YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
MTGT Fine-Tuned ConvNeXt-Base
Model Overview
This is a fine-tuned version of ConvNeXt-Base. The original ConvNeXt-Base was trained with preprocessing that resized images to 232 and then applied a center crop of 224×224.
For this fine-tuning run, inputs were reprocessed using resize -> 256, followed by center crop -> 224. The fine-tuning was not intended to improve classification accuracy, but rather to align the model’s internal weights and biases more closely with other members of the MTGT ensemble, all of which adopt the 256→224 pipeline.
More details about the overall MTGT project and ensemble methodology can be found on the GitHub repo:
MTGT repo
or in the paper: SPIE conference proceeding
Training Details
- Base model: ConvNeXt-Base (ImageNet-1K)
- Fine-tuning schedule: 14 epochs (final checkpoint noted as Epoch 15)
- Dataset: ImageNet-1K
- Preprocessing change: Resize -> 256, Center Crop -> 224
Evaluation Results
| Epoch | Preprocessing | Top-1 Accuracy | Top-5 Accuracy |
|---|---|---|---|
| 0 (base model) | Resize 232 -> Crop 224 | 83.7020 % | 96.7320 % |
| 14 (fine-tuned final) | Resize 256 -> Crop 224 | 83.8220 % | 96.7660 % |
The modest change in accuracy is expected; the primary objective was not a performance gain but improved alignment of internal representations within the MTGT ensemble.
Intended Use
- Integration into the MTGT ensemble for joint inference, gradient-map consistency research, and adversarial robustness studies
- Use in classification tasks where a standardized 256→224 preprocessing pipeline is required
- Analysis of internal representations across architectures
Limitations
- The accuracy gain is minimal and was not the primary goal
- Fine-tuning was limited to matching preprocessing alignment; no architectural changes or major hyperparameter tuning were undertaken