parameters.yml

pytorch_model:
  =: torch.nn.Sequential
  _:
    - =: ModuleConcat # -> [(5 + 5), 28, 28]
      _:
        - { =: TensorConv2d, in_channels: 1, out_channels: 5, kernel_size: 3 }
        - { =: TensorConv2d, in_channels: 1, out_channels: 5, kernel_size: 5 }
    - { =: torch.nn.ReLU, _ }
    - =: ModuleConcat # -> [(10 + 10), 14, 14]
      _:
        - { =: TensorMaxPool2d, stride: 2, kernel_size: 3 }
        - { =: TensorConv2d, stride: 2, in_channels: 10, out_channels: 10, kernel_size: 3 }
    - { =: torch.nn.ReLU, _ }
    - =: ModuleConcat # -> [(10 + 10), 7, 7]
      _:
        - { =: TensorConv2d, stride: 2, in_channels: 20, out_channels: 10, kernel_size: 3 }
        - { =: TensorConv2d, stride: 2, in_channels: 20, out_channels: 10, kernel_size: 5 }
    - { =: torch.nn.ReLU, _ }
    - =: ModuleConcat # -> [(20 + 20), 4, 4]
      _:
        - { =: TensorMaxPool2d, stride: 2, kernel_size: 3 }
        - { =: TensorConv2d, stride: 2, in_channels: 20, out_channels: 20, kernel_size: 3 }
    - { =: torch.nn.ReLU, _ }
    - { =: pipelinex.TensorFlatten, _ } # -> [640]
    - { =: torch.nn.Dropout, _ }
    - { =: torch.nn.Linear, in_features: 640, out_features: 10 }
    - { =: torch.nn.LogSoftmax, dim: -1 }

#pytorch_model:  # Resnet18
#  =: torch.nn.Sequential
#  _:
#    - {=: torch.nn.Conv2d, in_channels: 1, out_channels: 3, kernel_size: 3}
#    - {=: torch.nn.ReLU, _: }
#    - {=: torchvision.models.resnet.resnet18, num_classes: 10}
#    - {=: torch.nn.LogSoftmax, dim: -1}

#pytorch_model:   # Resnet9
#  =: torch.nn.Sequential
#  _:
#
#    - {=: torch.nn.Conv2d, in_channels: 1, out_channels: 3, kernel_size: 3}
#    - {=: torch.nn.ReLU, _: }
#
#    - =: torchvision.models.resnet._resnet
#      arch: resnet9
#      block: {=: torchvision.models.resnet.BasicBlock}
#      layers: [1, 1, 1, 1]
#      pretrained: False
#      progress:  # None
#      num_classes: 10
#
#    - {=: torch.nn.LogSoftmax, dim: -1}

#pytorch_model:
#  =: torch.nn.Sequential
#  _:
#    - {=: torch.nn.Conv2d, in_channels: 1, out_channels: 16, kernel_size: 3}  # (1, 28, 28) -> (16, 26, 26)
#    - {=: torch.nn.ReLU, _: }
#    - {=: torch.nn.AdaptiveMaxPool2d, output_size: 13}
#    - {=: torch.nn.Conv2d, in_channels: 16, out_channels: 32, kernel_size: 3}  # (16, 13, 13) -> (32, 11, 11)
#    - {=: torch.nn.ReLU, _: }
#    - {=: torch.nn.AdaptiveMaxPool2d, output_size: 6}
#    - {=: torch.nn.Conv2d, in_channels: 32, out_channels: 64, kernel_size: 3}  # (32, 6, 6) -> (64, 4, 4)
#    - {=: torch.nn.ReLU, _: }
#    - {=: torch.nn.AdaptiveMaxPool2d, output_size: 1}  # (64, 4, 4) -> (64, 1, 1)
#    - {=: pipelinex.TensorFlatten, _: }
#    - {=: torch.nn.Dropout, p: 0.2}
#    - {=: torch.nn.Linear, in_features: 64, out_features: 10}
#    - {=: torch.nn.LogSoftmax, dim: -1}

