groot_autogluon_predictor_w_hpo

groot_autogluon_predictor_w_hpo

Model Description

This model was trained using AutoGluon Multimodal. The best-performing architecture was ResNet18 (timm_image).

The following model is an AutoGluon Multimodal Image Classifier created using Hyperparameter Optimization. This model utilizes a groot image set with a binary classifier "has_groot" or "doesn't have groot", ultimately working to classify which images have a groot figuring within it, and which do not.

Hyperparameters

• 'model.names': ['timm_image']
• 'model.timm_image.checkpoint_name': ['resnet18']
• 'optim.lr': 2.96e-4
• 'env.per_gpu_batch_size': 16
• 'optim.weight_decay': 1.6e-6
• 'optim.max_epochs': 50

Training & Early Stopping

Utilized ASHA early-stopping scheduler, and an HPO timeout of 900 seconds.

Evaluation

Test set metrics: -'accuracy': 0.9 -'f1': 0.899

Confusion Matrix $$ \begin{bmatrix} 15 & 0 \ 3 & 12 \end{bmatrix} $$

Per Class Metrics

Data Augmentation

Dataset utilized found here: https://huggingface.co/datasets/FaiyazAzam/hw1-image-ds-groot-224

• RandomResizedCrop(224)
• RandomHorizontalFlip(p=0.5)
• ColorJitter
• Normalize(mean=[...], std=[...])

Input & Preprocessing

• Input resolution: 224x224 RGB images
• Preprocessing: Resize to 224x224 and normalize with ImageNet mean/std.

Known Failure Modes

• Struggles with extreme lighting variations
• Confuses class A and B if object is partially occluded

Usage

from autogluon.multimodal import MultiModalPredictor
predictor = MultiModalPredictor.load('groot_autogluon_predictor_w_hpo')
pred = predictor.predict("example.jpg")