Validation rules for manifest files
When you import a manifest file, Amazon Rekognition Custom Labels applies validation rules for limits, syntax, and semantics. The SageMaker Ground Truth schema enforces syntax validation. For more information, see Output. The following are the validation rules for limits and semantics.
Note
-
The 20% invalidity rules apply cumulatively across all validation rules. If the import exceeds the 20% limit due to any combination, such as 15% invalid JSON and 15% invalid images, the import fails.
-
Each dataset object is a line in the manifest. Blank/invalid lines are also counted as dataset objects.
-
Overlaps are (common labels between test and train)/(train labels).
Limits
| Validation | Limit | Error raised |
|---|---|---|
|
Manifest file size |
Maximum 1 GB |
Error |
|
Maximum line count for a manifest file |
Maximum of 250,000 dataset objects as lines in a manifest. |
Error |
|
Lower boundary on total number of valid dataset objects per label |
>= 1 |
Error |
|
Lower boundary on labels |
>=2 |
Error |
|
Upper bound on labels |
<= 250 |
Error |
|
Minimum bounding boxes per image |
0 |
None |
|
Maximum bounding boxes per image |
50 |
None |
Semantics
| Validation | Limit | Error raised |
|---|---|---|
|
Empty manifest |
Error |
|
|
Missing/in-accessible source-ref object |
Number of objects less than 20% |
Warning |
|
Missing/in-accessible source-ref object |
Number of objects > 20% |
Error |
|
Test labels not present in training dataset |
At least 50% overlap in the labels |
Error |
|
Mix of label vs. object examples for same label in a dataset. Classification and detection for the same class in a dataset object. |
No error or warning |
|
|
Overlapping assets between test and train |
There should not be an overlap between test and training datasets. |
|
|
Images in a dataset must be from same bucket |
Error if the objects are in a different bucket |
Error |
