Usage of Sagemaker Processing Job Manifest File

Question

I have a processing Job that uses input files saved in different folders of a S3 Bucket and use the Manifest file within the processing Job to copy it to /opt/ml/processing/input Folder.

This works perfectly fine when i have all the files in one folder but wont work when they are under the same prefix but under different folders.

Following the steps listed in the url https://docs.aws.amazon.com/sagemaker/latest/APIReference/API_S3DataSource.html

[ {"prefix": "s3://customer_bucket/some/prefix/"},

"relative/path/to/custdata-1",

"relative/path/custdata-2",

...

"relative/path/custdata-N"

]

If i have all the input files in "relative/path1/custdata-1" The job works fine but if i add another one "relative/path2/custdata-2", there is no file copied and my script fails with no such file or directory.

Any suggestions or advise on this will be very helpful.

Answer

I've used Processing Job ManifestFile inputs successfully in the past with multiple relative folders (for e.g. the "Extract clean input images" section of [this notebook](https://github.com/aws-samples/amazon-textract-transformer-pipeline/blob/main/notebooks/1.%20Data%20Preparation.ipynb) - sorry for citing a large/sprawling sample, there are probably simpler ones out there).

Not sure exactly what could be going wrong here so I'll try to describe using the feature as I think of it and hope that helps:

Given an S3 bucket containing:

```
s3://customer_bucket/some/prefix/relative/path1/custdata-1
s3://customer_bucket/some/prefix/relative/path2/custdata-2
```

...and a manifest file like:

```json
[ { "prefix":  "s3://customer_bucket/some/prefix/" },
  "relative/path1/custdata-1",
  "relative/path2/custdata-2"
]
```

...for a processing input something like the below (or equivalent if you're using boto3/etc instead of the [SageMaker Python SDK](https://sagemaker.readthedocs.io/en/stable/api/training/processing.html#sagemaker.processing.ProcessingInput)):

```python
ProcessingInput(
    destination="/opt/ml/processing/input/mycoolinput",
    input_name="mycoolinput",
    s3_data_type="ManifestFile",
    source="s3://path-to-your-manifest-file",
)
```

...I'd expect your processing job to see files:

```
/opt/ml/processing/input/mycoolinput/relative/path1/custdata-1
/opt/ml/processing/input/mycoolinput/relative/path2/custdata-2
```

So in this sense it is possible to have files under the same prefix with different subfolders. In the above mentioned sample, the `raw_s3uri` prefix contains credit card agreement PDFs categorized into folders by bank/provider - e.g. `{raw_s3uri}/Bank1/Card1.pdf`, `{raw_s3uri}/CreditUnion2/Disclosures.pdf`, etc.

To my knowledge it's not possible to have multiple `{ "prefix": "..." }` entries in your manifest, but as I understood it didn't sound like you were trying to do that.

Apart from double-checking this overall setup (and maybe using Python `os.walk()` to recursively `print()` out the folder contents as your Processing job sees them), the only other thing I could suggest is to check if your S3 object keys have any special characters in them that could be causing issues when mapping to a local filesystem - such as files/folders with spaces at the end, or characters that aren't usually allowed in filenames?

Usage of Sagemaker Processing Job Manifest File

관련 콘텐츠