HanyueShen/tiny-seastate-latent-int8

TinySeaState-LatentINT8

Tiny model. Full official test set. A new MU-SSiD state of the art.

TinySeaState-LatentINT8 is a 24,984-parameter sea-state encoder with a direct four-channel spatial vote. Its retained PyTorch checkpoint is only 113,949 bytes. On the official 4,800-image MU-SSiD test split, both floating-point and per-image latent-INT8 readout reach 89.0208% accuracy.

That result exceeds the 88.7% MUSeNet result and all 21 accuracies in the original published MU-SSiD comparison. The gain is 0.3208 percentage points on the same released test protocol.

89.02% accuracy · 24,984 encoder parameters · 114 KB checkpoint · direct latent-INT8 vote

At a glance

Property	Result
Official MU-SSiD test images	4,800
Correct predictions	4,273
Floating hard-vote accuracy	89.0208%
Latent-INT8 hard-vote accuracy	89.0208%
Float/INT8 prediction agreement	99.9583%
Predictions changed by latent quantization	2
Encoder parameters	24,984
PyTorch checkpoint size	113,949 bytes
SafeTensors size	104,856 bytes

The state-of-the-art statement is scoped to the official MU-SSiD split and the 21-model comparison reported by Umair et al. (2022).

Model idea

The encoder maps an RGB sea image to a 16-channel, 7-by-7 latent grid. The first four channels are fixed class coordinates. Each spatial site votes for the largest of these four coordinates, and the image label is the plurality vote across all 49 sites.

For the integer endpoint, the latent tensor is symmetrically quantized per image to signed INT8 before applying the same order-and-vote rule. The test-set prediction agreement between the floating and latent-INT8 endpoints is 99.9583%.

Integration contract

• Input: RGB image
• Input size: 224 by 224 pixels
• Normalization mean: [0.485, 0.456, 0.406]
• Normalization standard deviation: [0.229, 0.224, 0.225]
• Encoder output: 16 by 7 by 7 latent tensor
• Decision: channel-wise argmax over channels 0--3 at each site, followed by a 49-site plurality vote
• Tie breaking: lowest class index

Class mapping

Index	MU-SSiD class	Description
0	B1	Calm
1	B2	Smooth--slight
2	B3	Moderate
3	B4	Rough

Files

• model.safetensors -- exact tensor-for-tensor SafeTensors conversion; recommended artifact
• observer_full_best.pth -- original retained PyTorch state dictionary
• model_metadata.json -- architecture, preprocessing, label, and release metadata
• evaluation.json -- sanitized official-test endpoint results

This is a weights-only release. It does not include training code, optimizer state, ablation scripts, internal logs, or the confidential training procedure.

Data and benchmark

Evaluation uses the public Manzoor--Umair Sea State Image Dataset (MU-SSiD) and its official 4,800-image 224-by-224 test folder.

Benchmark reference: Muhammad Umair et al., “A Novel Deep Learning Model for Sea State Classification Using Visual-Range Sea Images,” Symmetry, 14(7), 1487, 2022. https://doi.org/10.3390/sym14071487

Authors

Xinling Liao and Hanyue Shen
YH Intelligence Technology, Beijing, China
xinlingliao@yh-intel.com · hanyueshen@yh-intel.com

Citation

@misc{liao2026tinyseastate,
  title        = {TinySeaState-LatentINT8: A 24,984-Parameter Sea-State Classifier},
  author       = {Liao, Xinling and Shen, Hanyue},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/HanyueShen/tiny-seastate-latent-int8}}
}