I'm working on a SageMaker labeling job with custom datatypes. For some reason though, I'm not getting the correct label in the AWS web console. It should have the selected label which is "Native", but instead, I'm getting the <labelattributename> which is "new-test-14".
After Ground Truth runs the post-annotation lambda, it seems to modify the metadata before returning a data object. The data object it returns doesn't contain a class-name key inside the metadata attribute, even when I hard-code the lambda to return an object that contains it.
My manifest file looks like this:
{"source-ref" : "s3://<file-name>", "text" : "Hello world"}
{"source-ref" : "s3://"<file-name>", "text" : "Hello world"}
And the worker response looks like this:
{"answers":[{"acceptanceTime":"2021-05-18T16:08:29.473Z","answerContent":{"new-test-14":{"label":"Native"}},"submissionTime":"2021-05-18T16:09:15.960Z","timeSpentInSeconds":46.487,"workerId":"private.us-east-1.ea05a03fcd679cbb","workerMetadata":{"identityData":{"identityProviderType":"Cognito","issuer":"https://cognito-idp.us-east-1.amazonaws.com/us-east-1_XPxQ9txEq","sub":"edc59ce1-e09d-4551-9e0d-a240465ea14a"}}}]}
That worker response gets processed by my post-annotation lambda which is modeled after this aws sample ground truth recipe. Here's my code:
import json
import sys
import boto3
from datetime import datetime
def lambda_handler(event, context):
# Event received
print("Received event: " + json.dumps(event, indent=2))
labeling_job_arn = event["labelingJobArn"]
label_attribute_name = event["labelAttributeName"]
label_categories = None
if "label_categories" in event:
label_categories = event["labelCategories"]
print(" Label Categories are : " + label_categories)
payload = event["payload"]
role_arn = event["roleArn"]
output_config = None # Output s3 location. You can choose to write your annotation to this location
if "outputConfig" in event:
output_config = event["outputConfig"]
# If you specified a KMS key in your labeling job, you can use the key to write
# consolidated_output to s3 location specified in outputConfig.
# kms_key_id = None
# if "kmsKeyId" in event:
# kms_key_id = event["kmsKeyId"]
# # Create s3 client object
# s3_client = S3Client(role_arn, kms_key_id)
s3_client = boto3.client('s3')
# Perform consolidation
return do_consolidation(labeling_job_arn, payload, label_attribute_name, s3_client)
def do_consolidation(labeling_job_arn, payload, label_attribute_name, s3_client):
"""
Core Logic for consolidation
:param labeling_job_arn: labeling job ARN
:param payload: payload data for consolidation
:param label_attribute_name: identifier for labels in output JSON
:param s3_client: S3 helper class
:return: output JSON string
"""
# Extract payload data
if "s3Uri" in payload:
s3_ref = payload["s3Uri"]
payload_bucket, payload_key = s3_ref.split('/',2)[-1].split('/',1)
payload = json.loads(s3_client.get_object(Bucket=payload_bucket, Key=payload_key)['Body'].read())
# print(payload)
# Payload data contains a list of data objects.
# Iterate over it to consolidate annotations for individual data object.
consolidated_output = []
success_count = 0 # Number of data objects that were successfully consolidated
failure_count = 0 # Number of data objects that failed in consolidation
for p in range(len(payload)):
response = None
dataset_object_id = payload[p]['datasetObjectId']
log_prefix = "[{}] data object id [{}] :".format(labeling_job_arn, dataset_object_id)
print("{} Consolidating annotations BEGIN ".format(log_prefix))
annotations = payload[p]['annotations']
# print("{} Received Annotations from all workers {}".format(log_prefix, annotations))
# Iterate over annotations. Log all annotation to your CloudWatch logs
annotationsFromAllWorkers = []
for i in range(len(annotations)):
worker_id = annotations[i]["workerId"]
anotation_data = annotations[i]["annotationData"]
annotation_content = anotation_data["content"]
annotation_content_json = json.loads(annotation_content)
annotation_job = annotation_content_json["new_test"]
annotation_label = annotation_job["label"]
consolidated_annotation= {
"workerId": worker_id,
"annotationData": {
"content": {
"annotatedResult": {
"instances": [{"label":annotation_label }]
}
}
}
}
annotationsFromAllWorkers.append(consolidated_annotation)
consolidated_annotation = {"annotationsFromAllWorkers": annotationsFromAllWorkers} # TODO : Add your consolidation logic
# Build consolidation response object for an individual data object
response = {
"datasetObjectId": dataset_object_id,
"consolidatedAnnotation": {
"content": {
label_attribute_name: consolidated_annotation,
label_attribute_name+ "-metadata": {
"class-name": "Native",
"confidence": 0.00,
"human-annotated": "yes",
"creation-date": datetime.strftime(datetime.now(), "%Y-%m-%dT%H:%M:%S"),
"type": "groundtruth/custom"
}
}
}
}
success_count += 1
# print("{} Consolidating annotations END ".format(log_prefix))
# Append individual data object response to the list of responses.
if response is not None:
consolidated_output.append(response)
failure_count += 1
print(" Consolidation failed for dataobject {}".format(p))
print(" Unexpected error: Consolidation failed." + str(sys.exc_info()[0]))
print("Consolidation Complete. Success Count {} Failure Count {}".format(success_count, failure_count))
print(" -- Consolidated Output -- ")
print(consolidated_output)
print(" ------------------------- ")
return consolidated_output
As you can see above, the do_consolidation method returns an object hard-coded to include a class-name of "Native", and the lambda_handler method returns that same object. Here's the post-annotation function response:
[{
"datasetObjectId": "4",
"consolidatedAnnotation": {
"content": {
"new-test-14": {
"annotationsFromAllWorkers": [{
"workerId": "private.us-east-1.ea05a03fcd679cbb",
"annotationData": {
"content": {
"annotatedResult": {
"instances": [{
"label": "Native"
}]
}
}
}
}]
},
"new-test-14-metadata": {
"class-name": "Native",
"confidence": 0,
"human-annotated": "yes",
"creation-date": "2021-05-19T07:06:06",
"type": "groundtruth/custom"
}
}
}
}]
As you can see, the post-annotation function return value has the class-name of "Native" in the metadata so I would expect the class-name to be present in the data object metadata, but it's not. And here's a screenshot of the data object summary:
It seems like Ground Truth overwrote the metadata, and now the object doesn't contain the correct label. I think perhaps that's why my label is coming through as the label attribute name "new-test-14" instead of as the correct label "Native". Here's a screenshot of the labeling job in the AWS web console:
The web console is supposed to show the label "Native" inside the "Label" column but instead I'm getting the <labelattributename> "new-test-14" in the label column.
Here is the output.manifest file generated by Ground Truth at the end:
{
"source-ref": "s3://<file-name>",
"text": "Hello world",
"new-test-14": {
"annotationsFromAllWorkers": [{
"workerId": "private.us-east-1.ea05a03fcd679ert",
"annotationData": {
"content": {
"annotatedResult": {
"label": "Native"
}
}
}
}]
},
"new-test-14-metadata": {
"type": "groundtruth/custom",
"job-name": "new-test-14",
"human-annotated": "yes",
"creation-date": "2021-05-18T12:34:17.400000"
}
}
What should I return from the Post-Annotation function? Am I missing something in my response? How do I get the proper label to appear in the AWS web console?