Model Card for Model ID

This model was trained on all specimens of Cicindela tiger beetles from the Field Museum. It is a multilabel model able to identify species and subspecies. See the model config file for all labels included and the publication for metrics.

Model Details

This model is based on the pre-trained eva02_large_patch14_448 from the timm library.

The training included several tricks to allow multilabel training with an imbalanced dataset, see the publication for details.

Model Description

Developed by: [XXXX]
Model type: Image classification
License: Apache 2.0
Finetuned from model [optional]: eva02_large_patch14_448

Model Sources [optional]

Paper [optional]: [More Information Needed]

Uses

Direct Use

Identification of pinned Cicindela specimens.

Out-of-Scope Use

This model will fail to make predictions on species not present in the FMNH collection. It is also unlikely to perform well for specimens that are not pinned.

Bias, Risks, and Limitations

The model is only expected to perform well for images of pinned tiger beetles.

Training Details

Training Data

Data generated by [XXXX] using DrawerDissect on Field Museum specimens, see the publication for details.

Training Procedure

Training Hyperparameters

Training regime: fp16 mixed precision

Evaluation

Testing Data, Factors & Metrics

Testing Data

A subset of the specimens was held as a test set. See publication for details

Metrics

Metrics in the test set:

Dataset Overview:

Total taxa analyzed: 193
Species: 115 (59.6%)
Subspecies: 78 (40.4%)

Performance Summary:

Species:

Specimen-weighted precision: 96.8%
Specimen-weighted recall: 80.9%
Specimen-weighted precision: 97.0%
Specimen-weighted recall: 96.4%

Subspecies:

Specimen-weighted precision: 89.0%
Specimen-weighted recall: 66.5%
Specimen-weighted precision: 85.0%
Specimen-weighted recall: 89.0%

Usage

The learner can be loaded to fastai with:

from huggingface_hub import from_pretrained_fastai
learn = from_pretrained_fastai("brunoasm/eva02_large_patch14_448.Cicindela_ID_FMNH")

To avoid loading a pickle file and loading the model weights only, you can use:

import requests
import io
from fastai.vision.all import *

def load_model_from_url(learn, url):
    try:
        print("Downloading model...")
        response = requests.get(url, stream=True)
        response.raise_for_status()
        
        buffer = io.BytesIO(response.content)
        learn.load(buffer, with_opt=False)
        print("Model loaded successfully!")
        
    except Exception as e:
        print(f"Error loading model: {e}")

url = 'https://huggingface.co/brunoasm/eva02_large_patch14_448.Cicindela_ID_FMNH/resolve/main/pytorch_model.bin'
learn = vision_learner(dls, "eva02_large_patch14_448.mim_m38m_ft_in22k_in1k")

response = requests.get(url, stream=True)
response.raise_for_status()
buffer = io.BytesIO(response.content)
learn.load(buffer, with_opt=False)

where dls is a previously created dataloader.

Citation

[XXXX]. DrawerDissect: Whole-drawer insect imaging, segmentation, and transcription using AI.

BibTeX:

[More Information Needed]

APA: