chunk_onnx

Runs an Onnx model over all rasters in the input to produce a single raster. This workflow is intended to apply an Onnx model over all rasters in the input to produce a single raster output. This can be used, for instance, to compute time-series analysis of a list of rasters that span multiple times. The analysis can be any computation that can be expressed as an Onnx model (for an example, see notebooks/crop_cycles/crop_cycles.ipynb). In order to run the model in parallel (and avoid running out of memory if the list of rasters is large), the input rasters are divided spatially into chunks (that span all times). The Onnx model is applied to these chunks and then combined back to produce the final output.

graph TD inp1>rasters] out1>raster] tsk1{{chunk_raster}} tsk2{{list_to_sequence}} tsk3{{compute_onnx}} tsk4{{combine_chunks}} tsk1{{chunk_raster}} -- chunk_series/chunk --> tsk3{{compute_onnx}} tsk2{{list_to_sequence}} -- rasters_seq/input_raster --> tsk3{{compute_onnx}} tsk3{{compute_onnx}} -- output_raster/chunks --> tsk4{{combine_chunks}} inp1>rasters] -- rasters --> tsk1{{chunk_raster}} inp1>rasters] -- list_rasters --> tsk2{{list_to_sequence}} tsk4{{combine_chunks}} -- raster --> out1>raster]

Sources

rasters: Input rasters.

Sinks

raster: Result of the Onnx model run.

Parameters

model_file: An Onnx model which needs to be deployed with “farmvibes-ai local add-onnx” command.
step: Size of the chunk in pixels.

Tasks

chunk_raster: Splits input rasters into a series of chunks.
list_to_sequence: Combines a list of Rasters into a RasterSequence.
compute_onnx: Runs the onnx model across chunks of the input rasters.
combine_chunks: Combines series of chunks into a final raster.

Workflow Yaml

name: chunk_onnx
sources:
  rasters:
  - chunk_raster.rasters
  - list_to_sequence.list_rasters
sinks:
  raster: combine_chunks.raster
parameters:
  model_file: null
  step: 100
tasks:
  chunk_raster:
    op: chunk_raster
    parameters:
      step_y: '@from(step)'
      step_x: '@from(step)'
  list_to_sequence:
    op: list_to_sequence
  compute_onnx:
    op: compute_onnx_from_chunks
    op_dir: compute_onnx
    parameters:
      model_file: '@from(model_file)'
      window_size: '@from(step)'
  combine_chunks:
    op: combine_chunks
edges:
- origin: chunk_raster.chunk_series
  destination:
  - compute_onnx.chunk
- origin: list_to_sequence.rasters_seq
  destination:
  - compute_onnx.input_raster
- origin: compute_onnx.output_raster
  destination:
  - combine_chunks.chunks
description:
  short_description: Runs an Onnx model over all rasters in the input to produce a
    single raster.
  long_description: This workflow is intended to apply an Onnx model over all rasters
    in the input to produce a single raster output. This can be used, for instance,
    to compute time-series analysis of a list of rasters that span multiple times.
    The analysis can be any computation that can be expressed as an Onnx model (for
    an example, see notebooks/crop_cycles/crop_cycles.ipynb). In order to run the
    model in parallel (and avoid running out of memory if the list of rasters is large),
    the input rasters are divided spatially into chunks (that span all times). The
    Onnx model is applied to these chunks and then combined back to produce the final
    output.
  sources:
    rasters: Input rasters.
  sinks:
    raster: Result of the Onnx model run.
  parameters:
    model_file: An Onnx model which needs to be deployed with "farmvibes-ai local
      add-onnx" command.
    step: Size of the chunk in pixels.