File size: 2,919 Bytes
d881e67 efcd089 d881e67 efcd089 c091c99 efcd089 c091c99 088afc9 c091c99 efcd089 c65c850 efcd089 c091c99 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | ---
library_name: transformers
tags:
- generated_from_trainer
model-index:
- name: dragon_interceptor
results: []
---
# 🐉 Dragon Interceptor (267k)
**Dragon Interceptor** is an ultra-compact, high-speed generative model designed for aviation-inspired structural synthesis. Optimized for **35W mobile hardware**, it bridges the gap between neural text generation and 2D pixel blueprints.
[]()
[]()
[]()
---
## 🚀 Performance HUD
Benchmarks recorded on **Intel Core i5-10210U** (Surface Pro setup):
* **Weight Loading:** 2055.47 it/s
* **Inference Speed:** ~286 iterations/sec
* **Full Image Gen (28x28):** ~0.88s
* **Brute Force Throughput:** 1.2 - 2.5 images/sec (Turbo Mode)
## 🛠️ Architecture
The model utilizes a **GPT-2 Causal LM** backbone, repurposed for spatial data:
- **Vocab Size:** 256 (Mapped to 8-bit grayscale intensity)
- **Sequence Length:** 784 (Fixed $28 \times 28$ positional embeddings)
- **Parameters:** 267,000 (Fits entirely within L2/L3 CPU cache)
## 🛰️ Key Features
### 1. Mosaic Synthesis
By utilizing a sliding-window context bridge, the model can bypass its 784-position limit to generate seamless, multi-tile blueprints.
- **Tile Resolution:** 28x28
- **Global Resolution:** 112x112 (4x4 Mosaic) or custom strips.
### 2. Seed Brute-Forcing
The "Titan DNA" protocol allows for mass-generation of 1,000+ seeds to map the latent space for specific aircraft parts (wings, fuselages, tail fins).
### 3. Thermal-Aware Inference
Optimized for the Surface Pro's 35W power envelope. Uses raw Torch inference with **KV-Caching** to maintain stable frame rates even under thermal pressure.
## 🖼️ Sample Inference & Visualization
Use the following code to generate a high-fidelity "Radar Scan" from a specific seed. This snippet is optimized for 4K displays and technical clarity.
```python
import torch
from transformers import AutoModelForCausalLM
import matplotlib.pyplot as plt
import os
# Load Dragon Interceptor
model = AutoModelForCausalLM.from_pretrained("MightyDragon-Dev/dragon_interceptor")
# Generate a 28x28 Blueprint (Random Seed)
seed_id = os.urandom(1)[0] % 10**6 # Random seed for variability
print(f"🚀 Generating Dragon Blueprint with Seed {seed_id}...")
input_ids = torch.tensor([[seed_id]])
output = model.generate(input_ids, max_length=784, min_length=784, do_sample=True, temperature=0.7)
# Reshape and Render
blueprint = output[0].view(28, 28).detach().numpy()
plt.figure(figsize=(8, 8), dpi=120)
plt.imshow(blueprint, cmap='magma', interpolation='lanczos')
plt.title(f"Dragon Interceptor: Sector Scan (Seed {seed_id})", color='white')
plt.style.use('dark_background')
plt.axis('off')
plt.show()
|