × 8

GeForce GTX 1080/PCIe/SSE2

Requirements

All requirements should be detailed in requirements.txt.

$ virtualenv venv --python=3.6
$ source ./venv/bin/activate
$ cd BirdClassifier
$ pip install -r requirements.txt

Official images can be downloaded from CodaLab competition

Repository Structure

Run the following command to build the directory.

$ mkdir model
$ mkdir data

The repository structure is:

BirdClassifier
  +- data                     # all file used in the program 
  +- model                    # all trained models
  +- result                   # confusion matrix of two models and answer.txt save here
  +- src            
  ∣- dataset.py               # set a dataset class for loading imgs
  ∣- inference.py             # reproduce my submission file or test your model
  ∣- train.py                 # for training model
  - requirements.txt          # txt file for establishing the environment

Dataset Preparation

After downloading images from CodaLab competition, we expect the data directory is structured as:

data
  +- training_data         # all training data from CodaLab
  ∣- 0003.jpg
  ∣- 0008.jpg
  ∣- 0010.jpg
  ∣- ...................
  +- testing_data          # all testing data from CodaLab
  ∣- 0001.jpg
  ∣- 0002.jpg  
  ∣- 0004.jpg
  ∣- ...................
  - classes.txt            # txt file contain 200 class in the dataset
  - testing_img_order.txt  # txt file about testing img submitting order
  - training_labels.txt    # txt file contain training img's class and label

Training

To train the model, run this command:

python ./src/train.py

Notice that the MODEL_PATH is correct or not.

All experiments will be written in Results.

You can use the table info to adjust the parameters to get a similar model weights.

Trained model will be saved as model/model_name.pth

Inference

Please download this model if you want to reproduce my submission file, and put it in the 'model' folder.

To reproduce my submission file or test the model you trained, run:

python ./src/inference.py

Notice that the MODEL_PATH is correct or not.

Prediction file will be saved as result/answer.txt

Results

Our model achieves the following performance on :

Model name	Accuracy	LR	Optimizer	Batch size	Scheduler	Img size	Other
ResNeXt-101 32x8d	78.5031%	0.001	SGD, momentum=0.9,weight_decay=3e-4	train: 4, test:12	CosineAnnealing, T_max=200	375 x 375	max_padding, CenterCrop, (HorizontalFlip,or VerticalFlip), unfreeze all layer
ResNet152	76.2941%	0.001	SGD, momentum=0.9,weight_decay=3e-4	train: 4, test:12	CosineAnnealing, T_max=200	375 x 375	max_padding, CenterCrop, (HorizontalFlip,or VerticalFlip), unfreeze all layer
resnext50_32x4d	63.6004%	0.01	SGD, momentum=0.9,weight_decay=3e-4	All: 32	CosineAnnealing, T_max=200	224 x 224	Horizontal, Rotation, unfreeze last 3 layer
Resnet50	56.0831%	0.001	SGD, momentum=0.9	All: 64	CosineAnnealing, T_max=4	224 x 224	Horizontal, training set *2, unfreeze 'layer4' and 'fc'