Skip to content

Latest commit

 

History

History
77 lines (61 loc) · 3.08 KB

ml_framework.md

File metadata and controls

77 lines (61 loc) · 3.08 KB
Raw

Machine learning training and inference

Train an ML model

To train an ML model, use the following command:

bin/train.sh model_config dataset
  • model_config: Name of ML model configuration to use. This should correspond to a configuration file named config/model_config.sh.
  • dataset: Dataset identifier. Check the variables train_file and eval_file in bin/train.sh to make sure that this maps to the correct data.

For example, to run the tutorial machine learning model on the baseline dataset:

bin/train.sh cnn2d_tutorial baseline

The training logs are written to a file of the form log/job_id.log. Take note of the job_id for it will be required at inference: the trained network is saved at DATAPATH/models/job_id/.

Configure ML model parameters

The parameters for an ML task can be configured by creating a corresponding configuration file: config/your_model_config.sh. Look at other ML model configuration files in config/ for examples.

Create a new ML model architecture

  • Create a new your_model.py file inside the ml_framework/model folder. Look at other models inside the folder for examples.
  • Reference your new model in ml_framework/model/__init__.py.
  • Set the model argument to your new model's name in your model configuration file config/your_model_config.sh.

Evaluate a ML model

Once the model is trained, evaluate the model's performance:

bin/evaluate.sh model_config dataset job_id
  • model_config: Name of ML model configuration to use. This should correspond to a configuration file named config/model_config.sh.
  • dataset: Dataset identifier. Check the variable eval_file in bin/evaluate.sh to make sure that this maps to the correct data.
  • job_id: The job identifier of the machine learning training from which to load the trained network.

This evaluates the model's performance on the test dataset and saves the corresponding logits to the following file: DATAPATH/models/job_id/eval_logits.npy The logits are saved in the same order as provided in the input data pipeline, which means that they correspond to each line of the TFRecord manifest file.

Run inference on continuous data

Once the model is trained, run inference on continuous data:

bin/predict.sh model_config dataset job_id
  • model_config: Name of ML model configuration to use. This should correspond to a configuration file named config/model_config.sh.
  • dataset: Dataset identifier. Check the variable test_file in bin/predict.sh to make sure that this maps to the correct data.
  • job_id: The job identifier of the machine learning training from which to load the trained network.

test_file should be a Unix glob pattern that matches the continuous data Numpy files on which to run inference. For each continuous data file path_to_continuous_data/filename.npy, a sliding window runs through the data and the correspoding logits are saved to path_to_continuous_data/filename_logits.npy. The amount of overlap between sliding windows can be set by adding an overlap argument to the model configuration.