#pytorch_model:  # https://github.com/slundberg/shap/blob/master/notebooks/deep_explainer/PyTorch%20Deep%20Explainer%20MNIST%20example.ipynb
#  =: torch.nn.Sequential
#  _:
#    - {=: torch.nn.Conv2d, in_channels: 1, out_channels: 10, kernel_size: 5}
#    - {=: torch.nn.MaxPool2d, kernel_size: 2}
#    - {=: torch.nn.ReLU, _: }
#    - {=: torch.nn.Conv2d, in_channels: 10, out_channels: 20, kernel_size: 5}
#    - {=: torch.nn.Dropout, _: }
#    - {=: torch.nn.MaxPool2d, kernel_size: 2}
#    - {=: torch.nn.ReLU, _: }
#    - {=: pipelinex.TensorFlatten, _: }
#    - {=: torch.nn.Linear, in_features: 320, out_features: 50}
#    - {=: torch.nn.ReLU, _: }
#    - {=: torch.nn.Dropout, _: }
#    - {=: torch.nn.Linear, in_features: 50, out_features: 10}
#    - {=: torch.nn.LogSoftmax, dim: -1}

#pytorch_model:  # https://github.com/slundberg/shap/blob/master/notebooks/deep_explainer/PyTorch%20Deep%20Explainer%20MNIST%20example.ipynb
#  =: torch.nn.Sequential
#  _:
#    - {=: torch.nn.Conv2d, in_channels: 1, out_channels: 10, kernel_size: 5}
#    - {=: torch.nn.MaxPool2d, kernel_size: 2}
#    - {=: torch.nn.ReLU, _: }
#    - {=: torch.nn.Conv2d, in_channels: 10, out_channels: 20, kernel_size: 5}
#    - {=: torch.nn.Dropout, _: }
#    - {=: torch.nn.MaxPool2d, kernel_size: 2}
#    - {=: torch.nn.ReLU, _: }
#    - {=: pipelinex.TensorFlatten, _: }
#    - {=: torch.nn.Linear, in_features: 320, out_features: 50}
#    - {=: torch.nn.ReLU, _: }
#    - {=: torch.nn.Dropout, _: }
#    - {=: torch.nn.Linear, in_features: 50, out_features: 10}
#    - {=: torch.nn.LogSoftmax, dim: -1}

#pytorch_model:
#  =: torch.nn.Sequential
#  _:
#    - {=: TensorConv2d, in_channels: 1, out_channels: 10, kernel_size: 5}  # -> [10, 28, 28]
#    - {=: torch.nn.ReLU, _: }
#    - {=: TensorMaxPool2d, stride: 2, kernel_size: 3}  # -> [10, 14, 14]
#    - {=: torch.nn.ReLU, _: }
#    - {=: TensorConv2d, stride: 2, in_channels: 10, out_channels: 20, kernel_size: 5}  # -> [20, 7, 7]
#    - {=: torch.nn.ReLU, _: }
#    - {=: TensorMaxPool2d, stride: 2, kernel_size: 3}   # -> [20, 4, 4]
#    - {=: torch.nn.ReLU, _: }
#    - {=: pipelinex.TensorFlatten, _: }   # -> [320]
#    - {=: torch.nn.Dropout, _: }
#    - {=: torch.nn.Linear, in_features: 320, out_features: 10}
#    - {=: torch.nn.LogSoftmax, dim: -1}

