- Once this is done, the script will create a hyperparameters.json file that can be edited to put the values you want for the training
- Once this is done, the script will create a hyperparameters.json file that can be edited to put the values you want for the training
- There you will find all parameters such as learning rates, batch size, number of epochs max (early stopping should stop before), the loss weight (how much impact the generator loss will have comparing to the discriminator loss),the tile size, the patience for the learning rate scheduler and early stopping and the validation set proportion.
- There you will find all parameters such as **learning_rates**, **batch size**, **number of epochs** max (early stopping should stop before), the **loss weight** (how much impact the generator loss will have comparing to the discriminator loss),the **tile size**, the **patience** for the learning rate scheduler, **early stopping** and the **validation set proportion**.
Here follows a sample hyperparameters.json file.
```json
{
"learning_rate_generator":1e-2,
"learning_rate_discriminator":1e-2,
"batch_size":64,
"num_epochs":1000,
"loss_weight":10,
"tile_size":256,
"patience":20,
"valid_size":0.1,
"data_paths":{
"clean":"./data/internship_GT",
"noisy":"./data/internship_low",
"test":"./data/test"
}
}
```
If you hardware is not able to run the code properly due to memory outage try lowering both tile size and batch size.
- The training script will just need the experiment path (so experiment folder + experiment name) and use the information provided earlier
- The training script will just need the experiment path (so experiment folder + experiment name) and use the information provided earlier
- The prediction script will predict all images located at the test data path location. If not provided in the new experiment script, you can still provide the path directly in the hyperparameters.json file.
- The prediction script will predict all images located at the test data path location. If not provided in the new experiment script, you can still provide the path directly in the hyperparameters.json file.
- A complete training on the provided dataset was done on a L40 card with a duration of 7 hours.
- A complete training on the provided dataset was done on a L40 card with a duration of 7 hours.
## 3 - Analysis : Shape classification and quantification
Once the segmentation is done, the segmentation result is passed to Fiji's MorpholibJ plugin to further post processing and 3D component labelling of objects (See paper Material & Methods section for more details).
A CSV file is generated and passed to the *Analysis/analysis.ipynb* where all shape classification and quantification are done and Figures are generated.
*__Running time__* : Depending on your hardware, the running time could vary a lot. The size of the dataset for training will also affect the training time.
*__Running time__* : Depending on your hardware, the running time could vary a lot. The size of the dataset for training will also affect the training time.
## 4 - License
## 3 - License
This code repository is release under the [CC-SA License ??](https://gitlab-lis-lab.fr/sicomp/mupix/LICENSE???)
This code repository is release under the [CC-SA License ??](https://gitlab-lis-lab.fr/sicomp/mupix/LICENSE???)
