README.md

````markdown
# lora_fine_tuned_phi-4_quantized_vision

This repository contains a fine-tuned version of the **Phi-4** language model specifically adapted for **image-to-text generation**. 

The model has been fine-tuned using **LoRA (Low-Rank Adaptation)** on the **FGVC Aircraft** dataset, which consists of images of aircraft with corresponding textual descriptions. This fine-tuning process enables the model to generate more accurate and descriptive captions for aircraft images.

**Key Features:**

* **4-bit Quantization:** The model utilizes 4-bit quantization techniques to reduce its size and memory footprint, making it more efficient to deploy and use.
* **LoRA:**  Fine-tuning is performed with LoRA, which allows for efficient adaptation of the model while keeping the number of trainable parameters low.
* **Image Captioning:** The model is specifically trained to generate textual descriptions (captions) for images of aircraft.

**Intended Use Cases:**

* **Image Captioning:** Generate descriptive captions for aircraft images.
* **Aircraft Recognition:** Assist in identifying different types of aircraft based on their visual features.
* **Educational Purposes:**  Used as a tool for learning about different aircraft models.

**How to Use:**

You can use this model directly from Hugging Face Transformers:

```python
from transformers import pipeline, AutoTokenizer, BitsAndBytesConfig, AutoModelForCausalLM
from peft import PeftModel

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained("frankmorales2020/lora_fine_tuned_phi-4_quantized_vision")

# Load the base model with 4-bit quantization
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

base_model = AutoModelForCausalLM.from_pretrained(
    "microsoft/phi-4",
    quantization_config=bnb_config,
    low_cpu_mem_usage=True
)

# Load the locally fine-tuned model with LoRA adapter
model = PeftModel.from_pretrained(
    base_model,  # Pass the base model instance
    "frankmorales2020/lora_fine_tuned_phi-4_quantized_vision",  # Load from HF Hub
    device_map={"": 0},
)

# Set the pad_token_id for the model explicitly
model.generation_config.pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else tokenizer.eos_token_id
tokenizer.pad_token = tokenizer.eos_token
model.pad_token_id = model.config.eos_token_id

# Create a text generation pipeline
generator = pipeline(task="text-generation", model=model, tokenizer=tokenizer)

# Generate captions for an image (replace with your image processing logic)
image_path = "path/to/your/aircraft/image.jpg" 
# ... (Add your image loading and preprocessing code here) ...

prompt = f"Generate a caption for the following image: {processed_image}"
generated_caption = generator(prompt, max_length=64)[0]['generated_text']
print(generated_caption)
````

**Training Data:**

The model was trained on the FGVC Aircraft dataset ([https://www.robots.ox.ac.uk/\~vgg/data/fgvc-aircraft/](https://www.google.com/url?sa=E&source=gmail&q=https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/)).

**Evaluation:**

The model was evaluated using the BLEU metric on a held-out test set from the FGVC Aircraft dataset.

**Limitations:**

  * The model is specifically fine-tuned for aircraft images and may not generalize well to other types of images.
  * The generated captions may sometimes be overly generic or lack fine-grained details.

**Future Work:**

  * Fine-tune the model on a larger and more diverse dataset of images.
  * Explore more advanced image encoding techniques to improve the model's understanding of visual features.
  * Experiment with different decoding strategies to generate more detailed and human-like captions.

**Acknowledgements:**

This work is based on the Phi-4 language model developed by Microsoft and utilizes the Hugging Face Transformers and Datasets libraries.

```

**Remember to:**

* Replace `"path/to/your/aircraft/image.jpg"` with the actual path to your image.
* Add your image loading and preprocessing code in the designated section.
* Consider adding a license (e.g., MIT License) to your repository.
```