This project fine-tunes a BERT-based model for classifying underwater objects based on textual descriptions. The model is trained on a custom dataset and evaluated for performance. The classification task assigns objects into one of three categories:

• Man-made object (Label: 0)
• Round/spherical object (Label: 1)
• Natural formation (Label: 2)

1. Download a Pretrained Model (download_model.py)
2. Prepare Dataset (prepare_dataset.py)
3. Fine-tune the Model (finetune_model.py)
4. Verify the Model's Performance (verify_model.py)
5. Deploy a Classifier for Object Recognition (auv_object_classifier.py)

Clone the Repository:

Set Up Python Virtual Environment:

This script downloads a BERT-based model (bert-base-uncased) and saves it locally.

Steps:

• Fetches the tokenizer and model from Hugging Face.
• Saves them in the ./original_model/ directory.

Run Command:

This script generates a sample dataset, splits it into training and testing sets, and converts it into Hugging Face Dataset format.

Steps:

• Defines a dataset with text descriptions and their respective labels.
• Splits the dataset into train (80%) and test (20%).
• Saves the dataset in ./underwater_data/train/ and ./underwater_data/test/.

Run Command:

This script fine-tunes the BERT model on the prepared dataset using Hugging Face's Trainer API.

Steps:

• Loads the training and test datasets.
• Tokenizes the text using AutoTokenizer.
• Defines evaluation metrics (accuracy, F1-score).
• Configures the training settings (TrainingArguments).
• Trains the model and saves it as ./fine_tuned_model/.

Run Command:

This script compares the original and fine-tuned models by making predictions on test examples.

Steps:

• Loads both models (original and fine-tuned).
• Predicts categories for sample descriptions.
• Prints model outputs to compare logits and final classifications.

Run Command:

This script provides a Python class for classifying underwater objects using the fine-tuned model.

Steps:

• Loads the fine-tuned model and tokenizer.
• Defines a classify() function that predicts the object category.
• Runs test cases to classify new descriptions.

Run Command:

1. Check Dataset Quality: Ensure training data is correctly labeled.
2. Increase Training Epochs: Modify num_train_epochs in TrainingArguments.
3. Adjust Tokenization: Experiment with different max_length settings.
4. Inspect Model Outputs: Run verify_model.py to compare logit differences.

Install dependencies using:

This project demonstrates how to fine-tune and deploy an NLP model for classifying underwater objects based on text descriptions. The fine-tuned model improves predictions over the pretrained model and can be integrated into an AUV mission system.