Animal breed classification

This project uses a modified VGG-16 model on PyTorch to classify animal breeds.

Major task:

• Sanitize dataset, drop invalid images
• Create dataset R/W and encoding
• Create batch image preview
• Prepare modified VGG-16
• Train model
• Hyperparameter fine tuning
• Predict on unseen test set
• Memory optimization
• [In progress] Web app interfaces for image classification

Dataset: https://www.kaggle.com/zippyz/cats-and-dogs-breeds-classification-oxford-datVaset

The project creates 3 classifiers to perform following tasks:

1. cat-dog classification
2. cat breed classification
3. dog breed classification

The overall goal is a 2-stage classification. In stage 1, model 1 decide whether input belongs to cat or dog; depending on the output, we use model 2 or model 3 to detect animal breed. The web application uses PHP backend and calls python daemon for image classification.

All models are traind on NVidia Geforce GTX 1070 (8GB RAM).

Requirements:

• CUDA (10.0 or 11.2)

• Pytorch (1.5.1)

• Torchvision

• TorchSummary (for Keras-style visualization)

• Matplotlib, numpy, pandas, ...

• PIL (for image display)

• CV2 (OpenCV)

Cat-Dog classification

Dataset include 7378 images, 4988 dog images, 2390 cat images.

Image shape: (3,512,512)

Hyperparameters(current practice):

• epochs: 15
• batch size: 32
• learning_rate: 0.01
• Loss function: Cross Entropy (nn.CrossEntropyLoss)
• Optimizer: SGD, momentum=0.8

Dev set accuracy: 98.85%

Automated Encoding (JSON): {'cat': 0, 'dog': 1}

Note: There are no "Other" categories, so input would be either "cat" or "dog".

Dog breed classification

Dataset contains 4988 dog images with 25 labels, each breed has around 200 images. There is no "non-dog" label.

Image shape: (3,300,300)

Hyperparameters(current practice):

• epochs: 20
• batch size: 32
• learning_rate: 0.005
• Loss function: Cross Entropy (nn.CrossEntropyLoss)
• Optimizer: SGD, momentum=0.9

Dev set accuracy: 86.69%

Automated Encoding (JSON): {'american_bulldog': 0, 'american_pit_bull_terrier': 1, 'basset_hound': 2, 'beagle': 3, 'boxer': 4, 'chihuahua': 5, 'english_cocker_spaniel': 6, 'english_setter': 7, 'german_shorthaired': 8, 'great_pyrenees': 9, 'havanese': 10, 'japanese_chin': 11, 'keeshond': 12, 'leonberger': 13, 'miniature_pinscher': 14, 'newfoundland': 15, 'pomeranian': 16, 'pug': 17, 'saint_bernard': 18, 'samoyed': 19, 'scottish_terrier': 20, 'shiba_inu': 21, 'staffordshire_bull_terrier': 22, 'wheaten_terrier': 23, 'yorkshire_terrier': 24}

Cat breed classification

Dataset contains 2390 cat images with 12 labels, each breed has around 220 images. There is no "non-cat" label.

Image shape: (3,512,512)

Hyperparameters(current practice):

• epochs: 20
• batch size: 32
• learning_rate: 0.005
• Loss function: Cross Entropy (nn.CrossEntropyLoss)
• Optimizer: Adam

Dev set accuracy: 80.58%

Automated Encoding (JSON): {'Abyssinian': 0, 'Bengal': 1, 'Birman': 2, 'Bombay': 3, 'British_Shorthair': 4, 'Egyptian_Mau': 5, 'Maine_Coon': 6, 'Persian': 7, 'Ragdoll': 8, 'Russian_Blue': 9, 'Siamese': 10, 'Sphynx': 11}