NanoVLM: A High-Performance "Nano" Vision-Language Model

NanoVLM is a robust, PyTorch-based implementation of a Vision-Language Model (VLM) inspired by OpenAI's CLIP architecture. While "Nano" in scale, it utilizes professional-grade techniques including pre-trained backbones, EMA weighting, and gradient accumulation to achieve state-of-the-art alignment quality on a single consumer GPU.

This project demonstrates how to bridge visual and linguistic semantics by aligning representations in a shared 256-dimensional latent space, trained natively on the massive 591k-pair MS-COCO dataset.

🚀 Key Features

• Advanced Multimodal Architecture:
  • Image Encoder: Utilizes a MobileNetV2 backbone (pretrained on ImageNet) for rich feature extraction. Upgraded with Spatial Self-Attention over a 7x7 patch grid (49 tokens) allowing the model to learn localized feature relevance before pooling.
  • Text Encoder: A deep 6-layer Transformer (4 attention heads) with Masked Mean Pooling to correctly handle padding tokens during semantic alignment.
• Superior Training Stability:
  • Optimized Regularization: Uses a weight decay of 0.05 and label smoothing (0.1) to ensure robust generalization and prevent overfitting on smaller datasets.
  • Gradient Accumulation: Simulates an effective batch size of 512, providing stable contrastive gradients without exceeding VRAM limits.
• Professional Pipeline:
  • Mixed Precision (AMP): Uses torch.amp for accelerated training.
  • Scheduling: Linear Warmup followed by Cosine Decay for optimal convergence.
  • Robust Early Stopping: Improved patience settings to allow the model to fully explore the loss landscape.

📊 Performance Results (MS-COCO)

After an extended training run on the full MS-COCO dataset (using Spatial Attention), the model achieved the following state-of-the-art metrics for its size class:

Metric	Result
Best Val Loss	2.6436
Image→Text Recall@1	19.2%
Image→Text Recall@5	68.7%
Image→Text Recall@10	84.8%

Note: Recall@10 of 84.8% indicates the correct matching caption is found in the top 10 results for nearly 85% of all images in the complex COCO validation set.

🛠️ Installation & Usage

1. Install Dependencies:

   pip install torch torchvision matplotlib pillow gradio
   

2. Run Inference: Ensure best_img_enc.pth, best_txt_enc.pth, and tokenizer_vocab.pth are in the root directory.

   python app.py
   

📂 Project Structure

• model.py: Core architecture, including tokenizer and dual-encoder definitions.
• nanoVLM.ipynb: Full training pipeline, including data ingestion, contrastive loop, and evaluation.
• app.py: Gradio-based web interface for real-time image-caption similarity matching.
• save_tokenizer.py: Utility script to serialize the training vocabulary for inference.