Training the Model
Developing new Menhaden age prediction models
The training scripts and routines described below remain untested as of September 2025. This page will be updated in the future as needed.
New image-only and multimodal age prediction models can be trained using the corresponding executable Python scripts, train-image-only.py and train-multimodal.py, respectively, and an accompanying configuration file, training-configs.yml. As with the model configuration file, this YAML file can be called anything as long as its internal structure is retained.
Training scripts and a sample configuration file can be found in FATES-BLH-ScaleAgeing/scripts/training_scripts.
Configuration settings
The configuration YAML file is structured as key: value pairs. The order in which these entries are presented in the file does not matter, but all keys need to be included and match exactly as expected. As before, absolute file paths are generally recommended to avoid unintended behavior but will vary from computer to computer. For example:
training-configs.yml
#-----------------------------------------------------------------------------------------------------
# Configuration for training scale ageing model
#-----------------------------------------------------------------------------------------------------
# Path to processed (cropped, padded, normalized (optional)) training images
train_image_path: "G:/Shared drives/NMFS SEFSC FATES Advanced Technology/BIOLOGY_LIFE_HISTORY_DATA/age_testing/cropped"
# Image type to be processed (should not need to be changed)
image_type: ".jpg"
# Path to CSV file with metadata with known target ages for the training data
train_csv: "train.csv"
# Path to CSV file with metadata with known target ages for the validation data
validation_csv: "validate.csv"
# Path where to store model results
model_out_path: "C:/Users/user.name/Documents/FATES-BLH-ScaleAgeing/scripts/weights"
# Number of epochs of training
epochs: 50
# Initial learning rate. Generally a small number much less than 1.
learning_rate: 0.001
# Training scheduler learning rate decay gamma. Should be less than 1.
gamma: 0.1
# Training scheduler checkpoints (learning rate multiplied by gamma at these epochs)
scheduler: [15, 35]These settings are described as follows:
| Key | Description |
|---|---|
train_image_path |
Path to processed (i.e., cropped and padded with any normalization applied) training images. Best to include the full path in quotations. Example: “G:/Shared drives/NMFS SEFSC FATES Advanced Technology/BIOLOGY_LIFE_HISTORY_DATA/age_testing/cropped” |
image_type |
File extension of image files to be used for training |
train_csv |
Path to a single CSV file containing the known age targets for all images in train_image_path that are to be used for training. (Validation examples will be indicated in a separate validation CSV, as described below.) If training a multimodal model, this file should also include the necessary metadata (length, weight, month of catch). Known ages must be in the last column, as shown below. Best to include the full path in quotations. |
validation_csv |
Path to a single CSV file containting the known age targets for all ages in train_image_path that are to be used for validation. These should all be separate from those files listed in the training CSV. If training a multimodal model, this file should also include the necessary metadata (length, weight, month of catch). Known ages must be in the last column, as shown below. Best to include the full path in quotations. |
model_out_path |
Path where to store model results. Best to include the full path in quotations. It will be created automatically if it does not already exist. |
epochs |
Number of training epochs |
learning_rate |
Initial learning rate for training. Generally a small number much less than 1. |
gamma |
Factor by which to decrease the learning rate at epoch checkpoints specified in scheduler. Should be less than 1. |
scheduler |
Training epoch number(s) at which the current learning rate is multiplied by gamma. Can be one or more integer values in increasing order, all less than epoch, and enclosed in brackets to indicate a Python list. Example: [15, 35] |
Training and Validation Subsets
The training scripts assume that all training and validation data are in a single directory passed to train_image_path above. Within this directory, the examples are split into training and validation subsets using two CSV files that each include the image file names to be used for each task. An example training CSV for an image-only model would be formatted as follows:
| image_name | age |
|---|---|
| 25120.jpg | 1 |
| 25235.jpg | 1 |
| 25349.jpg | 0 |
| … | … |
| 26120.jpg | 4 |
Similarly, an example validation CSV for a multimodal model would be formatted identically to the metadata CSV used to run the model except with a new “age” column appended to the end:
| image_name | length | weight | month | age |
|---|---|---|---|---|
| 25123.jpg | 208 | 147 | 6 | 2 |
| 25234.jpg | 219 | 176 | 5 | 2 |
| 25345.jpg | 222 | 183 | 6 | 3 |
| … | … | … | … | … |
| 26315.jpg | 211 | 168 | 5 | 4 |
The column headers can be anything, but the model expects these variables in this order. Also notice that each image file has an entry, each entry is on its own line, and that no example is contained in both the training and validation file lists. Generally, the majority of examples (say, 70%) are used for testing and a smaller subset (say, 15%) are used for validation. Any remaining examples are reserved for testing. Finally, the image file names in the first column do not need to include the file extension. If the extensions are missing, they will automatically be added using the image_type extension provided in training-configs.yml.
The column headers in metadata.csv can be called anything, but the variables must be provided in the order shown, especially for multimodal models: image name, length, weight, month of catch, age.
Usage
Training new models can be done by running either of the training scripts as follows:
Image-only Model
python train-image-only.py --config_path "path/to/training-configs.yml"Multimodal Model
python train-multimodal.py --config_path "path/to/training-configs.yml"Both of these scripts will write out a new model weights file called best_model.pth, which corresponds to the best-performing version of the model obtained during the training loop. Simply rename this file something more intuitive and use it as described in Using the Model.