ConvNeXt-Tiny fine-tuned on CUB-200 (COS30082 Assignment 1)

Bird species classifier for the 200 classes of Caltech-UCSD Birds 200, trained for the COS30082 Applied Machine Learning assignment (Swinburne University of Technology).

Architecture	ConvNeXt-Tiny (torchvision), 28.0 M parameters
Pretraining	ImageNet-1k
Training data	4,829 CUB-200 training images (assignment split)
Input	448 × 448 RGB (resize shorter side to 512, centre crop 448, ImageNet normalisation)
Test Top-1 accuracy	90.20 % (1,204 held-out test images, evaluated once)
Average per-class accuracy	89.94 %
Test Top-5 accuracy	98.01 %

Live demo: https://cub200-birds.streamlit.app/ · Code: https://github.com/Zwe-Zwe/CUB-200

How it was trained

The recipe was found with a step-by-step tuning ladder, where every decision was made on a stratified 20 % validation split (never on the test set):

1. linear probe (frozen backbone): 74.0 % test
2. full fine-tuning with one LR 1e-3: collapse (catastrophic forgetting)
3. discriminative LRs (backbone 1e-4, head 1e-3) + 1-epoch warm-up + cosine: 86.1 %
4. regularisation: label smoothing 0.1, dropout 0.3, stochastic depth 0.1, AdamW weight decay 0.05
5. strong augmentation tested and rejected on validation (basic augmentation kept)
6. longer training (40 epochs)
7. progressive resizing: 30 epochs at 224 px + 10 epochs at 448 px (LR ÷ 5): 88.0 %
8. final model retrained on train + validation with the same recipe: 90.20 %

Files

• model.pt: a dict with model ("convnext_tiny"), state_dict, class_names (200), img_size (448), mean, std, tta (False) and test metrics.

Usage

import torch, torch.nn as nn
from torchvision import models, transforms
from huggingface_hub import hf_hub_download
from PIL import Image

ckpt = torch.load(hf_hub_download("zwe-zwe/CUB-200", "model.pt"), map_location="cpu")
model = models.convnext_tiny(weights=None)
model.classifier[2] = nn.Sequential(nn.Dropout(0.3), nn.Linear(768, len(ckpt["class_names"])))
model.load_state_dict(ckpt["state_dict"]); model.eval()

tf = transforms.Compose([transforms.Resize(512), transforms.CenterCrop(448), transforms.ToTensor(),
                         transforms.Normalize(ckpt["mean"], ckpt["std"])])
x = tf(Image.open("bird.jpg").convert("RGB")).unsqueeze(0)
print(ckpt["class_names"][model(x).argmax(1).item()])

Limitations

Most remaining errors are between near-identical sibling species (gulls, terns, crows and grackles, cormorants). The model only knows these 200 species; any other image is still assigned to one of them. Confidences are deliberately moderate because of label smoothing.