Configuration

This document describes all configuration options available in the Spectrum Analyzer codebase.

Repository: GitHub

Dataset Configuration

Dataset configuration is specified in YAML files in the data/ directory.

Dataset YAML Structure

# Dataset root directory
path: /path/to/dataset

# Train, validation, and test paths (relative to path)
train: images/train
val: images/val
test: images/test

# Number of classes
nc: 11

# Class names
names: ['Rect', 'Barker', 'Frank', 'P1', 'P2', 'P3', 'P4', 'Px', 'ZadoffChu', 'LFM', 'FMCW']

Example: RadDet Configuration

# data/raddet.yaml
path: /home/ubuntu/RadDet/RadDet40k512HW009Tv2
train: images/train
val: images/val
test: images/test
nc: 11
names: ['Rect', 'Barker', 'Frank', 'P1', 'P2', 'P3', 'P4', 'Px', 'ZadoffChu', 'LFM', 'FMCW']

Model Configuration

Model architectures are defined in YAML files in the models/ directory.

Model YAML Structure

# Number of classes
nc: 11

# Depth and width multipliers
depth_multiple: 1.0
width_multiple: 1.0

# Anchor boxes for each detection layer
anchors:
  - [10,14, 23,27, 37,58]  # P4/16
  - [81,82, 135,169, 344,319]  # P5/32

# Backbone layers
backbone:
  [[-1, 1, Conv_1, [64, 7, 2]],
   [-1, 1, BasicBlock_2, [64, 3, 2]],
   # ... more layers
  ]

# Detection head
head:
  [[-1, 1, Detect, [nc, anchors]],]

Available Model Configurations

• resnet10.yaml: Lightweight ResNet-10 backbone
• resnet18.yaml: ResNet-18 backbone (default)
• resnet34.yaml: Deeper ResNet-34 backbone
• resnet34-cat.yaml: ResNet-34 with concatenation
• Custom architectures can be created

Hyperparameters Configuration

Training hyperparameters are defined in data/hyps/hyp.scratch.yaml.

Hyperparameters Structure

# Learning rate
lr0: 0.01  # initial learning rate
lrf: 0.1  # final learning rate (lr0 * lrf)

# Optimizer
momentum: 0.937  # SGD momentum
weight_decay: 0.0005  # weight decay

# Learning rate schedule
warmup_epochs: 3.0
warmup_momentum: 0.8
warmup_bias_lr: 0.1

# Loss weights
box: 0.05  # box loss gain
cls: 0.5  # classification loss gain
obj: 1.0  # objectness loss gain

# Data augmentation
hsv_h: 0.015  # HSV-Hue augmentation
hsv_s: 0.7  # HSV-Saturation augmentation
hsv_v: 0.4  # HSV-Value augmentation
degrees: 0.0  # rotation (+/- deg)
translate: 0.1  # translation (+/- fraction)
scale: 0.5  # scale (+/- gain)
shear: 0.0  # shear (+/- deg)
flipud: 0.0  # flip up-down probability
fliplr: 0.5  # flip left-right probability
mosaic: 1.0  # mosaic augmentation probability
mixup: 0.0  # mixup augmentation probability

Training Configuration

Training is configured via command-line arguments in train.py.

Training Parameters

Parameter	Type	Default	Description
--weights	str	None	Initial weights path
--cfg	str	models/resnet10.yaml	Model configuration file
--data	str	data/coco.yaml	Dataset configuration file
--hyp	str	data/hyps/hyp.scratch.yaml	Hyperparameters file
--epochs	int	300	Number of training epochs
--batch-size	int	16	Batch size
--imgsz	int	640	Image size (pixels)
--rect	flag	False	Rectangular training
--resume	str/bool	False	Resume training from checkpoint
--device	str	4,5,6,7	CUDA device(s) or 'cpu'
--workers	int	4	DataLoader workers
--project	str	runs/train	Project directory
--name	str	exp	Experiment name
--cache	str	None	Cache images in 'ram' or 'disk'
--multi-scale	flag	False	Vary image size +/- 50%
--single-cls	flag	False	Train as single-class
--adam	flag	False	Use Adam optimizer
--freeze	int	0	Number of layers to freeze
--save-period	int	-1	Save checkpoint every N epochs

Example Training Command

python train.py \
    --data data/raddet.yaml \
    --cfg models/resnet18.yaml \
    --imgsz 512 \
    --epochs 300 \
    --batch-size 64 \
    --workers 20 \
    --device 0 \
    --cache ram

Validation Configuration

Validation is configured via command-line arguments in val.py.

