EfficientDet-tensorflow2

A Tensorflow2.x implementation of EfficientDet as described in EfficientDet: Scalable and Efficient Object Detection, The project is based on official implementation of google/automl/efficientdet.

Update Log

[2021-03-03]

Package model’s preprocess part into inference model:The key benefit to doing this is that it makes your model portable and it helps reduce the training/serving skew.
Add TTA(Test Time Augmentation) in inference mode.

[2021-03-02]

Add Graph mode training with model.fit: this dramatically improve GPU utilization(over 95%).

[2021-03-01]

Add support for: efficientdet d0-d7,huber loss,focal loss.
Eager mode training with tf.GradientTape.
Add online coco evaluation callback.
Add ssd_random_crop;mosaic.
Support tensorboard.

Installation

1. Clone project

  git clone https://github.com/wangermeng2021/EfficientDet-tensorflow2.git
  cd EfficientDet-tensorflow2

2. Install environment

Install tesnorflow (skip this step if it’s already installed,test environment:tensorflow 2.3.0)
Install dependencies: pip install -r requirements.txt

Training:

Pretrained EfficientDet Checkpoints(google automl’s efficientdet):

Model	AP^test	AP₅₀	AP₇₅	AP_S	AP_M	AP_L	AP^val	#params	#FLOPs
EfficientDet-D0 (ckpt)	34.6	53.0	37.1	12.4	39.0	52.7	34.3	3.9M	2.54B
EfficientDet-D1 (ckpt)	40.5	59.1	43.7	18.3	45.0	57.5	40.2	6.6M	6.10B
EfficientDet-D2 (ckpt)	43.0	62.3	46.2	22.5	47.0	58.4	42.5	8.1M	11.0B
EfficientDet-D3 (ckpt)	47.5	66.2	51.5	27.9	51.4	62.0	47.2	12.0M	24.9B
EfficientDet-D4 (ckpt)	49.7	68.4	53.9	30.7	53.2	63.2	49.3	20.7M	55.2B
EfficientDet-D5 (ckpt)	51.5	70.5	56.1	33.9	54.7	64.1	51.2	33.7M	130B
EfficientDet-D6 (ckpt)	52.6	71.5	57.2	34.9	56.0	65.4	52.1	51.9M	226B
EfficientDet-D7 (ckpt)	53.7	72.4	58.4	35.8	57.0	66.3	53.4	51.9M	325B
EfficientDet-D7x (ckpt)	55.1	74.3	59.9	37.2	57.9	68.0	54.4	77.0M	410B

Download pretrain weight and place it under directory ‘./pretrain’.

For training on pothole dataset(No need to download dataset,it’s already included in project):

python train.py --model-type d0  --use-pretrain True --dataset-type voc --dataset dataset/pothole_voc --num-classes 1 --class-names dataset/pothole.names --voc-train-set dataset_1,train --voc-val-set dataset_1,val  --epochs 200 --batch-size 8 --augment ssd_random_crop

Tensorboard visualization:

Navigate to http://0.0.0.0:6006: you need to manually enable: “Setting”—>”Reload data” on tensorboard home page to automatically update data
mAP@0.5):" class="reference-link">Evaluation results(GTX2080,mAP@0.5):

model	pothole	VOC	COCO
efficientdet-d0(512)	0.798
Scaled-YoloV4-p5(416)	0.826

Evaluation on Pothole dataset:

Detection

For detection on Pothole dataset:

python3 detect.py --model-dir export/best_model_d0_189_0.798/1 --pic-dir images/pothole --class-names dataset/pothole.names --score-threshold 0.1

detection result: