peaceofcake 0.2.0

Simple vision model training and deployment. Apache 2.0.

peaceofcake

A simple Python wrapper for D-FINE object detection models. Pretrained weights are downloaded automatically. Includes an iOS demo app with real-time camera detection.

Installation

pip install -e .

Requirements: Python >= 3.9, PyTorch >= 2.0

Quick Start

from peaceofcake import DFINE

model = DFINE("dfine-n-coco")
results = model.predict("image.jpg", conf=0.3)
results[0].save("output.jpg")

Available Models

Model	Dataset	Size
dfine-n-coco	COCO	Nano (fastest)
dfine-s-coco	COCO	Small
dfine-m-coco	COCO	Medium
dfine-l-coco	COCO	Large
dfine-x-coco	COCO	XLarge (best accuracy)
dfine-s-obj2coco	Objects365+COCO	Small
dfine-m-obj2coco	Objects365+COCO	Medium
dfine-l-obj2coco	Objects365+COCO	Large
dfine-x-obj2coco	Objects365+COCO	XLarge

Weights are cached in ~/.cache/peaceofcake/weights/.

API

Inference

from peaceofcake import DFINE

model = DFINE("dfine-n-coco")

# From file path, PIL Image, numpy array, or list of paths
results = model.predict("image.jpg", conf=0.25, device="cpu", img_size=640)

Parameter	Default	Description
source	—	File path, list of paths, PIL Image, or numpy array
conf	0.25	Confidence threshold
device	auto	"cpu" or "cuda"
img_size	640	Input resolution

Results

r = results[0]
r.boxes       # (N, 4) bounding boxes in xyxy format
r.labels      # (N,) class indices
r.scores      # (N,) confidence scores
len(r)        # number of detections
print(r)      # human-readable summary

r.plot()      # returns PIL Image with drawn boxes
r.save("out.jpg")  # save visualization

Export

model.export("onnx")                # ONNX
model.export("coreml")              # CoreML (.mlpackage)
model.export("coreml", img_size=640, precision="FLOAT16", min_target="iOS17")
model.export("tensorrt")            # TensorRT (requires trtexec)

CoreML Export Options

Parameter	Default	Description
img_size	640	Input resolution
min_target	"iOS17"	"iOS16", "iOS17", "iOS18"
precision	"FLOAT16"	"FLOAT16", "FLOAT32"
compute_units	"ALL"	"ALL", "CPU_AND_GPU", "CPU_AND_NE", "CPU_ONLY"
output	"model.mlpackage"	Output path

CoreML model outputs:

• confidence — [N, 80] class scores
• coordinates — [N, 4] bounding boxes (normalized cxcywh)

iOS Demo App

The DFINEDemo/ directory contains a SwiftUI iOS app with:

• Real-time camera object detection
• Photo library detection
• Confidence threshold slider
• Model picker (when multiple models are bundled)

Setup

1. Export a CoreML model:

   from peaceofcake import DFINE
   model = DFINE("dfine-n-coco")
   model.export("coreml", output="dfine_n_coco.mlpackage")
   

2. Drag the .mlpackage into DFINEDemo/DFINEDemo/ in Xcode

3. Build and run on device (iOS 17+)

To use multiple models, add more .mlpackage files with dfine prefix. A model picker appears automatically in the toolbar.

Project Structure

peaceofcake/          # Python library
  models/dfine.py     # Model loading and registry
  engine/             # Predictor, exporter, trainer
  results/            # Detection results and plotting
  cfg/                # Model configs and defaults
third_party/dfine/    # Bundled D-FINE inference source
DFINEDemo/            # iOS demo app (SwiftUI)

License

Apache 2.0

Acknowledgments

This project wraps D-FINE by Peterande et al.

D-FINE: Redefine Regression Task of DETRs as Fine-grained Distribution Refinement.