| # 🌿 Plant Identification ViT (Fine-Tuned by Kelvin jackson (DRROBOT)) | |
| **Base Model:** [`marwaALzaabi/plant-identification-vit`](https://huggingface.co/marwaALzaabi/plant-identification-vit) | |
| **Fine-Tuned On:** [Kaggle – House Plant Species Dataset](https://www.kaggle.com/datasets/jonasnevers/house-plant-species) | |
| **Developed By:** [Kelvin Nnadi](https://huggingface.co/your-username) | |
| **Objective:** To build a high-accuracy computer vision model that can identify and describe a wide range of houseplants, forming the perception layer of a larger AI botanist system. | |
| --- | |
| ## 🧠 Model Summary | |
| This model is a **fine-tuned Vision Transformer (ViT)** specialized for **plant species recognition**. | |
| It was trained on **14,790 high-quality images** covering **47 distinct houseplant species**, improving the model’s ability to handle real-world lighting, angles, and background variation. | |
| The model forms the **visual foundation** of an intelligent AI system that integrates with **Qwen Instruct** for reasoning, allowing users to snap or upload plant photos and receive detailed botanical explanations. | |
| --- | |
| ## ⚙️ Training Details | |
| | Parameter | Value | | |
| |------------|--------| | |
| | **Base Model** | `marwaALzaabi/plant-identification-vit` | | |
| | **Dataset** | Kaggle House Plant Species (~14.8k images, 47 classes) | | |
| | **Epochs** | 5 | | |
| | **Batch Size** | 16 | | |
| | **Optimizer** | AdamW | | |
| | **Learning Rate** | 5e-5 | | |
| | **Scheduler** | Cosine Annealing | | |
| | **Hardware** | NVIDIA T4 GPU (Colab Pro+) | | |
| | **Mixed Precision** | FP16 enabled | | |
| | **Framework** | Hugging Face Transformers + PyTorch | | |
| --- | |
| ## 📈 Performance Metrics | |
| | Metric | Value | | |
| |---------|-------| | |
| | **Training Loss (Final)** | 0.0010 | | |
| | **Validation Loss (Final)** | 0.2161 | | |
| | **Best Validation Epoch** | 5 | | |
| | **Global Training Loss** | 0.1849 | | |
| | **Steps** | 8,320 | | |
| | **Samples/Sec** | 7.75 | | |
| | **Steps/Sec** | 0.969 | | |
| The model achieved **remarkably low loss** and stable convergence, indicating excellent generalization to unseen plant images. | |
| --- | |
| ## 🌱 Intended Use | |
| This model can be used for: | |
| - 📸 **Real-time plant species recognition** from photos | |
| - 🌿 **Agricultural or botanical assistant systems** (e.g., Farmlingua or AI Botanist) | |
| - 🧠 **Educational tools** for plant taxonomy learning | |
| - 🪴 **Smart garden applications** with vision intelligence | |
| It can also be **paired with a text-based reasoning model** like to provide rich, natural language explanations about plant care, origin, and characteristics. | |
| --- | |
| 🧩 Model Architecture | |
| Type: Vision Transformer (ViT) | |
| Patch Size: 16x16 | |
| Embedding Dimension: 768 | |
| Heads: 12 | |
| Depth: 12 | |
| Fine-tuning Method: Full fine-tuning (not LoRA) | |
| ⚖️ License | |
| This model is released under the Apache 2.0 License, allowing both commercial and research use with attribution. | |
| 💬 Citation | |
| If you use this model, please cite: | |
| java | |
| Copy code | |
| @model{kelvinnnadi_plant_vit_2025, | |
| title={Plant Identification ViT (Fine-Tuned)}, | |
| author={Kelvin Nnadi}, | |
| year={2025}, | |
| howpublished={Hugging Face}, | |
| url={https://huggingface.co/your-username/plant-identification-vit-finetuned} | |
| } | |
| 🏆 Highlights | |
| Fine-tuned with 47 classes of houseplants | |
| Highly generalized on real-world photos | |
| Seamlessly integrates with multimodal LLMs | |
| Production-grade architecture suitable for cloud APIs |