This is a demo about Faster R-CNN use GhostNet
This is a demo of Faster R-CNN for Windows. The backbone network uses GhostNet.
Faster R-CNN code from dBeker
GhostNet code from huawei-noah
This Demo is only for learning and communication, it is not certain that the code is completely correct. Due to limited computing resources, no suitable hyperparameters were found, and results did not perform well.
python 3.6
tensorflow 1.8.0
tensorpack 0.9.7
You need to make a dataset in VOC format
If it doesn't work, you need to set it up here and replace it with the corresponding file in the library
1. Put the dataset into './data'
2. in './lib/datasets/pascal_voc.py' to change class
<self.path = r'D:\Faster-R-CNN\labels_13_BJOYYX.txt'>
3. Set pre-training file path
<pretrained_model = r'F:\GhostNet\ghostnet\models\ghostnet_checkpoint'>
run train-ghostnet.py to train
run test-ghostnet.py to test
In my own dataset, batch=128, iteration=100000, results:
The results are not ideal, and I will continue to search for more suitable hyperparameters.
GhostNet is tiny.