isciml: train

Overview

The train subcommand in isciml is used for training machine learning models on the data generated from the generate command. This command allows users to configure various aspects of the training process, including model architecture, training parameters, and data handling.

Usage

singularity exec isciml.sif isciml train [OPTIONS]

Options

Option	Description	Default	Required
--sample_folder PATH	Folder containing sample files	-	Yes
--target_folder PATH	Folder containing target files	-	Yes
--n_blocks INTEGER	Number of blocks in UNet	4	No
--start_filters INTEGER	Number of start filters	32	No
--batch_size INTEGER	Batch size for training	1	No
--max_epochs INTEGER	Maximum number of epochs	1	No
--learning_rate FLOAT	Adam optimizer learning rate	0.001	No
--save_model PATH	File name to save the checkpoint at the end of training	pytorch_model.ckpt	No
--load_model PATH	Checkpoint file name to load at the beginning of training	-	No
--train_size FLOAT	Training size (proportion of data used for training)	0.8	No
--num_workers INTEGER	Number of workers for data loader	1	No
--n_gpus INTEGER	Number of GPUs used for training	1	No
--strategy TEXT	Distributed Data Parallel Strategy	auto	No
--checkpoint_folder PATH	Checkpoint folder	./lightning_checkpoint_folder	No
--every_n_epochs INTEGER	Number of epochs between checkpoints	-	No
--save_top_k INTEGER	Number of best models to save	1	No
`--reshape_base [two	eight]`	Reshape 1D to 2D using base 2 or 8	eight
--dim INTEGER	Dimension of the grid of solution	2	No
--help	Show the help message and exit	-	No

Description

The train subcommand allows you to train a UNet model on your generated data. It provides various options to customize the training process, including model architecture, training parameters, and data handling.

Example Usage

To train a model with default settings:

singularity exec isciml.sif isciml train \
    --sample_folder /path/to/samples \
    --target_folder /path/to/targets

To train a more complex model with custom settings:

singularity exec isciml.sif isciml train \
    --sample_folder /path/to/samples \
    --target_folder /path/to/targets \
    --n_blocks 5 \
    --start_filters 64 \
    --batch_size 32 \
    --max_epochs 100 \
    --learning_rate 0.0001 \
    --save_model custom_model.ckpt \
    --train_size 0.7 \
    --n_gpus 2 \
    --strategy ddp \
    --every_n_epochs 5 \
    --save_top_k 3

Notes

1. The sample and target folders should contain the input and output data generated using the generate command.

2. The UNet architecture is defined by n_blocks and start_filters. Increasing these values will create a more complex model.

3. batch_size, max_epochs, and learning_rate are key training parameters that affect the learning process.

4. The save_model option specifies where the final trained model will be saved.

5. Use load_model to continue training from a previously saved checkpoint.

6. train_size determines the split between training and validation data.

7. n_gpus and strategy are used for distributed training across multiple GPUs.

8. The checkpointing system saves models periodically and can keep the top K models based on validation performance.

9. reshape_base and dim options are used to reshape the input data, which may be necessary depending on your data format.

Related Commands

• generate-models: Used to create the physical models.

• generate: Used to create the training data.

• inference: Used to apply the trained model to new data.

See Also

For more information on model architectures, training strategies, and hyperparameter tuning, refer to the isciml documentation on machine learning models and training processes.