A pure WORKING Tensorflow2.0 implementation of EfficientDet for object detection. There are too many non-working versions of EfficientDet available. This attempt uses pure tf2.
| Dog | People |
|---|---|
 |
 |
| Bird | Plane |
|---|---|
 |
 |
git clone https://github.com/IdeaKing/EfficientDet.git
cd efficientdet
python -m venv venv
python -m pip install --upgrade pip
venv\Scripts\activate
pip install -r requirements.txtYou can train your own dataset or use the Pascal VOC dataset.
All labels must be in xml Pascal VOC format.
Use LabelImg to label and create your own dataset. Make sure to save in PascalVOC format.
pip install labelimg
labelmeThen copy the file names of the images and respesctive labels to the train.txt and val.txt files. The file names should not include the file type.
Download the PascalVOC 2012 dataset from host.robots.ox.ac.uk.
Then copy the VOCdevkit/VOC2012 folder to /data/datasets/VOC2012. Lastly, copy the train.txt and val.txt files from data/datasets/VOC2012/ImageSets to data/datasets/VOC2012.
Then create a file with the labels/classes that will be trained on.
cat
dog
cow
person
The structure of the directory is as follows:
+-- data
| +-- datasets
| +-- VOC2012
| +-- Annotations
| +-- ImageSets
| +-- JPEGImages
| +-- SegmentationClass
| +-- SegmentationObject
| +-- ImageSets
| +-- train.txt
| +-- val.txt
| +-- labels.txt
| +-- MyDataset
| +-- Labels
| +-- image-0001.xml
| +-- image-0002.xml
| +-- ...
| +-- Images
| +-- image-0001.jpg
| +-- image-0002.jpg
| +-- ...
| +-- Test
| +-- image-0001.jpg
| +-- image-0002.jpg
| +-- ...
| +-- train.txt
| +-- val.txt
| +-- labels.txt
Run the following command to train the model:
python main.py --dataset-path <path_to_dataset> \
--training-dir <path_to_training_dir> \
--model efficientdet-d0 \
--debug False \
--precision float32 \
--batch-size 8 \
--epochs 100 \
--learning-rate 0.0005 \
--optimizer ADAM \
--dataset-files train.txt \
--labels-file labels.txt \
--images-dir JPEGImages|MyDataset \
--labels-dir Annotations|Labels \
--augment-ds True (venv) D:\EfficientDet>python main.py -h
usage: EfficientDet [-h] [--dataset-path DATASET_PATH] [--training-dir TRAINING_DIR]
[--model MODEL] [--debug DEBUG] [--precision PRECISION]
[--batch-size BATCH_SIZE] [--epochs EPOCHS] [--optimizer OPTIMIZER]
[--learning-rate LEARNING_RATE]
[--optimizer-momentum OPTIMIZER_MOMENTUM]
[--from-checkpoint FROM_CHECKPOINT] [--dataset-files DATASET_FILES]
[--labels-file LABELS_FILE] [--images-dir IMAGES_DIR]
[--labels-dir LABELS_DIR] [--shuffle-size SHUFFLE_SIZE]
[--image-dims IMAGE_DIMS] [--augment-ds AUGMENT_DS]
[--print-loss PRINT_LOSS] [--log-every-step LOG_EVERY_STEP]
[--max-checkpoints MAX_CHECKPOINTS]
[--save-model-frequency SAVE_MODEL_FREQUENCY]
[--checkpoint-frequency CHECKPOINT_FREQUENCY]
Run EfficientDet Training or Tests
optional arguments:
-h, --help show this help message and exit
--dataset-path DATASET_PATH
Path to dataset
--training-dir TRAINING_DIR
Path to the training directory
--model MODEL Model name, can be efficientdet_d[0 - 7] (object detection)
--debug DEBUG Use debugging mode or not, NOTE: SEVERELY REDUCES PPERFORMANCE
--precision PRECISION
The precision type, can either be mixed_float16 or float32
--batch-size BATCH_SIZE
Batch size for training
--epochs EPOCHS The number of epochs to train the model(s)
--optimizer OPTIMIZER
The optimizer to train the model on, can be: SGD or ADAM
--learning-rate LEARNING_RATE
The initial learning rate for the optimizer
--optimizer-momentum OPTIMIZER_MOMENTUM
The momentum for the optimizer
--from-checkpoint FROM_CHECKPOINT
Continue training from checkpoint.
--dataset-files DATASET_FILES
Either filename of labeled_train.txt
--labels-file LABELS_FILE
Filename of the labels file.
--images-dir IMAGES_DIR
Folder name that holds images.
--labels-dir LABELS_DIR
Folder name that holds labels.
--shuffle-size SHUFFLE_SIZE
Shuffle the dataset steps, keep in powers of 2
--image-dims IMAGE_DIMS
Images dims to fit in the model
--augment-ds AUGMENT_DS
If augmentation is needed
--print-loss PRINT_LOSS
To print losses after each step
--log-every-step LOG_EVERY_STEP
Tensorboard logging every X steps
--max-checkpoints MAX_CHECKPOINTS
The maximum number of checkpoints.
--save-model-frequency SAVE_MODEL_FREQUENCY
Save model every X epochs.
--checkpoint-frequency CHECKPOINT_FREQUENCY
Save checkpoints every X epochs.Mixed precision training is supported. However, your GPU must have compute capability of at least 7.0.
To check your GPU's compute capability, run the following command:
D:\>nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 511.65 Driver Version: 511.65 CUDA Version: 11.6 |
|-------------------------------+----------------------+----------------------+
| GPU Name TCC/WDDM | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA GeForce ... WDDM | 00000000:01:00.0 Off | N/A |
| N/A 49C P0 20W / N/A | 409MiB / 6144MiB | 1% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+Finetuning can be accomplished by setting the --from-checkpoint flag to True.
To view Tensorboard, run the following command:
tensorboard --logdir=<path_to_training_dir/tensorboard>To test the trained model run the following command.
python inference.py --testing-image-dir <path_to_testing_images_dir> \
--save-image-dir <path_to_save_dir> \
--model <path_to_model> \
--labels-file <path_to_labels_file> \
--score-threshold 0.35 \
--iou-threshold 0.5(venv) D:\EfficientDet>python inference.py -h
usage: i-Sight [-h] [--testing-image-dir TESTING_IMAGE_DIR]
[--save-image-dir SAVE_IMAGE_DIR] [--model-dir MODEL_DIR]
[--image-dims IMAGE_DIMS] [--labels-file LABELS_FILE]
[--score-threshold SCORE_THRESHOLD] [--iou-threshold IOU_THRESHOLD]
Run i-Sight Tests
optional arguments:
-h, --help show this help message and exit
--testing-image-dir TESTING_IMAGE_DIR
Path to testing images directory.
--save-image-dir SAVE_IMAGE_DIR
Path to testing images directory.
--model-dir MODEL_DIR
Path to testing model directory.
--image-dims IMAGE_DIMS
Size of the input image.
--labels-file LABELS_FILE
Path to labels file.
--score-threshold SCORE_THRESHOLD
Score threshold for NMS.
--iou-threshold IOU_THRESHOLD
IOU threshold for NMS.Models were trained on PascalVOC2012, tf saved models they can only be used for inference. Saved weights can be used for transfer learning. More models and weights to come soon!
| Model | Model | Weights |
|---|---|---|
| EfficientDet-D0 | Link | Link |
[1] Official Repo [2] Paper [3] [https://github.com/joydeepmedhi/Anchor-Boxes-with-KMeans] [4] [https://github.com/fizyr/keras-retinanet]
Copyright 2020 Thomas Chia (IdeaKing)
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.