RectLabel
An offline image annotation tool for object detection and segmentation.
To download RectLabel apps.
Post the problem to our Github issues.
Have questions? Send an email to support@rectlabel.com.
How to Train a RF-DETR Instance Segmentation Model with Custom Data
We will show you how to train a RF-DETR instance segmentation model with your images and annotations and export to a Core ML model which can be used for auto labeling on RectLabel.
We recommend working through this blog post side-by-side with the Colab notebook of RF-DETR 1.5.1 Instance Segmentation and RF-DETR 1.6.0 Instance Segmentation.
Install RF-DETR.
# 1.5.1
pip install -q rfdetr==1.5.1
# 1.6.0
pip install -q rfdetr[train,loggers]==1.6.0
Download training images and annotations. You can use these or replace them with your own data.
mkdir datasets
cd datasets
wget -q https://huggingface.co/datasets/rectlabel/datasets/resolve/main/donut_coco.zip
unzip -q donut_coco.zip
cd ..
Fine-tune RF-DETR on custom dataset.
from rfdetr import RFDETRSegNano
model = RFDETRSegNano()
dataset_dir = "datasets/donut_coco"
model.train(dataset_dir=dataset_dir, epochs=10, batch_size=4, grad_accum_steps=4)
The trained model is checkpoint_best_total.pth.
ls -la /content/output
total 1956476
drwxr-xr-x 3 root root 4096 Mar 29 15:02 .
drwxr-xr-x 1 root root 4096 Mar 29 14:57 ..
-rw-r--r-- 1 root root 534398511 Mar 29 15:01 checkpoint0009.pth
-rw-r--r-- 1 root root 399873726 Mar 29 15:00 checkpoint_best_ema.pth
-rw-r--r-- 1 root root 401162890 Mar 29 14:59 checkpoint_best_regular.pth
-rw------- 1 root root 133248663 Mar 29 15:01 checkpoint_best_total.pth
-rw-r--r-- 1 root root 534387619 Mar 29 15:01 checkpoint.pth
drwxr-xr-x 2 root root 4096 Mar 29 14:57 eval
-rw-r--r-- 1 root root 4412 Mar 29 15:01 events.out.tfevents.1774796221.326de40b558e.1060.0
-rw-r--r-- 1 root root 111680 Mar 29 15:01 log.txt
-rw-r--r-- 1 root root 196723 Mar 29 15:02 metrics_plot.png
-rw-r--r-- 1 root root 755 Mar 29 15:02 results.json
If you installed RF-DETR 1.6.0, before exporting to a Core ML model, edit a line of transformer.py.
path = "/usr/local/lib/python3.12/dist-packages/rfdetr/models/transformer.py"
with open(path, "r") as f:
content = f.read()
modified_content = content.replace("mask_flatten, spatial_shapes_hw", "mask_flatten, spatial_shapes")
with open(path, "w") as f:
f.write(modified_content)
Install RF-DETR to CoreML.
git clone https://github.com/landchenxuan/rf-detr-to-coreml.git
cd rf-detr-to-coreml
pip install -q -e .
Move the best model to the current folder and export to a Core ML model.
mv /content/output/checkpoint_best_total.pth .
rfdetr-coreml --model seg-nano --weights checkpoint_best_total.pth
ls -la output
total 12
drwxr-xr-x 3 root root 4096 Mar 29 15:03 .
drwxr-xr-x 7 root root 4096 Mar 29 15:03 ..
drwxr-xr-x 3 root root 4096 Mar 29 15:03 rf-detr-seg-nano-checkpoint_best_total-fp32.mlpackage
Now you can auto label using the Core ML model on RectLabel.