The project is training an accurate object detection neteork using Faster R-CNN for SVHN.
We training and testing with Python 3.6, pytorch 1.4 and Need to reference timm, AutoAugment, DataAugmentationForObjectDetection and mAP.
First, this Faster R-CNN uses EfficientNet-b4 as a backbone to extract features.
If you want to use this network for training, you must generate a compliant dataset.
In addition, my model can be loaded for testing, and this is described in detail in the Testing model section.
Before generating the data, Upload the training images to /data/train and the test data to /data/train.
Make sure you have the mat file(default name: digitStruct.mat) in the /data/train folder.
Example:
python generate_dataset.py
important arguments about Generate Dataset in config.py
Default:
| Argument | Default value |
|---|---|
| annotation_file | SVHN_annotation.pkl |
When the program was finished, we will get a file SVHN_annotation.pkl in /model/.
Now we have one pkl file .
./models/SVHN_annotation.pkl
Example:
python train.py
important arguments about Traning Model in config.py
Default: These arguments is basic setting for model.
| Argument | Default value |
|---|---|
| model_name | faster_rcnn_efficientnetb4_v1.pth |
| batch_size | 32 |
| workers | 4 |
| num_classes | 11 (classes+background) |
| batch_size | 32 |
| max_size | 380 (Resize image ) |
| min_size | 380 (Resize image ) |
And, these is related to your training performance.
| Argument | Default value |
|---|---|
| epochs | 18 |
| learning_rate | 0.01 |
| momentum | 0.9 |
| num_classes | 11 (classes+background) |
| weight_decay | 5e-4 |
| milestones | [5, 12] (MultiStepLR) |
| gamma | 0.1 |
When the program was finished, we will get a traning model file in /model/.
./model/faster_rcnn_efficientnetb4_v1.pth
If we want to test image, make sure we have a model in /model/ first and confirm model_name.
In addition, if you want to use my trained model, download this model to /model/ and change the "model_name" to "faster_rcnn_efficientnetb4_v3_8" in config.py
Example:
python test.py
important arguments about Testing Model in config.py
Default:
| Argument | Default value |
|---|---|
| score_threshold | 0.0 |
| IoU_threshold | 1.5 (Eliminate overlapping boxes. Normally the range is between 0 ~ 1, 1.5 meaning that the function is not enabled.) |
| json_name | 0856566.json |
When the program was finished, we will get a json file in /result/.
./result/0856566.json
Adjust a few parameters to visualize the test images with bounding box. (In pynb, it works)
important arguments about Testing Model in config.py
Default:
| Argument | Default value |
|---|---|
| IoU_threshold | 1.5 (suggest 0.2~0.5) |
| plot_img | False (True, if you want to plot testing images) |
IoU_threshold=0.3 plot_img= True:
| Metrics | value |
|---|---|
| mAP |  |
| model performance |  |