The program uses several arguments to control the flow of the program. They are described in this section.
| Argument | Description | Shorthand | Type | Default | Notes |
|---|---|---|---|---|---|
--delete |
Delete files from after downloading or preprocessing them. | -e |
N/A | False | 1 |
--epochs |
The amount of epochs to train for. | -c |
Integer | 150 |
N/A |
--batches |
Sets the number of batches to train on per epoch. | -b |
Integer | 100 |
N/A |
--load-path |
The path to load images from | -p |
String | "./data/raw" |
4 |
--log |
Enables debug, or sets the output level. | -l |
String | "info" |
2 |
--mode |
Sets the mode the program is run in. | -m |
String | "generate" |
3 |
--model |
Sets the model to use for predicting or further training. | -o |
String | N/A | 5 |
--no-overwrite |
Prevents overwriting files during preprocessing. | -n |
N/A | False | N/A |
--save-path |
Where to save files after they've been downloaded or preprocessed. | -s |
String | "./data/raw" |
N/A |
--save-pickle |
Saves predicted images to pickles rather than PNG images. | None | N/A | False | N/A |
The delete flag does different things in different modes.
In the data generation mode (--mode generate), this flag ensures files are deleted from the Google Drive account after they have been downloaded.
In preprocess mode (--mode preprocess), this flag deletes the raw images from the disk after they've been preprocessed and saved as a pickle file.
The flag does nothing in the other two modes.
The log flag controls the output level of the console. Full logs are always written to file. It knows three different modes:
- If it's not present it defaults the output level to INFO level.
- If present without any value it sets the output level to DEBUG.
- If present with a value it sets the output level to the specified value.
The logging modes are as follows in reverse order of importance: debug, info, warning, error, and critical.
If you choose a mode with low importance like INFO, messages of higher importance like ERROR will still be printed.
The mode flag controls the mode the program is run in. Four modes have been implemented.
The default mode is data generation (generate), and uses Google Earth Engine to generate and acquire data files.
The preprocess mode (preprocess) prepares the generated images for use in the neural network.
The model training mode (train) trains a Generative Adversarial Model (GAN) on the acquired data. (not yet implemented)
The image prediction mode (predict) uses a trained model to predict/generate a new image from acquired data. (not yet implemented)
The load path flag informs the program where to load image files from.
Its behavior is dependent on the mode set (with the --mode flag).
In preprocessing mode (--mode preprocess) it selects the path to load raw .tif images from and preprocesses them to numpy data-cubes.
In training mode (--mode train) this is the path where the program will look for preprocessed images to use in training and testing.
In predict mode (--mode predict) this is the path where it will try to find the image to run a prediction on. The model to run in set by the --model flag.
The model flag selects which model to load.
What to do with the loaded model depends on the selected mode (--mode).
In training mode (--mode train) the loaded model is trained further, rather than training a new model from scratch.
In predict mode (--mode predict) the loaded model is used to predict images from input(s).
In the other modes (generate and preprocessing) this flag does nothing.
If the predict mode is used (--mode predict), but no load path is provided (--load-path),
the program will try to read the files to predict on from the input stream (stdin).
This means that either on the command-line or through a text file, this mode is able to work through a list of files.
The first line of the input stream should contain just the number n of files that you wish to process.
Then on the following n lines, you can enter one file to predict off per line.