# Polynomial Learning Rate Scheduler

A Python script for visualizing polynomial learning rate decay schedules with warmup. This tool helps machine learning practitioners visualize different learning rate decay curves to better understand and choose appropriate learning rate schedules for their training.

## Features

- Polynomial decay with configurable powers
- Linear warmup period
- Multiple visualization scales (linear and logarithmic)
- Scientific and decimal notation options
- Automatic filename generation with parameters
- Customizable output format and DPI
- Multiple preset configurations

## Installation

1. Clone the repository:
```bash
git clone https://github.com/yourusername/poly-lr-scheduler.git
cd poly-lr-scheduler
```

2. Create and activate a virtual environment (recommended):
```bash
# Windows
python -m venv lr_venv
lr_venv\Scripts\activate

# Linux/MacOS
python -m venv lr_venv
source lr_venv/bin/activate
```

3. Install dependencies:
```bash
pip install -r requirements.txt
```

## Usage

Basic usage:
```bash
python poly.py -p 0.5 1 2
```

### Command Line Arguments

| Flag | Description | Default | Example |
|------|-------------|---------|---------|
| `-p`, `--powers` | Space-separated polynomial powers | `0.5 0.8 1 1.5 2 3` | `-p 0.5 1 2` |
| `-lr`, `--learning-rate` | Initial learning rate | `1e-4` | `-lr 1e-3` |
| `-lre`, `--lr-end` | Final learning rate | `1e-7` | `-lre 1e-8` |
| `-s`, `--steps` | Total training steps | `4000` | `-s 5000` |
| `-w`, `--warmup` | Warmup steps | `50` | `-w 100` |
| `-o`, `--output` | Output filename | `lr_schedule.jpg` | `-o custom.jpg` |
| `--dpi` | Image DPI | `300` | `--dpi 600` |
| `-l`, `--log-scale` | Scale type (none/standard/fine/wide) | `none` | `-l standard` |
| `-n`, `--notation` | Y-axis notation (s/d) | `s` | `-n d` |

### Scale Presets

- `none`: Linear scale
- `standard`: Regular logarithmic scale
- `fine`: Dense logarithmic scale with more tick marks
- `wide`: Extended logarithmic scale with fewer tick marks

### Examples

1. Basic visualization with default settings:
```bash
python poly.py
```

2. Custom powers with standard log scale:
```bash
python poly.py -p 0.5 1 2 3 -l standard
```

3. Full customization:
```bash
python poly.py -p 0.5 0.8 1 1.5 2 3 5 -lr 1e-4 -lre 1e-7 -w 50 -s 4000 -l standard -n s
```

4. Different scale options:
```bash
# Linear scale
python poly.py -l none

# Standard log scale
python poly.py -l standard

# Fine-grained log scale
python poly.py -l fine

# Wide-range log scale
python poly.py -l wide
```

### Auto-generated Filenames

The script automatically generates descriptive filenames including all parameters when no output filename is specified. The format is:

```
lr_decay_p{powers}_lr{initial_lr}_lre{final_lr}_s{steps}_w{warmup}_scale_{scale_type}_not_{notation}_dpi{dpi}.jpg
```

Example:
```
lr_decay_p0.5-1-2_lr1e-04_lre1e-07_s4000_w50_scale_standard_not_s_dpi300.jpg
```

## Output

The script generates:
1. A plot showing all learning rate decay curves
2. Console output with:
   - Filename of saved plot
   - Initial and final learning rates
   - Scale type and notation used

### Plot Features

- Color-coded curves for different powers
- Grid lines for better readability
- Scientific notation on y-axis (default)
- Legend showing power values
- Optional warmup period visualization
- Customizable scales (linear/logarithmic)

## Requirements

- Python 3.6+
- NumPy
- Matplotlib

See `requirements.txt` for specific versions.

## License

MIT License - feel free to use and modify as needed.