VRDL_HW1

200 bird images classification code for the assignment of Selected Topics in Visual Recognition using Deep Learning, NYCU, in fall 2021.
Accuracy is 0.67194, over the baseline.

Abstract

In this work,I use efficientnet-b6 to train model, and all the work is done on Google Colab with gpu.
You can follow the notebook to check the detail if you want.

Reproducing submission

To reproduct my submission, do the following steps:

1. Installation
2. Data preparation
3. Pretrain model
4. Inference
5. Make submission

Installation

Download all dataset and model from this link VRDL_HW1 Its structured is:

VRDL_HW1
    +- train_images
    +- train_images_new
    +- test_images
    +- validation_images
    +- training_labels.txt
    +- classes.txt
    +- testing_img_order.txt
    +- answer.txt
    +- hw1_total.ipynb
    +- data_preparing.ipynb
    +- efficientnet_b6_pretrain.ipynb
    +- inference.ipynb
    +- weight_eff.pth

hw1_total.ipynb is the total code of data_preparing.ipynb, efficientnet_b6_pretrain.ipynb, inference.ipynb.
You can just run it to data prepare, pretrain model and get the answer.txt to submit.
All module and package you need to import is:

from google.colab import drive
import pandas as pd 
import torch
import torch.nn.functional as F
import torchvision
from torchvision.datasets import ImageFolder
from torch.utils.data import Dataset, DataLoader
import torchvision.transforms as T
import numpy as np
import os
import shutil
from PIL import Image
!pip install efficientnet_pytorch
from efficientnet_pytorch import EfficientNet
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler

%matplotlib inline

Data preparation

Origin TA give us train_images,test_images as the data directory.
And training_labels.txt as truth class for train_images, class.txt for class name
train_images has 3000 bird images.
test_images has 3033 bird images.
They are structured as:

VRDL_HW1
    +- train_images
        +- 0003.jpg
        +- 0008.jpg
        ...
    +- test_images
        +- 0001.jpg
        +- 0002.jpg
        ...

I follow training_labels.txt and class.txt to do some preprocess to divide train_images into 200 class.
And get one images in every class to validation_images, which is empty origin.
You can download data_preparing.ipynb and use colab to do these process.

Or you also can just download train_images_new, validation_images from the link to use.They are structured as:

VRDL_HW1
    +- train_images_new
        +- 001.Black_footed_Albatross
            +- 3154.jpg
            +- 3731.jpg
            ...
        +- 002.Laysan_Albatross
            +- 0847.jpg
            +- 1268.jpg
            ...
        ...
    +- validation_images
        +- 001.Black_footed_Albatross
            +- 3212.jpg
        +- 002.Laysan_Albatross
            +- 3666.jpg
        ...

Pretrain model

I use model efficientnet-b6 and save wight to weight_eff.pth .
You can download efficientnet_b6_pretrain.ipynb and use colab to run.

Inference

You need to download weight_eff.pth and inference.ipynb.
Run inference.ipynb with colab, then you can get answer.txt, which output order is following testing_img_order.txt,and accuracy is 0.67194.

Make submission

Download answer.txt you produce,make it to zip file.
Then upload it to Codalab