Rice Phenotype Prediction

Self-contained skill for predicting 10 rice agronomic traits via pre-trained MMoE models. All models, data, and scripts are inside this directory — give users this one folder.

Setup

First-time check

python <SKILL_DIR>/scripts/check_env.py

This verifies Python dependencies and data integrity. If packages are missing:

pip install -r <SKILL_DIR>/requirements.txt

Required: torch>=2.0 numpy pandas scikit-learn scipy requests GPU is optional — CPU works (just slower). If GPU is present, cuda:0 is used automatically.

<SKILL_DIR> convention

Throughout this file, <SKILL_DIR> means the absolute path to this skill's root directory (the folder containing this SKILL.md). When running commands, substitute with the actual path. --base_dir is optional; if omitted, scripts auto-detect it from their own location.

Supported Traits

Supported Locations (7 built-in stations)

Any input lat/lon is auto-matched to the nearest station via Haversine distance. For locations with internet, daily weather data can also be fetched from NASA POWER API for the exact coordinates.

Stress Types

Prediction Commands

Full prediction (recommended)

python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --sample sample1

Genotype-only / environment-only

python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --sample sample1 --mode gene
python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --sample sample1 --mode env

Specific traits

python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --sample sample1 --trait PH,Y

With stress

python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --sample sample1 --stress high_temp
python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --sample sample1 --stress high_temp --stress_delta 5.0

Multiple samples

python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --sample "sample1,sample2,sample3"

Custom genotype file

python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --genotype_file /path/to/user_vae.csv

Format: CSV with 1024 columns (VAE-encoded features), first column = sample index.

Force CPU / specific device

python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --sample sample1 --device cpu
python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --sample sample1 --device cuda:0

Human-readable table

python <SKILL_DIR>/scripts/predict.py --lat 30.5 --lon 114.3 --sample sample1 --output table

All CLI arguments

Handling User Requests

1. Extract location

• "经纬度30.5, 114.3" → --lat 30.5 --lon 114.3
• "武汉" → --lat 30.58 --lon 114.27
• "北纬25度，东经103度" → --lat 25 --lon 103

2. Map trait names

• 株高/plant height → PH
• 产量/yield → Y
• 粒长/grain length → GL
• 抽穗期/heading date → HD
• 千粒重/1000-grain weight → TGW
• 穗长/panicle length → PL
• 结实率/seed setting rate → SSR
• 每穗粒数/grains per panicle → GP
• 粒宽/grain width → GW
• 分蘖数/tiller number → TN

3. Map stress requests

• 高温/heat → high_temp
• 低温/cold/chilling → low_temp
• 干旱/drought → drought
• 洪涝/flooding → flood
• 阴天/寡照/low light → low_light
• "高温+5度" → --stress high_temp --stress_delta 5.0

4. Genotype data

• Built-in samples: --sample sample1 (3925 available: sample1..sample3925)
• User file: --genotype_file /path/to/file.csv

5. Interpreting output

JSON contains: location, genotype_prediction, environment_prediction, stress_prediction, trait_info.

Report environment_prediction as primary (has environmental context). Compare genotype_prediction as baseline. For stress, compare normal vs stressed values.

Rounding: HD/TN/GP → integer, PH/PL/TGW/SSR → 1 decimal, GL/GW → 2 decimals, Y → integer.

Directory Structure

rice_prediction/                   ← give users this folder
├── SKILL.md                       ← this file
├── requirements.txt               ← pip dependencies
├── data/
│   ├── grid_points.json           ← 7 station coordinates
│   ├── vae_features.csv           ← 3925 built-in genotype samples (1024-dim VAE)
│   ├── season_history.csv         ← historical season data for normalization
│   ├── env_cache/                 ← cached daily weather (auto-populated)
│   ├── models_env/                ← 10 trait-specific env+gene models (~4.6MB each)
│   └── models_gene/               ← 7 location-specific genotype models (~8MB each)
└── scripts/
    ├── predict.py                 ← main entry point
    ├── check_env.py               ← dependency checker
    ├── model_def.py               ← MMoE model architectures
    ├── grid_manager.py            ← nearest grid point finder
    ├── env_data_fetcher.py        ← NASA POWER API fetcher + cache
    ├── env_processor.py           ← environmental feature engineering
    └── stress_simulator.py        ← stress scenario simulation

Architecture (for reference)

• Model: Multi-gate Mixture-of-Experts (MMoE) with ResidualMLP experts
• Genotype features: 1024-dim VAE latent encoding of genomic data
• Environment features: 53 season-aggregated variables from daily weather
• Environmental data: NASA POWER API (auto-fetched and cached locally)