README_MODAL.md

Running on Modal + Image Rendering Guide

1. Running Training on Modal

Setup

pip install modal
modal setup  # Authenticate with your Modal token

Create a Modal Secret for HuggingFace

modal secret create huggingface-token HF_TOKEN=your_hf_token_here

Deploy & Run

The repo includes modal_train.py. Simply:

cd parametric-floorplan-generator
modal run modal_train.py

This will:

1. Spin up a CPU container to generate the 5,000-example synthetic dataset (saved to a Modal Volume)
2. Spin up an A10G GPU container to fine-tune Qwen2.5-1.5B-Instruct with LoRA
3. Push the trained model to HuggingFace Hub

Customize GPU

Edit modal_train.py and change the GPU:

@app.function(gpu="A100-40GB", ...)  # or "T4", "H100"

Check Progress

modal app logs floorplan-trainer

---

2. Generating Floorplan Images

The model outputs JSON with room polygons. To convert to visual plans:

Option A: SVG (Vector, best for CAD/printing)

# Generate a floorplan first
python generate.py \
  --plot_length 15 --plot_width 12 \
  --setback_front 1.5 --setback_rear 1.0 \
  --setback_left 1.0 --setback_right 1.0 \
  --road_side N --num_bedrooms 3 --toilets 3 \
  --parking --has_pooja --has_balcony \
  --num_floors 2 --city Delhi > myhouse.json

# Render to SVG
python render_floorplan.py --input myhouse.json --output myhouse.svg

The SVG includes:

• Plot boundary (thick black line)
• Buildable boundary (dashed gray)
• Rooms color-coded by type (living=blue, bedroom=orange, kitchen=purple, etc.)
• Door openings (green lines)
• Windows (blue dashed lines)
• Room labels with names and areas
• Dimension annotations
• North arrow
• Legend

Option B: PNG (Raster, best for web/presentations)

pip install cairosvg
python render_floorplan.py --input myhouse.json --output myhouse.png

Option C: Interactive Web Viewer (Gradio)

Deploy as a HuggingFace Space:

import gradio as gr
import json
from render_floorplan import render_floorplan_svg

def generate_and_render(params_json):
    floorplan = json.loads(model_output)
    svg = render_floorplan_svg(floorplan, width=1200)
    return svg

gr.Interface(
    fn=generate_and_render,
    inputs=gr.JSON(label="Project Parameters"),
    outputs=gr.HTML(label="Floorplan SVG"),
    title="Parametric Floorplan Generator"
).launch()

Option D: CAD Export (DXF)

For professional CAD output, extend render_floorplan.py to write DXF using ezdxf:

pip install ezdxf

Then iterate over the JSON polygons and write them as DXF polylines.

---

3. Complete Pipeline on Modal

You can also run inference + rendering as a Modal endpoint:

import modal

app = modal.App("floorplan-api")

@app.function(gpu="T4", image=modal.Image.debian_slim().pip_install("transformers", "torch", "accelerate", "cairosvg"))
@modal.web_endpoint(method="POST")
def generate(params: dict):
    # 1. Run model inference
    # 2. Render SVG
    # 3. Convert to PNG
    # 4. Return image
    pass

This gives you an HTTP API that takes your ProjectCreate JSON and returns a PNG floorplan.

---

4. Summary of Commands

# Clone repo
git clone https://huggingface.co/Karthik8nitt/parametric-floorplan-generator
cd parametric-floorplan-generator

# Install locally
pip install transformers trl torch datasets peft accelerate trackio
pip install cairosvg  # for PNG rendering

# Train on Modal
modal run modal_train.py

# Generate floorplan
python generate.py --plot_length 15 --plot_width 12 ... > plan.json

# Render
python render_floorplan.py --input plan.json --output plan.svg
python render_floorplan.py --input plan.json --output plan.png