# Standalone Float32 Training Script for 16-bit TIFFs

## Overview

This standalone script (`train_float32_standalone.py`) trains YOLO models on 16-bit grayscale TIFF datasets with **no data loss**.

- Loads 16-bit TIFFs with `tifffile` (not PIL/cv2)
- Converts to float32 [0-1] on-the-fly (preserves full 16-bit precision)
- Replicates grayscale → 3-channel RGB in memory
- **No disk caching required**
- Uses custom PyTorch Dataset + training loop

## Quick Start

```bash
# Activate virtual environment
source venv/bin/activate

# Train on your 16-bit TIFF dataset
python scripts/train_float32_standalone.py \
    --data data/my_dataset/data.yaml \
    --weights yolov8s-seg.pt \
    --epochs 100 \
    --batch 16 \
    --imgsz 640 \
    --lr 0.0001 \
    --save-dir runs/my_training \
    --device cuda
```

## Arguments

| Argument | Required | Default | Description |
|----------|----------|---------|-------------|
| `--data` | Yes | - | Path to YOLO data.yaml file |
| `--weights` | No | yolov8s-seg.pt | Pretrained model weights |
| `--epochs` | No | 100 | Number of training epochs |
| `--batch` | No | 16 | Batch size |
| `--imgsz` | No | 640 | Input image size |
| `--lr` | No | 0.0001 | Learning rate |
| `--save-dir` | No | runs/train | Directory to save checkpoints |
| `--device` | No | cuda/cpu | Training device (auto-detected) |

## Dataset Format

Your data.yaml should follow standard YOLO format:

```yaml
path: /path/to/dataset
train: train/images
val: val/images
test: test/images  # optional

names:
  0: class1
  1: class2

nc: 2
```

Directory structure:
```
dataset/
├── train/
│   ├── images/
│   │   ├── img1.tif  (16-bit grayscale TIFF)
│   │   └── img2.tif
│   └── labels/
│       ├── img1.txt  (YOLO format)
│       └── img2.txt
├── val/
│   ├── images/
│   └── labels/
└── data.yaml
```

## Output

The script saves:
- `epoch{N}.pt`: Checkpoint after each epoch
- `best.pt`: Best model weights (lowest loss)
- Training logs to console

## Features

✅ **16-bit precision preserved**: Float32 [0-1] maintains full dynamic range  
✅ **No disk caching**: Conversion happens in memory  
✅ **No PIL/cv2**: Direct tifffile loading  
✅ **Variable-length labels**: Handles segmentation polygons  
✅ **Checkpoint saving**: Resume training if interrupted  
✅ **Best model tracking**: Automatically saves best weights  

## Example

Train a segmentation model on microscopy data:

```bash
python scripts/train_float32_standalone.py \
    --data data/microscopy/data.yaml \
    --weights yolov11s-seg.pt \
    --epochs 150 \
    --batch 8 \
    --imgsz 1024 \
    --lr 0.0003 \
    --save-dir data/models/microscopy_v1
```

## Troubleshooting

### Out of Memory (OOM)
Reduce batch size:
```bash
--batch 4
```

### Slow Loading
Reduce num_workers (edit script line 208):
```python
num_workers=2  # instead of 4
```

### Different Image Sizes
The script expects all images to have the same dimensions. For variable sizes:
1. Implement letterbox/resize in dataset's `_read_image()`
2. Or preprocess images to same size

### Loss Computation Errors
If you see "Cannot determine loss", the script may need adjustment for your Ultralytics version. Check:
```python
# In train() function, the preds format may vary
# Current script assumes: preds is tuple with loss OR dict with 'loss' key
```

## vs GUI Training

| Feature | Standalone Script | GUI Training Tab |
|---------|------------------|------------------|
| Float32 conversion | ✓ Yes | ✓ Yes (automatic) |
| Disk caching | ✗ None | ✗ None |
| Progress UI | ✗ Console only | ✓ Visual progress bar |
| Dataset selection | Manual CLI args | ✓ GUI browsing |
| Multi-stage training | Manual runs | ✓ Built-in |
| Use case | Advanced users | General users |

## Technical Details

### Data Loading Pipeline

```
16-bit TIFF file
    ↓ (tifffile.imread)
uint16 [0-65535]
    ↓ (/ 65535.0)
float32 [0-1]
    ↓ (replicate channels)
float32 RGB (H,W,3) [0-1]
    ↓ (permute to C,H,W)
torch.Tensor (3,H,W) float32
    ↓ (DataLoader stack)
Batch (B,3,H,W) float32
    ↓
YOLO Model
```

### Precision Comparison

| Method | Unique Values | Data Loss |
|--------|---------------|-----------|
| **float32 [0-1]** | ~65,536 | None ✓ |
| uint16 RGB | 65,536 | None ✓ |
| uint8 | 256 | 99.6% ✗ |

Example: Pixel value 32,768 (middle intensity)
- Float32: 32768 / 65535.0 = 0.50000763 (exact)
- uint8: 32768 → 128 → many values collapse!

## License

Same as main project.