Writing a new training pipeline

This tutorial will walk you through how to set up your first custom training pipeline. The goal of this example pipeline is to orchestrate a classifier that can distinguish between daisies and roses.

In this tutorial we will - learn which configuration yaml files we have to write to get our pipeline running - have a look how the dataset, training, prediction, etc. configurations are connected with each other - train a binary classifier with niceML

Prerequisites

Before we get started, please ensure that you have followed the Getting Started tutorial and initialized a new niceML project.

Step 1: Data ingestion

Download the images from Kaggle and store the roses and daisies images in a data folder in your niceML project.

Step 2: Configure the pipeline job

Let us start configuring at the highest level of niceML: The job configuration. For this, navigate to configs/jobs/job_train/job_train_cls and create a new file job_train_cls_binary_flowers.yaml.

Paste the content of configs/jobs/job_train/job_train_cls/job_train_cls_binary.yaml which is a good starting point for our own pipeline.

Adjust job parameters in job_train_cls_binary_flowers.yaml. Many parameters are pre-configured and do not require adjustment. However, we want to configure our own - train and prediction ops (/ops/train@ops.train.config, /ops/train@ops.prediction.config), - experiment_id (globals -> exp_name and exp_prefix) - and input data location (globals -> data_location).

In the job configuration, you can already point to the location where the Kaggle images are stored. Also, you can change the path to the dagster train and prediction operation configurations you are going to write in the next two steps.

# train binary classification
defaults:
  # experiment
  - ops/experiment@ops.experiment.config: op_experiment_default.yaml
  # train
  - /ops/train@ops.train.config: op_train_cls_binary_flowers.yaml
  # prediction
  - /ops/prediction@ops.prediction.config: op_prediction_cls_binary_flowers.yaml
  # analysis
  - /ops/analysis@ops.analysis.config: op_analysis_cls_binary.yaml
  # experiment tests
  - /ops/exptests@ops.exptests.config.tests: exptests_default.yaml
  # experiment locations
  - shared/locations@globals: exp_locations.yaml
  - _self_

hydra:
  searchpath:
    - file://configs

globals:
  exp_name: FlowersClsBinary
  exp_prefix: FLOWERSCLB
  data_location:
    uri: ${oc.env:DATA_URI,./data}/kaggle_flowers

Step 3: Configure the train operation

Navigate to configs/ops/train and create op_train_cls_binary_flowers.yaml.

You can use the content of op_train_cls_binary_flowers.yaml as a starting point and change the definition where the training data is stored, what the image size is and which classes you use as labels.

This means, you have to edit - /shared/datasets@data_train - /shared/datasets@data_validation - classes - target_size width and target_size height

For the time being, use the OwnMobileNetModel model of Tensorflow and do not implement your own model definition.

defaults:
  - op_train_base.yaml@_here_
  - /shared/datasets@data_train: dataset_kaggle_flowers.yaml
  - /shared/datasets@data_validation: dataset_kaggle_flowers.yaml
  - _self_
train_params:
  epochs: 5
data_description:
  _target_: niceml.data.datadescriptions.clsdatadescription.ClsDataDescription
  classes:
  - "daisy"
  - "roses"
  use_binary: true
  target_size:
    _target_: niceml.utilities.imagesize.ImageSize
    width: 180
    height: 180

data_train:
  datainfo_listing:
    sub_dir: train
  set_name: train
  shuffle: true
data_validation:
  datainfo_listing:
    sub_dir: validation
  set_name: validation
model:
  _target_: niceml.dlframeworks.keras.models.mobilenet.OwnMobileNetModel
  final_activation: sigmoid
model_load_custom_objects:
  _target_: niceml.mlcomponents.modelcompiler.modelcustomloadobjects.ModelCustomLoadObjects
learner:
  _target_: niceml.dlframeworks.keras.learners.defaultlearner.DefaultLearner
  model_compiler:
    _target_: niceml.dlframeworks.keras.modelcompiler.defaultmodelcompiler.DefaultModelCompiler
    loss: binary_crossentropy
    # Optimizer used in the experiment
    optimizer:
      _target_: tensorflow.keras.optimizers.Adam
      learning_rate: 0.0001
    metrics: [ "accuracy" ]   

You might have noticed, that you did not give a plain directory definition to tell niceml where the training images are stored. Instead, you define the datasets within a configuration yaml in configs/shared/datasets/dataset_kaggle_flowers.yaml. For now, use the same configuration as in dataset_cls_test.yaml

Step 4: Configure the prediction operation

Create configs/ops/prediction/op_prediction_cls_binary_flowers.yaml and use op_prediction_cls.yaml as a template.

Here, only change the definition of the training, validation and test datasets with the yaml we wrote in the step before (dataset_kaggle_flowers.yaml). The prediction configuration should look like this:

      defaults:
   - /shared/datasets@datasets.validation: dataset_kaggle_flowers.yaml
   - /shared/datasets@datasets.test: dataset_kaggle_flowers.yaml
   - /shared/datasets@datasets.train_eval: dataset_kaggle_flowers.yaml
   - prediction_handler: prediction_handler_vector.yaml
   - datasets: datasets_generic_default.yaml
   - op_prediction_base.yaml@_here_
   - _self_

(Optional) Step 5: Add a custom model

If you want to add a custom model have a look at the How to add a custom model.

Step 6: Start the pipeline

Run niceML and hand over the top-level jon configuration file via

niceml train configs/jobs/job_train/job_train_cls/job_train_cls_binary_flowers.yaml