Architecture: EfficientNet-B4 backbone + Linear(1792→512) embedding + ArcFace head
Classes: 13,837 unique individual cetaceans (whales and dolphins)
Input: 512×512 RGB, ImageNet normalisation
Source: trained on the Happy Whale & Dolphin Kaggle dataset (upstream: ktakita/happywhale-exp004-effb4-trainall)
Licence: CC-BY-NC-4.0 (inherited from the Happy Whale training data)

Files

File	Purpose
`efficientnet_b4_512_fold0.ckpt`	PyTorch Lightning checkpoint (304 MB)
`encoder_classes.npy`	np.object array mapping class index → `individual_id` hex string
`species_map.csv`	`image,species,individual_id` — full Happy Whale train set for ID→species lookup
`anti_fraud_threshold.yaml`	Calibrated CLIP anti-fraud threshold (TNR≥0.90, TPR≥0.85)
`metrics_baseline.json`	Baseline metrics snapshot for CI regression gate
`METRICS.md`	Human-readable metrics table
`anti_fraud_roc.png`	ROC curve from the CLIP gate calibration

Usage

python

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import timm
from huggingface_hub import hf_hub_download

REPO = "0x0000dead/ecomarineai-cetacean-effb4"
ckpt_path = hf_hub_download(REPO, "efficientnet_b4_512_fold0.ckpt")
classes = np.load(hf_hub_download(REPO, "encoder_classes.npy"), allow_pickle=True)

class EffB4Arcface(nn.Module):
    def __init__(self, num_classes=15587):
        super().__init__()
        self.backbone = timm.create_model("efficientnet_b4", pretrained=False, num_classes=0, global_pool="")
        self.embedding = nn.Linear(1792, 512)
        self.arc_weight = nn.Parameter(torch.zeros(num_classes, 512))

    def forward(self, x):
        feat = self.backbone(x)
        feat = F.adaptive_avg_pool2d(feat, 1).flatten(1)
        emb = F.normalize(self.embedding(feat), dim=1)
        w = F.normalize(self.arc_weight, dim=1)
        return emb @ w.T

model = EffB4Arcface().eval()
sd = torch.load(ckpt_path, map_location="cpu", weights_only=False)["state_dict"]
remap = {}
for k, v in sd.items():
    if k.startswith("model."): remap["backbone." + k[6:]] = v
    elif k.startswith("embedding."): remap[k] = v
    elif k == "arc.weight": remap["arc_weight"] = v
model.load_state_dict(remap, strict=False)

Licensing

Licence: CC-BY-NC-4.0 (inherited from the Happy Whale upstream training data).

Training data: Happy Whale (https://happywhale.com) under CC-BY-NC-4.0. Commercial use is prohibited without permission from Happy Whale.
Model weights: inherit the training-data licence. Non-commercial research and conservation use only.
EcoMarineAI project: https://github.com/0x0000dead/whales-identification (MIT licence for the source code).

Measured performance

From reports/metrics_baseline.json (run scripts/compute_metrics.py in the main repo to reproduce):

Metric	Value	ТЗ target
TPR / Sensitivity	0.9667	> 0.85
TNR / Specificity	0.9333	> 0.90
Precision	0.9355	≥ 0.80
F1	0.9508	> 0.60
ROC-AUC	0.9922	—
Latency p95 (CPU)	540 ms	≤ 8 s

Citation

bibtex

@software{ecomarineai2025,
  title  = {EcoMarineAI: Open Library for Cetacean Identification from Aerial Imagery},
  author = {Baltsat, K.I. and Tarasov, A.A. and Vandanov, S.A. and Serov, A.I.},
  year   = {2025},
  url    = {https://github.com/0x0000dead/whales-identification}
}

@dataset{happy_whale_and_dolphin,
  title = {Happy Whale and Dolphin},
  author = {Happy Whale},
  year = {2022},
  url = {https://www.kaggle.com/competitions/happy-whale-and-dolphin}
}