Skip to main content

​
Overview

BeamFinder uses three main configuration points:
  1. Training configuration - parameters for model fine-tuning in train.py
  2. Detection configuration - parameters for inference in detect.py
  3. Dataset configuration - data paths and class definitions in data.yaml

​
Training Configuration

The training pipeline (train.py) uses Ultralytics YOLO with optimized parameters for A100 GPU training.

​
Model Selection

​
model
string
default:"yolo26s.pt"
Base YOLO26 model to fine-tune. Available variants: yolo26n.pt, yolo26s.pt, yolo26m.pt, yolo26l.pt, yolo26x.pt

​
Training Arguments

All training arguments are passed to model.train() in train.py:16-36:
​
data
string
default:"data.yaml"
Path to dataset configuration file. Defines train/val/test splits and class names.
​
epochs
integer
default:"100"
Number of training epochs. The model trains for 100 epochs with early stopping via patience.
​
imgsz
integer
default:"960"
Input image size (square). All images are resized to 960×960 pixels during training and inference.
​
batch
float
default:"0.90"
Batch size as fraction of GPU memory (0.90 = 90% utilization). Automatically scales based on available VRAM.
​
patience
integer
default:"20"
Early stopping patience. Training stops if validation mAP doesn’t improve for 20 consecutive epochs.
​
cache
string
default:"ram"
Dataset caching strategy. Options: "ram" (cache in memory), "disk" (cache on SSD), or False (no caching).
​
workers
integer
default:"8"
Number of data loader workers for parallel image preprocessing.
​
cos_lr
boolean
default:"true"
Use cosine learning rate scheduler instead of linear decay.
​
deterministic
boolean
default:"false"
Enable deterministic mode for reproducible results. Set to false for faster training with cudnn.benchmark.
​
compile
boolean
default:"true"
Enable torch.compile() for 10-30% faster training on A100 + PyTorch 2.x.
​
rect
boolean
default:"true"
Rectangular training - batches images by aspect ratio to reduce padding and increase efficiency.
​
save_period
integer
default:"10"
Save checkpoint every N epochs (in addition to best.pt and last.pt).

​
Output Paths

​
project
string
default:"runs"
Root directory for training outputs. Resolved to absolute path via SCRIPT_DIR / "runs".
​
name
string
default:"drone_detect"
Experiment name. Results saved to {project}/{name}/.
​
exist_ok
boolean
default:"true"
Allow overwriting existing experiment directory.

​
Data Augmentation

BeamFinder applies geometric augmentations during training:
​
degrees
float
default:"15.0"
Random rotation range in degrees (±15°).
​
flipud
float
default:"0.5"
Probability of vertical flip (50%).
​
scale
float
default:"0.9"
Random scale range. 0.9 allows scaling from 90% to 110% of original size.
​
translate
float
default:"0.2"
Random translation as fraction of image size (±20%).

​
GPU Optimizations

The training script enables TF32 and cudnn benchmark for A100 GPUs (train.py:10-12): 
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cudnn.benchmark = True
These optimizations provide ~30% speedup on A100 but may reduce numerical precision slightly.

​
Validation Arguments

After training completes, BeamFinder runs validation and test evaluation (train.py:38-42):
​
imgsz
integer
default:"960"
Validation image size (must match training).
​
half
boolean
default:"true"
Use FP16 precision for faster validation.
​
split
string
default:"val"
Dataset split to evaluate. Options: "val" or "test".

​
Detection Configuration

The detection script (detect.py) loads a trained model and runs inference on test images.

​
Model & Paths

Configuration variables defined at detect.py:8-14:
​
MODEL
string
default:"runs/drone_detect/weights/best.pt"
Path to trained model checkpoint. Points to best weights from training.
​
IMAGE_DIR
Path
default:"data/images/test"
Directory containing images to process.
​
OUTPUT_DIR
Path
default:"output"
Directory for detection results (CSV + annotated images).
​
CONF
float
default:"0.4"
Confidence threshold for detections. Only boxes with confidence ≥ 0.4 are saved.
​
IMGSZ
integer
default:"960"
Inference image size (must match training).

​
Prediction Arguments

Passed to model.predict() at detect.py:34-38:
​
source
string
required
Path to image directory or single image file.
​
conf
float
default:"0.4"
Minimum confidence threshold (0-1). Lower values increase recall but add false positives.
​
imgsz
integer
default:"960"
Input image size for inference.
​
save
boolean
default:"true"
Save annotated images with bounding boxes drawn.
​
project
string
default:"output"
Root directory for saving annotated images.
​
name
string
default:"annotated"
Subdirectory name for annotated images.
​
exist_ok
boolean
default:"true"
Overwrite existing output directory.
​
half
boolean
default:"true"
Use FP16 precision for faster inference.
​
batch
integer
default:"16"
Batch size for inference. Process 16 images at a time.

​
CSV Output

Detection results are written to output/detections.csv with 7 columns (see Output Format for details).

​
Dataset Configuration

The data.yaml file defines dataset paths and class labels for YOLO training.

​
Structure

path: data
train: images/train
val: images/validation
test: images/test

nc: 1
names: ["drone"]

​
Fields

​
path
string
default:"data"
Root directory for dataset. All other paths are relative to this.
​
train
string
default:"images/train"
Training images directory (7,970 images).
​
val
string
default:"images/validation"
Validation images directory (1,708 images).
​
test
string
default:"images/test"
Test images directory (1,709 images).
​
nc
integer
default:"1"
Number of classes. BeamFinder detects a single class (drone).
​
names
array
default:"[\"drone\"]"
Class names list. Must have length equal to nc.
Each image in images/train, images/validation, and images/test must have a corresponding label file in labels/train, labels/validation, and labels/test with the same filename but .txt extension.

​
Label Format

YOLO format: one bounding box per line 
<class_id> <x_center> <y_center> <width> <height>
All coordinates are normalized to [0, 1] relative to image dimensions. Example: 
0 0.5234 0.6891 0.1245 0.0934
  • class_id: 0 (drone is the only class)
  • x_center, y_center: center of bounding box
  • width, height: box dimensions

​
Environment Setup

​
PyTorch Backend Configuration

Both train.py and detect.py configure PyTorch for optimal A100 performance: 
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cudnn.benchmark = True
cudnn.benchmark=True is optimized for fixed input sizes. If you change imgsz, restart training to re-tune convolution algorithms.

​
Hardware Requirements

  • GPU: NVIDIA A100 40GB recommended (training uses ~36GB VRAM with batch=0.90)
  • CPU: 8+ cores for workers=8 data loading
  • RAM: 32GB+ for cache="ram" (caches full dataset in memory)
  • Storage: 10GB+ for dataset + model checkpoints

​
Python Dependencies

From requirements.txt: 
ultralytics
The Ultralytics package includes PyTorch, CUDA libraries, and all YOLO dependencies.

Build docs developers (and LLMs) love

Get started for freeTalk to us