Add python code block to workflows
#509
Conversation
inference/core/workflows/execution_engine/dynamic_blocs/block_scaffolding.py
Outdated
Show resolved
Hide resolved
| WorkflowBlockManifest, | ||
| ) | ||
|
|
||
| IMPORTS_LINES = [ |
There was a problem hiding this comment.
If your custom code block needs other available dependencies (eg time, json, os, requests, cv2, shapely, our CLIP Model, etc) is there a way to add them to the import list?
There was a problem hiding this comment.
one may import more in the string with python code, like:
import cv2
# any other import
def function(x, y, z):
passwe could also extend the imports, yet for instance CLIP should probably be used via model manager as manual init in multiple blocks may end up OOM on multiple instances on the same model. We can obviously add more imports which I will do, the question is what we suggest people regarding model usage - I believe it would be better to suggest using models within existing blocks if possible and only use model in-place when there is no other option
There was a problem hiding this comment.
ok, added separate field in manifest to declare optional imports - that is going to give us more visibility
| description="Definition of manifest for dynamic block to be created in runtime by " | ||
| "workflows execution engine." | ||
| ) | ||
| code: PythonCode = Field( |
There was a problem hiding this comment.
Probably also needs an init function
There was a problem hiding this comment.
👍 changed the way how it will work to have init function and run function with self injected as first param which would have access to self._init_results
| description="Accepted dimensionality offset for parameter. Dimensionality works the same as for " | ||
| "traditional workflows blocks.", | ||
| ) | ||
| selector_types: List[SelectorType] = Field( |
There was a problem hiding this comment.
How do you accept an image (whether it be an output of a previous step or an input to the workflow)?
There was a problem hiding this comment.
u choose union of
class SelectorType(Enum):
INPUT_IMAGE = "input_image"
INPUT_PARAMETER = "input_parameter"
STEP_OUTPUT = "step_output"
which is almost 1-1 mapping to types: WorkflowImageSelector, WorkflowParameterSelector, StepOutputSelector
PawelPeczek-Roboflow
requested review from
grzegorz-roboflow and
probicheaux
as code owners
There was a problem hiding this comment.
Probably the file name should be test_dimension_collapse.py
Description
This is the preview of custom python code block in
workflows- so far implemented quite ugly - but showcasing what will be possible.Methodology:
init()function to define state)"steps"listPOST /workflows/blocks/describewhich takes optional payload with list of blocks - once it is done -dynamic blocks schemas and connections are generated into output -GET /workflows/blocks/describeremained, yet deprecated to be removed laterGET /workflows/blocks/describeandPOST /workflows/blocks/describewill return OpenAPI schema for manifest of dynamic blockType of change
Please delete options that are not relevant.
How has this change been tested, please provide a testcase or example of how you tested the change?
Any specific deployment considerations
Docs