Validation Parameters

Parameter	Type	Default	Description
--weights	str	Required	Model weights path
--data	str	data/coco.yaml	Dataset configuration
--batch-size	int	32	Batch size
--imgsz	int	640	Image size (pixels)
--conf-thres	float	0.001	Confidence threshold
--iou-thres	float	0.6	NMS IoU threshold
--task	str	val	Task: train/val/test
--device	str	''	CUDA device or 'cpu'
--augment	flag	False	Augmented inference
--verbose	flag	False	Report mAP by class
--save-txt	flag	False	Save results to *.txt
--save-json	flag	False	Save COCO-JSON results

Example Validation Command

python val.py \
    --weights runs/train/exp28/weights/best.pt \
    --data data/raddet.yaml \
    --task test \
    --imgsz 512 \
    --conf-thres 0.001

Detection/Inference Configuration

Detection is configured via command-line arguments in detect.py.

Detection Parameters

Parameter	Type	Default	Description
--weights	str	Required	Model weights path
--source	str	data/images	Source: file/dir/URL
--imgsz	int	640	Inference image size
--conf-thres	float	0.25	Confidence threshold
--iou-thres	float	0.45	NMS IoU threshold
--max-det	int	1000	Maximum detections per image
--device	str	''	CUDA device or 'cpu'
--save-txt	flag	False	Save results to *.txt
--save-img	flag	False	Save inference images
--save-conf	flag	False	Save confidences in labels
--classes	list	None	Filter by class IDs
--agnostic-nms	flag	False	Class-agnostic NMS
--augment	flag	False	Augmented inference
--half	flag	False	FP16 half-precision
--line-thickness	int	3	Bounding box thickness
--hide-labels	flag	False	Hide labels
--hide-conf	flag	False	Hide confidences

Example Detection Command

python detect.py \
    --weights runs/train/exp28/weights/best.pt \
    --source path/to/images \
    --imgsz 512 \
    --conf-thres 0.25

Detection Parameters

Parameter	Type	Default	Description
--weights	str	Required	Model weights path
--source	str	data/images	Source: file/dir/URL
--imgsz	int	640	Inference image size
--conf-thres	float	0.25	Confidence threshold
--iou-thres	float	0.45	NMS IoU threshold
--max-det	int	1000	Maximum detections per image
--device	str	''	CUDA device or 'cpu'
--save-txt	flag	False	Save results to *.txt
--save-img	flag	False	Save inference images
--save-conf	flag	False	Save confidences in labels
--classes	list	None	Filter by class IDs
--agnostic-nms	flag	False	Class-agnostic NMS
--augment	flag	False	Augmented inference
--half	flag	False	FP16 half-precision
--line-thickness	int	3	Bounding box thickness
--hide-labels	flag	False	Hide labels
--hide-conf	flag	False	Hide confidences

Example Detection Command

python detect.py \
    --weights runs/train/exp28/weights/best.pt \
    --source path/to/images \
    --imgsz 512 \
    --conf-thres 0.25 \
    --save-img true

Model Conversion Configuration

Model conversion from SiLU to binary spiking is configured in convert_to_binary_spiking.py.

Conversion Parameters

Parameter	Type	Default	Description
--weights	str	Required	Path to trained model weights (.pt file)
--output	str	*_binary.pt	Output path for converted model (adds _binary suffix if not specified)
--diagram	str	model_architecture.txt	Output file for architecture diagram
--device	str	cuda	Device to load model on (cpu or cuda)
--data	str	None	Path to dataset YAML file for forward pass tracing (recommended for accurate diagram)
--img-size	int	640	Input image size for forward pass tracing (used as fallback if --data not provided)
--no-trace	flag	False	Disable forward pass tracing (show all modules, including unused ones)

Conversion Behavior

Default (without --data):

• Uses dummy data for forward pass
• Shows all layers including unused shortcuts and conditional layers
• May not accurately represent actual execution paths

