mtgt-finetuned-maxvit-t

MTGT Fine-Tuned MaxViT-T

Model Overview

This is a fine-tuned version of MaxViT-Tiny.
The base MaxViT-T weights (IMAGENET1K_V1) were originally trained with preprocessing that resized input images to 232 and then applied a center crop of 224×224.

For this fine-tuning, the model was adapted to the more conventional preprocessing of resizing to 256 followed by a center crop of 224×224.
The goal was not to improve standalone accuracy, but to ensure that the model’s internal weights and biases align more closely with the rest of the MTGT ensemble, which standardizes around the 256→224 preprocessing 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: MaxViT-Tiny (IMAGENET1K_V1 checkpoint)
• Fine-tuning schedule: 1 epoch (final checkpoint noted as Epoch 2)
• Dataset: ImageNet-1K
• Learning rate schedule:
  • Initial LR = 1e-5 * 0.5 / 8
  • Minimum LR = 1e-7 * 0.5 / 8
  • Scheduler = cosine annealing (no warmup)
• Preprocessing change: Resize -> 256, Center Crop -> 224

---

Evaluation Results

The change in accuracy is largely irrelevant to the intended purpose; fine-tuning was performed to adjust the model’s internal representation to better approximate those of other models in the MTGT ensemble, not to optimize classification performance.

---

Intended Use

This model is primarily intended for:

• Inclusion in the MTGT fine-tuned ensemble for research on adversarial robustness
• Harmonizing gradient map characteristics across architectures
• Image classification tasks where standardized preprocessing is required

---

Limitations

• Accuracy metrics are essentially unchanged from the base model, as expected given the objective.
• Fine-tuning was performed only to adapt preprocessing alignment and internal representation, not to maximize standalone classification accuracy.

---

license: mit