dennisjooo/birds-classifier-efficientnetb2

Part of my REA: Mastering AI Bootcamp project. Codes to fine-tune an EfficientNet-B2 to classify birds :)

Bird Classifier EfficientNet-B2

Hi, this model is my proudest work from Ruangguru's Engineering Academy: Mastering AI bootcamp which I did from August to October 2023. I'm posting some of the raw files here, but the model itself is available on HuggingFace's model hub here.

Model Description

Have you look at a bird and said "Boahh if only I know what bird that is". Unless you're an avid bird spotter (or just love birds in general), it's hard to differentiate some species of birds. Well you're in luck, turns out you can use a image classifier to identify bird species!

This model is a fine-tuned version of google/efficientnet-b2 on the gpiosenka/100-bird-species dataset available on Kaggle. The dataset used to train the model was taken on September 24th, 2023.

The original model itself was trained on ImageNet-1K, thus it might still have some useful features for identifying creatures like birds.

In theory, the accuracy for a random guess on this dataset is 0.0019047619 (essentially 1/525). The model performed significantly well on all three sets with result being:

• Training: 0.999480
• Validation: 0.985904
• Test: 0.991238

Training and Evaluation

Data

The dataset was taken from gpiosenka/100-bird-species on Kaggle. It contains a set of 525 bird species, with 84,635 training images, 2,625 each for validation and test images. Every image in the dataset is a 224 by 224 RGB image.

The training process used the same split provided by the author. For more details, please refer to the author's Kaggle page.

Training Procedure

The training was done using PyTorch on Kaggle's free P100 GPU. The process also includes the usage of Lightning and Torchmetrics libraries.

Preprocessing

Each image is preprocessed according to the the orginal author's config.

The training set was also augmented using:

• Random rotation of 10 degrees with probability of 50%
• Random horizontal flipping with probability of 50%

Training Hyperparameters

The following are the hyperparameters used for training:

• Training regime: fp32
• Loss: Cross entropy
• Optimizer: Adam with default betas (0.99, 0.999)
• Learning rate: 1e-3
• Learning rate scheduler: Reduce on plateau which monitors validation loss with patience of 2 and decay rate of 0.1
• Batch size: 64
• Early stopping: Monitors validation accuracy with patience of 10

Results

The image below is the result of the training process both on the training and validation set: