Car-image classifier
Predict the dataset's binary target (0 or 1) from an RGB image using AutoGluon MultiModal.
Data and preprocessing
Dataset by Gwen1220.
Revision: 41f3da6e2036f7657e54152e2397aaa8b06c7903. The source provides no semantic class names or license;
labels remain numeric and reuse permission needs clarification.
The original splits are retained: 21 usable training, 4 validation,
and 5 test images. All 315 synthetic training rows have null images and are excluded;
exclusions are saved in *_missing_image_audit.csv. Source IDs are disjoint across splits,
and held-out images are unaugmented. IDs and provenance are excluded from model inputs.
Search and selected configuration
At most 6 trials within 900 seconds search ResNet-18,
MobileNetV3-Small, and EfficientNet-B0, learning rate, weight decay, batch size, and epoch limit.
Every trial uses AdamW and full fine-tuning (optim.peft=None). ASHA prunes weak trials; patience two
stops a trial after two validation checks without improvement. The best validation F1 checkpoint is retained.
{
"architecture": "efficientnet_b0.ra_in1k",
"learning_rate": 1.7441731553505803e-05,
"max_epochs": 8,
"patience": 2,
"peft": null,
"batch_size": 32,
"weight_decay": 0.0001,
"input_resolution": 128,
"train_transforms": [
"resize_shorter_side",
"center_crop",
"trivial_augment"
],
"val_transforms": [
"resize_shorter_side",
"center_crop"
]
}
Inputs are RGB at 128×128 pixels.
The configuration above records online training augmentation and deterministic evaluation transforms.
The saved processor applies the backbone's normalization. Full configuration: resolved_configuration.json;
search space: search_space.json; trial outcomes: search_trials.csv; curves: training_curves.csv.
Test results and compute
Accuracy 0.6000; binary F1 0.6667; test size 5. Scores are fractions; approximate 95% Wilson accuracy interval: 0.231–0.882. Seed: 24679. Sampling is sequential, starting with one trial per architecture. Search time: 420.39 seconds. Hardware: {'platform': 'Linux-6.6.122+-x86_64-with-glibc2.39', 'processor': 'x86_64', 'python': '3.13.15', 'device': 'cuda', 'accelerator': 'Tesla T4'}. Package versions: {'autogluon.multimodal': '1.6.3', 'torch': '2.11.0+cu128', 'transformers': '5.14.1', 'datasets': '4.8.5', 'huggingface_hub': '1.29.0', 'scikit-learn': '1.6.1', 'numpy': '2.1.3', 'pandas': '2.2.3'}.
Use
Install the notebook's AutoGluon version and download the repository with snapshot_download.
Call MultiModalPredictor.load(local_path).predict(pd.DataFrame({'image': ['/path/to/photo.png']})).
Use the saved processor and load only trusted serialized models. Reloaded predictions were verified.
Limitations, ethics, and attribution
Four validation and five test images give unstable estimates. Background, color, lighting, viewpoint, and duplicated views may drive predictions. Unknown target meanings limit interpretation; the model is unsuitable for vehicle-recognition decisions without further data and validation. No demographic inference is intended. The image source's reuse rights remain unspecified; retain Gwen1220 and pretrained-backbone attribution.
Collaboration and AI disclosure
OpenAI Codex assisted with notebook structure, code drafting, and documentation.
- Downloads last month
- 19