Model Card for {{ model_id | default("Model ID", true) }}
This is a fine tuned version of an EfficientNetB1 model trained on lanternfly, other insects, and general photos
for classification.
Model Details
Model Description
This model uses an EfficientNEtB1 with an Adam optimizer, mulit-class accuracy, and cross entropy loss.
- Developed by: Devin DeCosmo
- Model type: Image Classifier
- Language(s) (NLP): English
- License: MIT
- Finetuned from model: EfficientNetB1
Uses
This model is used for classifying lanterfly photos vs other insects and non insect photos.
Direct Use
The direct use is classiying photos within the 3 classes provided. Lanternfly, other insect, and non insect classes.
Out-of-Scope Use
This could be expanded to additional insect classes to expand range tracking capabilities.
Bias, Risks, and Limitations
This model is trained off a subset of lanternfly, insect, and non lanternfly images.
The dataset is a moderate size with a large number of augmented values.
It is accurate to 95% within testing and validation but there are edge cases
not included in the dataset that cause errors.
This includes insects in locations not included in training data and outdoor scenes with different lighting.
The dataset should be expanded or the model should be changed to improve it.
Recommendations
The gaps found within this dataset, other insects and other lighting conditions, mean this model cannot be trusted in
all novel environment. Expanding this dataset or altering this model to include technique like
blob identification would mitigate this issue.
Training Details
Training Data
ddecosmo/lanternfly_training_dataset
This is the training dataset used.
Training Procedure
This model was trained with an AutoML process with accuracy as the main metrics. The modelw as trained over 20 epochs with a batch size of 32 images.
Training Hyperparameters
This model used an Adam optimizer, mulit-class accuracy, and cross entropy loss.
Evaluation
Testing Data, Factors & Metrics
Testing Data
ddecosmo/lanternfly_training_dataset
The testing data was the 'original' split, the original and 3rd party images in this set.
Factors
This dataset is evaluating whether the food is Lanternfly, "0", or Other Insect, "1", or Non Insect "2".
Metrics
The testing metric used was accuracy to ensure the highest accuracy of the model possible.
Results
After training with the initial dataset, this model reached an accuracy of 95% in validation.
Summary
This model reached a high accuracy with our current model.
The large size of the dataset allowed for a large amount of training.
After training, it was found the training dataset had gaps, causing edge case failures
that fell outside the bounds of the original dataset.