With --data (Recommended):

• Uses real test data from dataset
• Performs actual forward pass
• Only shows layers and operations actually executed
• Provides accurate architecture representation

Example Conversion Commands

Basic conversion:

python convert_to_binary_spiking.py \
    --weights runs/train/exp28/weights/best.pt

With forward pass tracing (accurate diagram):

python convert_to_binary_spiking.py \
    --weights runs/train/exp28/weights/best.pt \
    --output arch/model_binary.pt \
    --diagram arch/architecture_diagram.txt \
    --data data/raddet.yaml \
    --img-size 512 \
    --device cuda

Show all layers (disable tracing):

python convert_to_binary_spiking.py \
    --weights runs/train/exp28/weights/best.pt \
    --no-trace

Model Quantization Configuration

Quantization is configured in quantize_model.py. For detailed quantization guide, see Quantization Documentation.

Quantization Parameters

Parameter	Type	Default	Description
--weights	str	Required	Path to model weights (.pt file)
--quant-type	str	Required	Quantization type: fp16, dynamic_int8, static_int8, or per_channel_int8
--output	str	quantized_models	Output directory for quantized models
--data	str	None	Dataset YAML file (required for static/per-channel INT8)
--device	str	cpu	Device to use (cpu or cuda)
--imgsz	int	640	Image size for calibration
--batch-size	int	16	Batch size for calibration data
--num-calib-batches	int	100	Number of batches for calibration (static quantization)
--workers	int	4	Number of dataloader workers

Quantization Types

FP16 (Half Precision):

• ~2x compression
• No calibration needed
• Best for GPU deployment

Dynamic INT8:

• ~4x compression
• No calibration needed
• Weights quantized, activations FP32

Static INT8:

• ~4x compression
• Requires calibration data
• Best for production deployment

Per-Channel INT8:

• ~4x compression
• Requires calibration data
• Best accuracy among INT8 methods

Example Quantization Commands

FP16 Quantization:

python quantize_model.py \
    --weights arch/model_binary.pt \
    --quant-type fp16 \
    --output quantized_models/ \
    --device cuda

Static INT8 Quantization:

python quantize_model.py \
    --weights arch/model_binary.pt \
    --quant-type static_int8 \
    --data data/raddet.yaml \
    --num-calib-batches 50 \
    --batch-size 16 \
    --imgsz 512 \
    --output quantized_models/

Per-Channel INT8 Quantization:

python quantize_model.py \
    --weights arch/model_binary.pt \
    --quant-type per_channel_int8 \
    --data data/raddet.yaml \
    --num-calib-batches 50 \
    --batch-size 16 \
    --imgsz 512 \
    --output quantized_models/

Image Size Configuration

Image size is configurable and affects:

• Model Input: All images are resized to this size
• Detection Resolution: Higher resolution = better detection but slower
• Memory Usage: Larger images require more GPU memory

Common Sizes:

• 320×320: Fast, lower accuracy
• 512×512: Balanced (default for RadDet)
• 640×640: Higher accuracy, slower
• Custom: Any size divisible by 32 (YOLO requirement)

Set via --imgsz parameter in training, validation, and detection.

Device Configuration

Single GPU:

--device 0

Multiple GPUs:

--device 0,1,2,3

CPU:

--device cpu

Output Configuration

Training Output:

• runs/train/{name}/weights/: Model checkpoints
• runs/train/{name}/results.png: Training curves
• runs/train/{name}/PR_curve.png: Precision-Recall curve
• runs/train/{name}/P_curve.png: Precision-Confidence curve
• runs/train/{name}/R_curve.png: Recall-Confidence curve

Detection Output:

• runs/detect/{name}/: Detection result images
• runs/detect/{name}/labels/: Detection labels (if --save-txt)

Model Conversion Output:

• *_binary.pt: Converted model with binary spiking activations
• *_binary_state_dict.pt: Model state dictionary
• model_architecture.txt: Detailed architecture diagram

Configure via --project and --name parameters.

Related Documentation

• Architecture: Understanding the model architecture
• Training: Training configuration and usage
• Detection: Detection/inference configuration
• Model Conversion: Converting models for deployment
• Quantization: Model quantization options
• RadDet Use Case: Example configuration for RadDet dataset