| # Thundernet |
| Thundernet is a semantic segmentation model that processes RGB input using convolutional networks to extract key features. |
|
|
| ### USE OF THE REPOSITORY - TRAINING |
|
|
| To execute the train file: python train_config.py |
| |
| You can change the parameters adding them as flags in the previous command or changing them directly in the Thundernet_config.py file |
| The parameters that you can include (that are relevant for training) are: |
| - **train_path (str)**: path to the train data (default: "C:/Users/user/Documents/pruned_training/training/") |
| - **val_path (str)**: path to the val data (default: "C:/Users/user/Documents/pruned_training/val/") |
| - **model_dir (str)**: path to save the trained model (default: "C:/Users/user/Documents/Thundernet/pruebas_modelos/" ) |
| - **model_weights (str)**: NOT APPLICABLE FOR TRAINING path to the trained weights (default: "C:/Users/user/Documents/Thundernet/pruebas_modelos/32_ppm/BS4_lossBCE_weights_lr_0.00013713842558297858_reg-1.1743577101671763e-05-ep-13-val_loss0.11463435739278793-train_loss0.053004469722509384-val_iou0.8959722518920898-train_iou0.9606077075004578.hdf5") |
| - **batch_size (int)**: batch size (default: 4) |
| - **loss (str)**: type of loss to use (default: "BCE") |
| - **classes (int)**: number of classes (default: 2). The original purpose of the code was to segment egocentric bodies, so two classes were involved (body and background) |
| - **pretrained (bool)**: start from a pretrained model (default: False) |
| - **pretrained_weights (str)**: path to the pretrained model (default: None) |
| - **lr (float)**: learning rate (default: 1e-4) |
| - **epochs (int)**: number of epochs (default: 15) |
| - **resolution (str)**: resolution of input images (defualt: 640x480) |
| - **kernel_regularizer (float)**: kernel regularizer (default: 2e-4) |
| |
| Note: there are more parameters in the file, but they are no use for training |
| |
| |
| ### USE OF THE REPOSITORY - EVALUATION |
| |
| To execute the evaluation, you can execute: python inference_config.py |
|
|
| You can change the parameters adding them as flags in the previous command or changing them directly in the Thundernet_config2.py file |
| (that are relevant for evaluation) are: |
| - **model_weights (str)**: NOT APPLICABLE FOR TRAINING path to the trained weights (default: "C:/Users/user/Documents/Thundernet/pruebas_modelos/32_ppm/BS4_lossBCE_weights_lr_0.00013713842558297858_reg-1.1743577101671763e-05-ep-13-val_loss0.11463435739278793-train_loss0.053004469722509384-val_iou0.8959722518920898-train_iou0.9606077075004578.hdf5") |
| - **batch_size (int)**: batch size (default: 4) |
| - **resolution (str)**: resolution of input images (defualt: 640x480) |
| |
| Note: there are more parameters in the file, but they are no use for evaluation |
| |
| In the inference_config.py file you can execute the main with a "show=True" to display some predictions. However, you will have to close the image after every prediction. |
| |
| |
| ### DATA PREPARATION |
| The data must be stored in a path with the following structure: |
| |
| - data |
| + images (folder with RGB images in a .jpg format) |
| -- example1.jpg |
| + labels (folder with label images in a .png format) |
| -- example1.png |
| |
| The name for the files should be the same for the image and it's corresponding label. |
| |
| |
| ### OPTIMIZE HYPERPARAMETERS |
| To find the best hyperparameters for training, you can execute the train_optuna.py file. This file will find the optimal values for batch_size, learning_rate and kernel_regularizer. The rest of the hyperparameters values will be obtained from the "thundernet_config.py" file. |
| |
| |
| #### Compare models |
| To compare if two models are the same (the weights are equal), the file "compare_models.py" can be executed. The path to the model's weights should be changed in the "weights_path1" and "weights_path2" variables. |