train_params:
  epochs: 1 # number of epochs to train
  time_limit: 32400 # sec
  model_checkpoint: { =: pipelinex.extras.ops.ignite.handlers.flexible_checkpoint.FlexibleModelCheckpoint }
  model_checkpoint_params:
    dirname: ../checkpoint
    filename_prefix: "%Y-%m-%dT%H-%M-%S"
    offset_hours: 8
    n_saved: 1
    atomic: True
    require_empty: True
    create_dir: True
    save_as_state_dict: False
  early_stopping_params:
    metric: loss
    minimize: True
    patience: 1000
  scheduler: { =: ignite.contrib.handlers.param_scheduler.LinearCyclicalScheduler }
  #  scheduler: {=: ignite.contrib.handlers.param_scheduler.CosineAnnealingScheduler}
  scheduler_params:
    param_name: lr
    start_value: 0.001
    end_value: 0.01
    cycle_epochs: 2 # cycle_size: int(cycle_epochs * len(train_loader))
    cycle_mult: 1.0
    start_value_mult: 1.0
    end_value_mult: 1.0
    save_history: False
  optimizer: { =: torch.optim.Adam }
  optimizer_params:
    weight_decay: { =: operator.truediv, _: [0.00001, { $: train_params.train_data_loader_params.batch_size }] }
  loss_fn: { =: torch.nn.functional.nll_loss }
  evaluation_metrics:
    accuracy: { =: ignite.metrics.Accuracy, _ }
    loss:
      =: ignite.metrics.Loss
      loss_fn: { =: torch.nn.functional.nll_loss }
  train_data_loader_params:
    batch_size: 1000 # input batch size for training
    num_workers: 2
  val_data_loader_params:
    batch_size: 1000 # input batch size for validation
    num_workers: 2
  # evaluate_train_data: EPOCH_COMPLETED # Temporarily commented out to avoid AttributeError
  # evaluate_val_data: EPOCH_COMPLETED # Temporarily commented out to avoid AttributeError
  progress_update: True
  seed: 0 #
  train_dataset_size_limit: 10000 # Only for prototyping; Remove to use the whole data
  val_dataset_size_limit: # None

test_batch_size: 1000

RUN_CONFIG:
  pipeline_name: __default__
  only_missing: False
  runner: SequentialRunner # None
  tags: # None
  node_names: # None
  from_nodes: # None
  to_nodes: # None
  from_inputs: # None
  load_versions: # None

PIPELINES:
  __default__:
    =: pipelinex.FlexiblePipeline
    module: # None
    decorator: pipelinex.log_time
    nodes:
      - inputs: parameters
        func: demo.mnist_demo.generate_datasets
        outputs:
          - train_dataset
          - val_dataset

      - inputs:
          - "params:pytorch_model"
          - train_dataset
          - val_dataset
        func:
          =: pipelinex.NetworkTrain
          train_params: { $: train_params }
        outputs: model

      - inputs:
          - model
          - val_dataset # test_dataset in actual use
        func: demo.mnist_demo.infer
        outputs: submission

      - inputs:
          - model
          - train_dataset
          - val_dataset
        func:
          =: pipelinex.ExplainModel
          train_data_loader_params: { batch_size: 100 }
          val_data_loader_params: { batch_size: 3 }
          output_transform:
            =: torch.nn.Sequential
            _:
              - { =: pipelinex.TensorExp, _ }
              - { =: pipelinex.TensorSlice, start: 0, end: 10 }
        outputs: shap_image

HOOKS:
  - =: pipelinex.MLflowBasicLoggerHook # Configure and log duration time for the pipeline 
    enable_mlflow: True # Enable configuring and logging to MLflow
    uri: sqlite:///mlruns/sqlite.db
    experiment_name: experiment_001
    artifact_location: ./mlruns/experiment_001
    offset_hours: 0 # Specify the offset hour (e.g. 0 for UTC/GMT +00:00) to log in MLflow

  - =: pipelinex.MLflowArtifactsLoggerHook # Log artifacts of specified file paths and dataset names
    enable_mlflow: True # Enable logging to MLflow
    filepaths_before_pipeline_run: # Optionally specify the file paths to log before pipeline is run
      - conf/base/parameters.yml
    datasets_after_node_run: # Optionally specify the dataset names to log after the node is run
      - model
    filepaths_after_pipeline_run: # None  # Optionally specify the file paths to log after pipeline is run

#
# Command to run mlflow server:
# $ mlflow server --host 0.0.0.0 --backend-store-uri sqlite:///mlruns/sqlite.db --default-artifact-root ./mlruns/experiment_001